AI MonitorPrototype
One mark per launch from OpenAI, Anthropic, Google, xAI, Meta, China, Europe and the rest, with open source shown apart. Higher means a bigger change in practical work for people building an AI business brain, as scored by three AIs: Claude, ChatGPT and Jev. The last 30 days in detail; earlier years only the big moves. The Story tab explains where it is all heading.
Company productOpen source or open weightsOutside the filterRange of the three AI raters
Newest first. At most five launches a week, the biggest first. The log follows every filter above.
No launches for this filter.
Desktop app that sets itself up by chat, connectors to apps like Linear and Notion, local open models with no account, 600 reviewed plugins.
For AIOS builders: Running your own agents gets close to a normal app install, and local models keep sensitive work on your machine.
AI raters: Claude 75 · ChatGPT 85 · Jev 70 (Jev 59% sure) · pushes: owner control, local
Meta released Muse Wonder (API and Muse Code) and Muse Wonder Full (Muse Code only), with Active Memory and Active Focus for long agent tasks.
For AIOS builders: Meta's API now has a frontier option aimed at agents, with memory across conversations; no open weights.
AI raters: Claude 54 · ChatGPT 65 · Jev 75 (Jev 43% sure) · pushes: platform hold, cloud
Workspaces can choose where AI processes their requests, including EU-only inference.
For AIOS builders: EEA businesses can build with AI while keeping prompts and data processed inside the EU, which eases GDPR reviews.
AI raters: Claude 54 · ChatGPT 66 · Jev 63 (Jev 83% sure) · pushes: owner control, cloud
Built-in Google Ads and Meta ads: connect or create accounts, add tracking and consent, make the ads and launch campaigns after approval. All plans, no fee on ad spend.
For AIOS builders: Ad campaigns move into the same chat that builds the app; check the approval step, Google campaigns spend once approved.
AI raters: Claude 52 · ChatGPT 59 · Jev 64 (Jev 75% sure) · pushes: platform hold, cloud
Gemini 4 Argon reached the API, AI Studio, Antigravity and the Gemini app at $2/$10 per million tokens until Nov 30, then $4/$20.
For AIOS builders: A new frontier model at a low intro price; the Gemini app also gained a tasks mode that keeps working while you are away.
AI raters: Claude 56 · ChatGPT 56 · Jev 63 (Jev 69% sure) · pushes: cloud
At Gemini at Work, Google introduced the Gemini agent; coworker agents get their own Workspace account, email, calendar and Drive, plus skills registries.
For AIOS builders: An agent can hold a real seat in the company directory with its own permissions, instead of acting as a person.
AI raters: Claude 60 · ChatGPT 81 · Jev 72 (Jev 43% sure) · pushes: platform hold, cloud
GPT-6 Instant (Sol for paid plans, Luna for Free and Go from Oct 8) became the Chat model; answers can include charts, buttons, forms and small tools.
For AIOS builders: Answers can become small interactive tools, so simple calculators and forms no longer need a separate app.
AI raters: Claude 52 · ChatGPT 71 · Jev 76 (Jev 41% sure) Raters disagree · pushes: platform hold, cloud
Claude Haiku 5.5 launched with 1M context, 128k output and effort levels, costing about 75% less to run than Haiku 4.5.
For AIOS builders: High-volume steps like triage, extraction and sub-agents can run on Claude at much lower cost.
AI raters: Claude 54 · ChatGPT 73 · Jev 66 (Jev 85% sure) · pushes: cloud
Claude builds live dashboards from Snowflake, BigQuery, Databricks or Salesforce data, and Motion makes editable animated explainers exported as MP4.
For AIOS builders: Recurring reports can be a live Claude dashboard with visible queries instead of a manual export each week.
AI raters: Claude 52 · ChatGPT 65 · Jev 67 (Jev 80% sure) · pushes: platform hold, cloud
Docs can open, edit and co-edit .md files without converting them, and Drive shows rendered Markdown previews, for all Workspace and personal accounts.
For AIOS builders: Rules, skills and memory kept as .md files can be read and edited by non-technical staff in Docs.
AI raters: Claude 54 · ChatGPT 80 · Jev 64 (Jev 85% sure) · pushes: owner control, cloud
At DevDay OpenAI launched Dots: persistent ChatGPT agents on GPT-6 Astra, each with a cloud computer and browser, working between chats with approvals.
For AIOS builders: Background agents that keep working on a goal became a ChatGPT feature, not a custom build. Pro and Business Premium first.
AI raters: Claude 62 · ChatGPT 83 · Jev 81 (Jev 84% sure) · pushes: platform hold, cloud
ChatGPT Space is a shared home for files and project context for teams and agents, replacing Library; Pages are documents people and agents co-edit.
For AIOS builders: Shared business context can live in one ChatGPT workspace that both staff and agents read and update.
AI raters: Claude 58 · ChatGPT 76 · Jev 74 (Jev 35% sure) · pushes: platform hold, cloud
Google announced skills: reusable instructions run with a slash, stackable and buildable from chats, in the SKILL.md format. Gems will be phased out.
For AIOS builders: Gemini now uses the same skill format as Claude and others, so one skill library can serve several assistants.
AI raters: Claude 70 · ChatGPT 76 · Jev 66 (Jev 69% sure) · pushes: owner control, cloud
Claude Sonnet 5.5 launched at $2/$10 per million tokens, over 30% faster than Sonnet 5, scoring 70.6% on Terminal-Bench 4.0 against 10.3%.
For AIOS builders: The mid-priced Claude became strong enough for most agent and coding work, using fewer tokens per task.
AI raters: Claude 54 · ChatGPT 72 · Jev 70 (Jev 59% sure) · pushes: cloud
GPT-6.1 Sol launched in ChatGPT Work, Codex and the API at $2/$10 per million tokens, close to GPT-6 Astra on coding and computer use.
For AIOS builders: Strong agent quality at mid-tier prices; it became the Codex default on Oct 8.
AI raters: Claude 54 · ChatGPT 75 · Jev 67 (Jev 81% sure) · pushes: cloud
n8n launched Agents (preview): goal-driven agents with MCP tools, reusable skills, memory across sessions, and triggers from chat, schedules or workflows.
For AIOS builders: A widely used automation tool now hosts agents with skills and memory next to existing workflows, close to the AIOS pattern.
AI raters: Claude 66 · ChatGPT 79 · Jev 78 (Jev 71% sure)
New Copilot: Home joins chat and Cowork, Code builds apps and workflows, and Autopilot is a persistent cloud agent with its own identity. Frontier users first.
For AIOS builders: Microsoft 365 teams get a native always-on agent and plugin registry, a direct alternative to building on Claude or ChatGPT.
AI raters: Claude 58 · ChatGPT 76 · Jev 69 (Jev 49% sure) · pushes: platform hold, cloud
Anthropic opened Claude Marketplace: over 2,000 connectors and plugins, partner agents payable from Anthropic spend, and consulting partners.
For AIOS builders: Finding a ready connector or plugin got easier than building one, and partner agents can be bought on existing contracts.
AI raters: Claude 58 · ChatGPT 69 · Jev 72 (Jev 41% sure) · pushes: platform hold, cloud
GPT-6 Sol and GPT-6 Luna came to the API, ChatGPT Work and Codex for paid plans, priced about 50% below GPT-5.6, with higher usage limits.
For AIOS builders: GPT-6 quality became cheap enough for routine agent steps, not only hard tasks.
AI raters: Claude 54 · ChatGPT 72 · Jev 65 (Jev 92% sure) · pushes: cloud
Claude Opus 5.5 launched at $4/$20 per million tokens, at about Fable 5.1 level on most work, over 30% faster than Opus 5 and with cheaper cache reads.
For AIOS builders: Near-flagship quality for daily agent and coding work at about 40% lower cost than Opus 5.
AI raters: Claude 54 · ChatGPT 71 · Jev 63 (Jev 79% sure) · pushes: cloud
Redesigned Claude projects (beta) use a coordinator that splits a goal into parallel threads, with shared project memory and a file library.
For AIOS builders: A project can work like a small team on one goal, closer to how an AIOS runs multi-step business work.
AI raters: Claude 60 · ChatGPT 77 · Jev 75 (Jev 47% sure) · pushes: platform hold, cloud
Notion 3.7 added a shared skills library with SKILL.md export, sub-agents for Custom Agents, custom MCP connections and an Agent SDK in public beta.
For AIOS builders: Team skills can live in Notion and be exported in the same SKILL.md format other agents use.
AI raters: Claude 66 · ChatGPT 77 · Jev 74 (Jev 40% sure) · pushes: owner control, cloud
Gemini in Workspace got MCP connectors for Asana, Atlassian Rovo, HubSpot, Mailchimp, QuickBooks, monday and Salesforce, on by default with admin control.
For AIOS builders: Gemini in Gmail, Docs and Chat can read and act on CRM and project data without custom integration work.
AI raters: Claude 62 · ChatGPT 73 · Jev 71 (Jev 53% sure) · pushes: platform hold, cloud
Jev, TypeSafe AI's first System One model, entered early access. It returns typed choices, scores and probabilities at $0.042 per million input tokens.
For AIOS builders: Routing, triage and yes/no checks can use a fast decision model instead of parsing chat output. OpenAI, AWS and others followed.
AI raters: Claude 52 · ChatGPT 65 · Jev 66 (Jev 80% sure) · pushes: cloud
Claude chat and Cowork became one app that picks the mode itself; Claude Docs and Slides launched in beta, and Claude Design moved into chat.
For AIOS builders: One place for chat, agent work and documents; builders no longer route tasks between apps by hand.
AI raters: Claude 54 · ChatGPT 64 · Jev 65 (Jev 84% sure) · pushes: platform hold, cloud
Public beta of the Agents API: OpenAI runs the agent loop on the Codex harness, in OpenAI-hosted or your own sandboxes, with skills and plugins.
For AIOS builders: Builders no longer need their own agent loop and sandbox to run OpenAI agents with skills in production.
AI raters: Claude 66 · ChatGPT 82 · Jev 80 (Jev 83% sure) · pushes: owner control, cloud
DeepSeek released V4.1-Flash, a 552B MoE with native vision and published weights, with lower API prices and a much smaller KV cache.
For AIOS builders: Cheaper cache hits suit agents that resend long context, and open weights allow self-hosting.
AI raters: Claude 58 · ChatGPT 73 · Jev 71 (Jev 51% sure) · pushes: owner control, local
GPT-6 Astra rolled out to trusted cyber partners, then Pro, Business and Enterprise in ChatGPT Work and Codex, and to the API at $10/$50 with 1M context.
For AIOS builders: A new top model for long agent runs in ChatGPT Work and Codex; it later powered ChatGPT Dots.
AI raters: Claude 58 · ChatGPT 68 · Jev 72 (Jev 43% sure) · pushes: cloud
Google released Gemini 3.8 Flash at $0.75/$3.75 per million tokens (intro price until Dec 31) across the API, Gemini app, Sheets and Gemini Enterprise.
For AIOS builders: A cheap, fast model good enough for most agent steps, also available inside Workspace.
AI raters: Claude 54 · ChatGPT 70 · Jev 64 (Jev 80% sure) · pushes: cloud
Anthropic released Claude Fable 5.1 for paid plans and the API at $10/$50 per million tokens, with Mythos 5.1 limited to vetted organizations.
For AIOS builders: The strongest Claude for long agent work got cheaper cache reads and per-message effort control.
AI raters: Claude 53 · ChatGPT 56 · Jev 57 (Jev 53% sure) · pushes: cloud
Named bots with their own memory, a shared cloud computer, connectors, approvals, saved skills and scheduled routines.
For AIOS builders: Close to a ready-made business brain: role memory, skills taught once, routines on a schedule and approval gates in one product.
AI raters: Claude 58 · ChatGPT 78 · Jev 80 (Jev 77% sure) · pushes: platform hold, cloud
Agent Skills, the Files API and computer use became generally available on the Claude Platform, and a new browser use tool launched.
For AIOS builders: Skills-based agents built on the API could move to production without beta headers.
AI raters: Claude 68 · ChatGPT 73 · Jev 77 (Jev 57% sure) · pushes: platform hold, cloud
Alibaba released open weights for Qwen3.8-2.4T-A95B, an always-thinking text model with up to about 1M context; Qwen3.8-27B followed on Aug 14.
For AIOS builders: A very large open model became available for self-hosted agents where data must stay in house.
AI raters: Claude 66 · ChatGPT 76 · Jev 72 (Jev 44% sure) · pushes: owner control, local
Claude memory became shared between chat and Cowork and saves topics as you chat, on by default for Free, Pro and Max, with view and edit controls.
For AIOS builders: Persistent context no longer depends on hand-written project files alone; builders now also curate what Claude remembers.
AI raters: Claude 60 · ChatGPT 70 · Jev 68 (Jev 68% sure) · pushes: platform hold, cloud
DeepSeek-V4-Pro GA improved agent ability, added Responses API support and Codex adaptation, effort levels and half-price off-peak pricing.
For AIOS builders: A low-cost model could plug into Codex-style agent harnesses directly, cutting the cost of batch and overnight jobs.
AI raters: Claude 54 · ChatGPT 66 · Jev 66 (Jev 88% sure) · pushes: cloud
The AI Act became generally applicable: AI Office enforcement, GPAI duties and Article 50 transparency rules. High-risk rules were moved to Dec 2027.
For AIOS builders: EU businesses must tell people when they deal with an AI and mark AI-generated content; high-risk use cases got more time.
AI raters: Claude 62 · ChatGPT 77 · Jev 65 (Jev 39% sure)
The fifth MCP spec release moved to stateless request/response, put Apps and Tasks under versioned extensions and aligned auth with OAuth 2.0 and OIDC.
For AIOS builders: Connectors got easier to host on serverless infrastructure and to put behind company sign-in such as Entra or Okta.
AI raters: Claude 80 · ChatGPT 84 · Jev 73 (Jev 45% sure) · pushes: owner control
GPT-5.6 Sol, Terra and Luna became generally available, and ChatGPT Work added an agent mode in one desktop app with Chat and Codex.
For AIOS builders: Long business tasks could run inside ChatGPT itself, with three price tiers to match each workflow step to a model.
AI raters: Claude 56 · ChatGPT 73 · Jev 74 (Jev 35% sure) · pushes: platform hold, cloud
Anthropic released Claude Opus 5 at $5/$25 per million tokens, close to Claude Fable 5 on coding and knowledge work, and the new Max default.
For AIOS builders: Near top-tier quality became affordable as the everyday model for agents, coding and document work.
AI raters: Claude 55 · ChatGPT 70 · Jev 65 (Jev 81% sure) · pushes: cloud
Kimi K3 (2.8T MoE, about 104B active, 1M context, vision) launched in the API, Kimi Work and Kimi Code, with open weights promised.
For AIOS builders: A frontier-class Chinese model with promised weights gave builders a cheaper fallback for long agent and coding runs.
AI raters: Claude 58 · ChatGPT 65 · Jev 73 (Jev 33% sure) · pushes: owner control, cloud
Most agentic Sonnet yet, default for Free and Pro, API at $2/$10 per million tokens, cyber safeguards on by default, higher rate limits.
For AIOS builders: Close to flagship agent quality at a lower price, making always-on business agents cheaper to run.
AI raters: Claude 57 · ChatGPT 71 · Jev 68 (Jev 71% sure) · pushes: cloud
Spark beta on macOS for US AI Ultra users, desktop file tasks, new connected apps (Tasks, Keep, Canva, Dropbox) and custom MCP connections.
For AIOS builders: Google's background agent can now reach your own tools via MCP, the same connector standard used across other agent stacks.
AI raters: Claude 58 · ChatGPT 69 · Jev 62 (Jev 68% sure) · pushes: platform hold, cloud
Claude Fable 5 for general use at $10/$50 with 1M context, Mythos 5 for vetted partners. Access was suspended Jun 12 and restored Jul 1.
For AIOS builders: Top-tier capability became available for hard agent work, at a higher price and with some security and bio queries routed to Opus 4.8.
AI raters: Claude 55 · ChatGPT 64 · Jev 61 (Jev 39% sure) · pushes: cloud
Multiagent orchestration, Outcomes and webhooks in public beta. Dreams research preview reorganizes agent memory between sessions. Memory beta came Apr 23.
For AIOS builders: Hosted agents can coordinate sub-agents and curate long-term memory from past sessions, a core need for a business brain.
AI raters: Claude 64 · ChatGPT 78 · Jev 78 (Jev 69% sure) · pushes: platform hold, cloud
Opus upgrade at $5/$25, dynamic workflows research preview with hundreds of parallel subagents, effort control in claude.ai and Cowork, cheaper fast mode.
For AIOS builders: Very large jobs can fan out to many subagents, and system messages can change instructions mid-task.
AI raters: Claude 55 · ChatGPT 73 · Jev 75 (Jev 49% sure) · pushes: platform hold, cloud
128B dense open-weight model (modified MIT), 256k context, $1.5/$7.5. Vibe remote coding agents in the cloud and a Work mode preview in Le Chat.
For AIOS builders: A self-hostable European model on four GPUs plus async cloud agents, relevant for EU data residency needs.
AI raters: Claude 60 · ChatGPT 72 · Jev 73 (Jev 40% sure) · pushes: owner control, cloud
Skills in Lovable package instructions you repeat so the agent applies them across projects.
For AIOS builders: The same skills idea as Claude and Gemini reaches app building, so house rules can travel into every app you build.
AI raters: Claude 54 · ChatGPT 61 · Jev 65 (Jev 87% sure) · pushes: platform hold, cloud
Gemini 3.5 Flash generally available and default in Gemini app and AI Mode. Gemini Spark, a background personal agent, went to trusted testers.
For AIOS builders: Fast model that Google says beats 3.1 Pro on coding and agent tests, suited to high-volume agent steps.
AI raters: Claude 53 · ChatGPT 60 · Jev 66 (Jev 79% sure) · pushes: platform hold, cloud
Codex app update with computer use on macOS, in-app browser, image generation, memory preview, 90+ plugins and background automations.
For AIOS builders: Codex grew from a coding tool into a general work agent that remembers preferences and connects to business apps.
AI raters: Claude 60 · ChatGPT 78 · Jev 78 (Jev 69% sure) · pushes: platform hold
Four sizes: E2B, E4B, 26B MoE and 31B dense. First Gemma under Apache 2.0, with function calling and on-device use for the small models.
For AIOS builders: Capable local models with a permissive license, useful for private data, offline steps and cheap agent tool calls.
AI raters: Claude 78 · ChatGPT 81 · Jev 67 (Jev 85% sure) Raters disagree · pushes: owner control, local
V4-Pro (1.6T total, 49B active) and V4-Flash (284B, 13B active), open weights, API live the same day, 1M context default, thinking and non-thinking modes.
For AIOS builders: Near-frontier open weights with 1M context, usable self-hosted or via API as a fallback or cost layer in agent pipelines.
AI raters: Claude 77 · ChatGPT 77 · Jev 76 (Jev 53% sure) · pushes: owner control, local
Hosted agent harness and APIs on the Claude Platform for cloud agents, with state, memory, permissions and scheduled execution handled by Anthropic.
For AIOS builders: Builders could run cloud agents without writing their own loop, sandbox or scheduler, which shortens the path to production.
AI raters: Claude 62 · ChatGPT 79 · Jev 74 (Jev 36% sure) · pushes: platform hold, cloud
New flagship in ChatGPT and Codex, API from Apr 24 at $5/$30 with 1M context. Reports 82.7% on Terminal-Bench 2.0 and 78.7% on OSWorld-Verified.
For AIOS builders: Stronger planning, tool use and computer use with fewer tokens per task, raising the ceiling for long agent runs.
AI raters: Claude 56 · ChatGPT 71 · Jev 72 (Jev 41% sure) · pushes: cloud
Shared agents in ChatGPT Business, Enterprise and Edu, described as an evolution of GPTs, powered by Codex and running in the cloud.
For AIOS builders: Teams could build shared agents that keep working in the background, a direct step toward a company-wide AI brain.
AI raters: Claude 58 · ChatGPT 76 · Jev 69 (Jev 65% sure) · pushes: platform hold, cloud
Opus upgrade at unchanged price, new xhigh effort level, images up to about 3.75 MP, task budgets in API beta, /ultrareview in Claude Code.
For AIOS builders: Better results on long agent tasks and screenshots, with budgets to cap how much an agent spends per task.
AI raters: Claude 53 · ChatGPT 65 · Jev 64 (Jev 84% sure) · pushes: cloud
MIT-licensed skill that transcribes, cuts false starts and filler, burns in captions and adds overlays with HyperFrames or Remotion.
For AIOS builders: Video editing becomes a repeatable skill with your own cut and caption rules, reviewed instead of done by hand.
AI raters: Claude 52 · ChatGPT 53 · Jev 56 (Jev 64% sure) · pushes: owner control
GPT-5.4 Thinking and Pro in ChatGPT, API and Codex. First general OpenAI model with native computer use, up to 1M tokens in API and Codex, $2.50/$15.
For AIOS builders: Agents could operate apps and websites without custom integrations, and one model covered chat, coding and long documents.
AI raters: Claude 68 · ChatGPT 83 · Jev 83 (Jev 84% sure) · pushes: cloud
Computer use research preview in Cowork for Pro and Max, with Dispatch letting Claude use your computer while you are away from it.
For AIOS builders: Tasks could be started from a phone and finished on the desktop, including apps that have no API or connector.
AI raters: Claude 58 · ChatGPT 81 · Jev 72 (Jev 41% sure) · pushes: platform hold, cloud
First tagged public release: one agent reachable from Telegram, Slack, WhatsApp, Discord and email, with MCP tools, 70+ skills and cron jobs. MIT license.
For AIOS builders: A self-hosted business brain runtime: your own agents on your own server, reached from the chat apps the team already uses.
AI raters: Claude 72 · ChatGPT 85 · Jev 68 (Jev 72% sure) · pushes: owner control, local
119B MoE with about 6B active, 256k context, text and image input, configurable reasoning effort. Open weights under Apache 2.0.
For AIOS builders: A European, Apache-licensed model light enough to self-host for agents that need reasoning and image input.
AI raters: Claude 64 · ChatGPT 72 · Jev 71 (Jev 49% sure) · pushes: owner control, local
Smaller, cheaper GPT-5.4 models for high-volume and low-latency work in the API.
For AIOS builders: Cheap models good enough for sub-agents, routing, classification and other high-volume steps in a pipeline.
AI raters: Claude 52 · ChatGPT 65 · Jev 64 (Jev 92% sure) · pushes: platform hold, cloud
Microsoft 365 Copilot feature that plans multi-step tasks, works in the background with checkpoints and waits for approval. Research preview at launch.
For AIOS builders: Agentic task delegation arrived inside the Microsoft 365 suite where many business teams already keep files and mail.
AI raters: Claude 58 · ChatGPT 73 · Jev 63 (Jev 77% sure) · pushes: platform hold, cloud
New default for Free and Pro at $3/$15, 1M context beta. API: code execution, memory, programmatic tool calling and tool search became generally available.
For AIOS builders: Near-flagship agent quality at mid price, and the API pieces for memory and large tool catalogs became production ready.
AI raters: Claude 68 · ChatGPT 83 · Jev 80 (Jev 83% sure) · pushes: cloud
New flagship with 1M-token context (beta), adaptive thinking with effort levels, context compaction, 128k output, agent teams in Claude Code.
For AIOS builders: Long-running agents could keep a whole knowledge base in context and compact old turns, so multi-hour sessions became practical.
AI raters: Claude 70 · ChatGPT 79 · Jev 81 (Jev 89% sure) · pushes: cloud
General-purpose agent for Max subscribers that splits a goal into tasks, delegates to sub-agents and can run for hours. Enterprise Max planned.
For AIOS builders: Another hosted option for delegating whole research and admin outcomes rather than single prompts.
AI raters: Claude 56 · ChatGPT 78 · Jev 78 (Jev 67% sure) · pushes: platform hold, cloud
Qwen3.5-397B-A17B open weights under Apache 2.0, natively multimodal, 201 languages, plus Qwen3.5-Plus on Alibaba Cloud. Smaller sizes followed.
For AIOS builders: Apache-licensed multimodal model with broad language coverage, a strong option for self-hosted or low-cost agents, including smaller languages.
AI raters: Claude 66 · ChatGPT 76 · Jev 70 (Jev 61% sure) · pushes: owner control, local
Desktop command center for running several Codex agents in parallel by project, with worktrees, skills and scheduled automations that feed a review queue.
For AIOS builders: Running many agents at once on separate tasks became a desktop workflow, with skills and scheduled jobs built in.
AI raters: Claude 62 · ChatGPT 80 · Jev 69 (Jev 69% sure) · pushes: platform hold, cloud
Feb 24: Cowork plugin marketplace with admin controls for Team and Enterprise. Feb 25: recurring and on-demand scheduled tasks in Cowork.
For AIOS builders: Repeatable business routines could run on a schedule inside Cowork, packaged as plugins and skills that a team can share.
AI raters: Claude 66 · ChatGPT 77 · Jev 65 (Jev 95% sure) · pushes: platform hold, cloud
Upgraded core model in preview across Gemini API, AI Studio, Gemini CLI, Antigravity, Vertex AI, the Gemini app and NotebookLM. 77.1% on ARC-AGI-2.
For AIOS builders: A top reasoning model on Google's stack with 1M context, available the same day in dev tools and consumer apps.
AI raters: Claude 54 · ChatGPT 64 · Jev 68 (Jev 69% sure) · pushes: platform hold, cloud
Coding model for Codex app, CLI, IDE and web on paid ChatGPT plans, about 25% faster than GPT-5.2-Codex. API access came later.
For AIOS builders: Stronger and faster code agent for building and maintaining the scripts, integrations and tools an AIOS depends on.
AI raters: Claude 54 · ChatGPT 65 · Jev 59 (Jev 36% sure) · pushes: platform hold, cloud
1T-parameter MoE (32B active) with vision, open weights under a modified MIT license, Agent Swarm beta with up to 100 sub-agents, and the Kimi Code CLI.
For AIOS builders: A large open model that can be self-hosted or used via API for agent work, including parallel sub-agents.
AI raters: Claude 62 · ChatGPT 78 · Jev 81 (Jev 85% sure) · pushes: owner control, local
Research preview in the Claude macOS app for Max users: Claude works on local files in an isolated VM with MCP connectors. Pro access followed on Jan 16.
For AIOS builders: Agentic file and tool work moved from the terminal to a desktop app, so non-developers could run multi-step business tasks with skills and connectors.
AI raters: Claude 77 · ChatGPT 84 · Jev 68 (Jev 59% sure) Raters disagree · pushes: local
Self-hosted agent platform that runs on a laptop, homelab or VPS and works through WhatsApp, Telegram, Discord, Slack and Teams. Renamed from Clawd and Moltbot.
For AIOS builders: Made the always-on personal agent on your own server, reached by chat, a common pattern with skills and memory under your control.
AI raters: Claude 76 · ChatGPT 86 · Jev 53 (Jev 85% sure) Raters disagree · pushes: owner control, local
US beta for AI Pro and Ultra subscribers: Gemini can draw on Gmail, Google Photos, YouTube and Search history for personal answers. Opt-in per app.
For AIOS builders: Showed the value of an assistant that reads your own mail, photos and history with consent. Personal accounts only, not Workspace.
AI raters: Claude 53 · ChatGPT 62 · Jev 55 (Jev 73% sure) · pushes: platform hold, cloud
Skills spec opened for other platforms, with org-wide skill management for admins and a partner skills directory.
For AIOS builders: Skills written once can move between tools and vendors, and teams can roll them out centrally.
AI raters: Claude 86 · ChatGPT 89 · Jev 71 (Jev 52% sure) Raters disagree · pushes: owner control
V3.2 (app, web, API, open weights) and API-only V3.2-Speciale, reasoning-first models for agents with tool use while thinking.
For AIOS builders: Cheap open reasoning model that can call tools mid-reasoning, after a 50%+ API price cut with V3.2-Exp in September.
AI raters: Claude 60 · ChatGPT 73 · Jev 75 (Jev 47% sure) · pushes: owner control, local
Fast, lower-cost Gemini 3 model, the new default in the Gemini app and AI Mode in Search, also in the API.
For AIOS builders: Put near-frontier reasoning at Flash speed and price into high-volume agent work.
AI raters: Claude 54 · ChatGPT 65 · Jev 60 (Jev 47% sure) · pushes: platform hold, cloud
Open-source engine on CAMEL-AI's OASIS: seed material becomes a knowledge graph and agents with personas react on two simulated platforms.
For AIOS builders: A cheap pre-test of offers and launch pages before real customers see them; no published accuracy data yet.
AI raters: Claude 52 · ChatGPT 53 · Jev 60 (Jev 49% sure) · pushes: owner control, local
Gemini 3 Pro across the Gemini app, Search and API, plus Google Antigravity, an agent-first development platform.
For AIOS builders: Raised the bar for reasoning and coding and added a new agent workspace option for builders.
AI raters: Claude 66 · ChatGPT 77 · Jev 78 (Jev 64% sure) · pushes: platform hold, cloud
Tool-calling model with 2M context plus an Agent Tools API with server-side web search, X search and code execution.
For AIOS builders: Agents can search the web and run code through the API without the builder hosting those tools.
AI raters: Claude 54 · ChatGPT 70 · Jev 75 (Jev 41% sure) · pushes: platform hold, cloud
Reasoning version of Kimi K2 (1T MoE, 32B active) with open weights, built to run long chains of tool calls.
For AIOS builders: Open model that reasons between many tool calls, so long agent tasks can run without a closed vendor.
AI raters: Claude 60 · ChatGPT 73 · Jev 69 (Jev 65% sure) · pushes: owner control, local
Opus 4.5 for coding, agents and computer use, priced at $5/$25 per million tokens, with an effort parameter in the API.
For AIOS builders: Made top-tier Claude affordable for everyday agent work instead of only special cases.
AI raters: Claude 66 · ChatGPT 73 · Jev 65 (Jev 93% sure) · pushes: cloud
Image generation and editing with better text rendering, more control and world knowledge.
For AIOS builders: Readable text in images made infographics, slides and ads practical to generate.
AI raters: Claude 51 · ChatGPT 56 · Jev 60 (Jev 44% sure) · pushes: cloud
Skills are folders of instructions, scripts and resources that Claude loads only when relevant, in apps, Claude Code and API.
For AIOS builders: Gave a standard format for repeatable business procedures without filling every prompt with them.
AI raters: Claude 84 · ChatGPT 88 · Jev 71 (Jev 48% sure) Raters disagree · pushes: owner control
Composer, Cursor's own fast coding model, and an agent-centered interface for running many agents in parallel.
For AIOS builders: Running several agents side by side became a normal way to work, not only one chat at a time.
AI raters: Claude 53 · ChatGPT 69 · Jev 72 (Jev 40% sure) · pushes: platform hold, cloud
Open-weights model for agents and coding, API at $0.30 input and $1.20 output per 1M, about 8% of Claude Sonnet 4.5 price.
For AIOS builders: Low-cost agent backend with open weights, useful for high-volume tool-calling jobs.
AI raters: Claude 56 · ChatGPT 69 · Jev 66 (Jev 95% sure) · pushes: owner control, local
Reasoning model with open weights under MIT, six distilled smaller models, API at $0.55 input and $2.19 output per 1M.
For AIOS builders: o1-level reasoning became cheap and self-hostable, and visible reasoning traces made agent steps easier to inspect.
AI raters: Claude 82 · ChatGPT 88 · Jev 73 (Jev 39% sure) Raters disagree · pushes: owner control, local
New Responses API with built-in web search, file search and computer use, plus the open-source Agents SDK.
For AIOS builders: Gave developers standard building blocks for tool-using agents instead of custom orchestration code.
AI raters: Claude 72 · ChatGPT 84 · Jev 80 (Jev 85% sure) · pushes: cloud
gpt-oss-120b and gpt-oss-20b, open-weight reasoning models sized for a single GPU or a high-end laptop.
For AIOS builders: Enabled private, local or self-hosted AI setups with strong tool use and no per-token API fees.
AI raters: Claude 78 · ChatGPT 85 · Jev 76 (Jev 51% sure) · pushes: owner control, local
Agent mode with its own virtual computer to research, fill forms and build slides and spreadsheets, under user control.
For AIOS builders: Brought multi-step task execution across websites and files to Pro, Plus and Team users.
AI raters: Claude 62 · ChatGPT 84 · Jev 78 (Jev 63% sure) · pushes: platform hold, cloud
Reasoning models that use search, Python, files and images inside ChatGPT; open-source Codex CLI for the terminal.
For AIOS builders: Reasoning models became tool users, and Codex CLI brought a local coding agent to builders.
AI raters: Claude 62 · ChatGPT 80 · Jev 78 (Jev 61% sure) · pushes: owner control
Hybrid reasoning model with optional extended thinking, plus Claude Code, a command-line agent for coding tasks.
For AIOS builders: Claude Code turned a folder of files into a working agent environment, a common home for AI business setups.
AI raters: Claude 78 · ChatGPT 85 · Jev 70 (Jev 60% sure) Raters disagree · pushes: cloud
Apps that run inside ChatGPT via the MCP-based Apps SDK, AgentKit with Agent Builder, and Codex general availability.
For AIOS builders: Made MCP a route into ChatGPT and added visual tools for building and shipping agents.
AI raters: Claude 70 · ChatGPT 81 · Jev 77 (Jev 57% sure) · pushes: platform hold, cloud
Claude Opus 4 and Sonnet 4 for long agent tasks; Claude Code GA with VS Code, JetBrains and GitHub Actions.
For AIOS builders: Agents could run for hours on multi-step work, and Claude Code became a supported daily tool.
AI raters: Claude 72 · ChatGPT 81 · Jev 77 (Jev 60% sure) · pushes: cloud
Sonnet 4.5 for coding and long agent runs, plus the Claude Agent SDK, the building blocks behind Claude Code.
For AIOS builders: Let builders reuse Claude Code's agent harness for their own non-coding business agents.
AI raters: Claude 76 · ChatGPT 83 · Jev 72 (Jev 48% sure) · pushes: owner control, cloud
Claude remembers project and work context across chats, with project-scoped memory users can view and edit.
For AIOS builders: Added built-in, editable memory to Claude, kept separate per project so client work stays apart.
AI raters: Claude 62 · ChatGPT 77 · Jev 74 (Jev 39% sure) · pushes: platform hold, cloud
Platform with a no-code workbench to build and run agents across company data, plus prebuilt Google agents.
For AIOS builders: Gave companies a managed place to build, share and govern business agents.
AI raters: Claude 58 · ChatGPT 73 · Jev 74 (Jev 35% sure) · pushes: platform hold, cloud
Unified system that switches between fast answers and deeper reasoning; available to all ChatGPT users and in the API.
For AIOS builders: Raised the floor for coding and agentic tool use and removed most manual model picking.
AI raters: Claude 66 · ChatGPT 78 · Jev 73 (Jev 32% sure) · pushes: platform hold, cloud
Open-weights MoE model with 1T total and 32B active parameters, trained for tool use and agentic tasks, modified MIT license.
For AIOS builders: An open model good at long tool-calling chains, usable as a cheaper backend for coding and agent harnesses.
AI raters: Claude 66 · ChatGPT 76 · Jev 73 (Jev 36% sure) · pushes: owner control, local
Bugbot PR review, Background Agent for all users, Memories (beta) per project, and one-click MCP install with OAuth.
For AIOS builders: Project memory, remote agents and easy MCP setup brought core AIOS ideas into a mainstream coding tool.
AI raters: Claude 58 · ChatGPT 73 · Jev 77 (Jev 63% sure) · pushes: platform hold, cloud
A cloud software engineering agent that works on many tasks in parallel in sandboxes and proposes pull requests.
For AIOS builders: Made it practical to hand off code changes to an agent and review the results.
AI raters: Claude 58 · ChatGPT 77 · Jev 73 (Jev 33% sure) · pushes: platform hold, cloud
Gemini 2.5 Pro, a reasoning model that debuted at #1 on LMArena, shipped with a 1M-token context window.
For AIOS builders: Gave builders a top reasoning and coding model with very long context at competitive prices.
AI raters: Claude 64 · ChatGPT 78 · Jev 72 (Jev 44% sure) · pushes: cloud
Two MoE models (235B-A22B, 30B-A3B) and six dense 0.6B to 32B, Apache 2.0, switchable thinking mode, stronger MCP support.
For AIOS builders: One open family for fast replies and deep reasoning, with tool use via MCP; a default pick for local agents.
AI raters: Claude 68 · ChatGPT 83 · Jev 71 (Jev 53% sure) · pushes: owner control, local
One model with thinking and non-thinking modes, stronger tool use and multi-step agent skills, 128K context, Anthropic API format.
For AIOS builders: Could be plugged into tools built for Anthropic's API, such as coding agents, as a much cheaper backend.
AI raters: Claude 57 · ChatGPT 70 · Jev 74 (Jev 39% sure) · pushes: owner control, cloud
Qwen3-Coder-480B-A35B open model with 256K context (1M with extrapolation) and the open-source Qwen Code command-line agent.
For AIOS builders: An open coding agent stack similar to closed CLI tools, for building and maintaining automations with own models.
AI raters: Claude 60 · ChatGPT 78 · Jev 70 (Jev 57% sure) · pushes: owner control, local
Claude apps connect to remote MCP servers such as Jira, Zapier and Asana, with advanced Research across them.
For AIOS builders: Brought MCP tools into the chat app, so non-developers could connect business systems to Claude.
AI raters: Claude 70 · ChatGPT 77 · Jev 68 (Jev 65% sure) · pushes: owner control, cloud
Agent mode preview that iterates on its own code, fixes errors and suggests terminal commands; Copilot Edits became GA.
For AIOS builders: Agentic multi-step coding moved into the most used editor, so non-specialists could build tools by asking.
AI raters: Claude 56 · ChatGPT 76 · Jev 68 (Jev 75% sure) · pushes: platform hold, cloud
An agent that browses and reasons for minutes to produce cited multi-step research reports; Pro first, then Plus and Team.
For AIOS builders: Turned market, competitor and topic research into a delegated task that returns a report.
AI raters: Claude 66 · ChatGPT 79 · Jev 68 (Jev 75% sure) · pushes: platform hold, cloud
Grok 4 Fast in reasoning and non-reasoning API variants, each with a 2M-token context, from $0.20 input and $0.50 output per 1M.
For AIOS builders: Very long context at low cost makes it practical to load large document sets or long histories into one call.
AI raters: Claude 53 · ChatGPT 67 · Jev 67 (Jev 77% sure) · pushes: cloud
Public beta of plugins that bundle slash commands, subagents, MCP servers and hooks, installed via /plugin.
For AIOS builders: Made a whole AI setup shareable and installable as one package across a team.
AI raters: Claude 64 · ChatGPT 79 · Jev 66 (Jev 88% sure) · pushes: owner control
Lovable added a built-in backend (database, auth, storage) and AI features apps can call without setting up accounts or keys.
For AIOS builders: A business owner can ship a working app with data and AI in it, without a developer wiring services together.
AI raters: Claude 52 · ChatGPT 72 · Jev 66 (Jev 84% sure) · pushes: platform hold, cloud
Claude creates and edits Excel, Word, PowerPoint and PDF files directly in Claude.ai and the desktop app.
For AIOS builders: Moved output from chat text to finished business documents ready to send.
AI raters: Claude 60 · ChatGPT 76 · Jev 66 (Jev 84% sure) · pushes: platform hold, cloud
GLM-4.5 (355B total, 32B active) and GLM-4.5-Air (106B, 12B active), MIT license, with switchable reasoning.
For AIOS builders: Another strong open agentic model with a permissive license, usable as a low-cost backend for coding and tool-calling agents.
AI raters: Claude 58 · ChatGPT 71 · Jev 66 (Jev 89% sure) · pushes: owner control, local
Assign an issue to Copilot; it works in a GitHub Actions environment and returns a draft pull request; MCP servers supported.
For AIOS builders: Delegating routine changes as tickets became normal, with review through pull requests and tools via MCP.
AI raters: Claude 56 · ChatGPT 72 · Jev 66 (Jev 76% sure) · pushes: platform hold, cloud
Invite-only research preview of an agent that plans multi-step tasks and runs tools in a cloud virtual computer.
For AIOS builders: Showed a general agent doing whole work tasks end to end, setting a template for agent loops, files and tools.
AI raters: Claude 54 · ChatGPT 70 · Jev 66 (Jev 56% sure) · pushes: platform hold, cloud
Agent Mode (beta) lets Lovable explore the codebase, plan, search the web and fix its own errors before answering.
For AIOS builders: App building moved from one prompt per change to an agent that finishes a task, closer to delegating work.
AI raters: Claude 50 · ChatGPT 65 · Jev 69 (Jev 69% sure) · pushes: platform hold, cloud
Grok 3 and Grok 3 mini beta with reasoning (Think) modes, plus the DeepSearch research agent, on X Premium and Grok.com.
For AIOS builders: Another strong reasoning model and research agent to test for analysis and web research steps in business workflows.
AI raters: Claude 52 · ChatGPT 65 · Jev 66 (Jev 85% sure) · pushes: platform hold, cloud
Reasoning model for agentic coding at $0.20/$1.50 per 1M tokens, free for a limited time in GitHub Copilot, Cursor, Cline and others.
For AIOS builders: A low-cost, fast option inside the coding tools builders already use for scripts, skills and automations.
AI raters: Claude 52 · ChatGPT 64 · Jev 65 (Jev 83% sure) · pushes: cloud
Open-source (Apache 2.0) command-line agent with free use up to 60 requests per minute and 1,000 per day.
For AIOS builders: A free terminal agent for files, code and scripts, widening access to the Claude Code style of work.
AI raters: Claude 64 · ChatGPT 76 · Jev 64 (Jev 89% sure) · pushes: owner control
Native image generation in GPT-4o with accurate text rendering and strong instruction following.
For AIOS builders: Made usable business visuals, mockups and graphics with readable text a routine chat task.
AI raters: Claude 54 · ChatGPT 68 · Jev 61 (Jev 52% sure) · pushes: platform hold, cloud
Stockholm-based Lovable added team workspaces, a chat mode agent for planning and debugging, security scan and visual edits.
For AIOS builders: Non-developers could build and ship internal tools and front ends for their business brain by chatting.
AI raters: Claude 50 · ChatGPT 60 · Jev 65 (Jev 96% sure) · pushes: platform hold, cloud
Google folded Gemini into Workspace Business and Enterprise plans without a separate add-on, with plan price changes.
For AIOS builders: Workspace teams got Gemini in Gmail, Docs and Drive by default, removing a separate buying decision.
AI raters: Claude 52 · ChatGPT 65 · Jev 60 (Jev 43% sure) · pushes: platform hold, cloud
Image model that blends images, keeps characters consistent and edits by text prompt, at $0.039 per image.
For AIOS builders: Made consistent product and brand image editing cheap enough to automate.
AI raters: Claude 52 · ChatGPT 58 · Jev 61 (Jev 64% sure) · pushes: cloud
Grok 4 released in apps and the xAI API on launch day, with a SuperGrok Heavy tier for the multi-agent Grok 4 Heavy.
For AIOS builders: A frontier-level reasoning model available by API from day one, worth benchmarking against other vendors for hard tasks.
AI raters: Claude 52 · ChatGPT 58 · Jev 66 (Jev 79% sure) · pushes: cloud
Veo 3 video with native audio, Imagen 4, an Agent Mode preview in the Gemini app, and Google AI Pro and Ultra plans.
For AIOS builders: Video with sound became usable for marketing content, and Gemini moved toward doing tasks for you.
AI raters: Claude 51 · ChatGPT 58 · Jev 61 (Jev 49% sure) · pushes: platform hold, cloud
Models that edit images from text instructions and reference images without finetuning; [pro] and [max] by API, [dev] later open.
For AIOS builders: Consistent, instruction-based image edits made brand and product visuals repeatable inside automated workflows.
AI raters: Claude 50 · ChatGPT 56 · Jev 63 (Jev 73% sure) · pushes: cloud
Open protocol and SDKs for connecting AI assistants to data sources and tools through reusable servers.
For AIOS builders: Became the common way to plug tools and data into AI systems across vendors, a base layer for many AI setups.
AI raters: Claude 93 · ChatGPT 95 · Jev 93 (Jev 84% sure) · pushes: owner control
Llama 3.1 405B plus upgraded multilingual 8B and 70B, 128K context, with a license allowing outputs to train other models.
For AIOS builders: Frontier-class quality with open weights and long context; cheap hosting via many providers and a source for distillation.
AI raters: Claude 80 · ChatGPT 85 · Jev 76 (Jev 51% sure) · pushes: owner control, local
o1-preview and o1-mini think before answering, with large gains on math, science and coding; in ChatGPT and the API.
For AIOS builders: Started the reasoning-model era: better planning and harder analysis, at the cost of speed.
AI raters: Claude 77 · ChatGPT 84 · Jev 76 (Jev 53% sure) · pushes: platform hold, cloud
Claude 3 models can call developer-defined tools and APIs in production on the Anthropic API and cloud platforms.
For AIOS builders: Let builders connect Claude to their own systems, a precondition for agent workflows.
AI raters: Claude 70 · ChatGPT 77 · Jev 78 (Jev 64% sure) · pushes: cloud
Open-weights MoE model and paper, about 60 tokens/s, API at $0.27 input and $1.10 output per 1M tokens from Feb 2025.
For AIOS builders: Near-frontier quality at a fraction of usual API prices, with weights anyone can host; reset cost expectations.
AI raters: Claude 76 · ChatGPT 80 · Jev 73 (Jev 37% sure) · pushes: owner control, local
Claude can operate a computer by viewing the screen, moving the cursor, clicking and typing; public beta in the API.
For AIOS builders: First frontier model able to work in ordinary software interfaces, opening agent work beyond APIs.
AI raters: Claude 76 · ChatGPT 84 · Jev 75 (Jev 45% sure) · pushes: cloud
Natively multimodal model across text, vision and audio, faster and cheaper than GPT-4 Turbo; GPT-4-level tools opened to free users.
For AIOS builders: Put GPT-4-class models, file analysis and GPTs in front of free users and cut API costs for builders.
AI raters: Claude 66 · ChatGPT 84 · Jev 74 (Jev 33% sure) · pushes: cloud
Gemini 1.5 Pro handled up to 1 million tokens in a limited developer preview, far beyond other models then.
For AIOS builders: Made it realistic to load whole codebases, archives or hours of video into one prompt.
AI raters: Claude 64 · ChatGPT 80 · Jev 74 (Jev 45% sure) · pushes: cloud
Gemini 2.0 Flash Experimental with native tool use, and Deep Research in Gemini Advanced for multi-step cited reports.
For AIOS builders: First deep research agent in a mainstream assistant, plus a fast model built for agent use.
AI raters: Claude 66 · ChatGPT 74 · Jev 72 (Jev 45% sure) · pushes: platform hold, cloud
Gemini 1.5 Flash, a fast low-cost model; 1.5 Pro with 2M-token context via waitlist; Gemma 2 and Project Astra shown.
For AIOS builders: Cheap, fast long-context models made bulk document processing affordable.
AI raters: Claude 58 · ChatGPT 74 · Jev 72 (Jev 43% sure) · pushes: platform hold, cloud
First Llama 3 models at 8B and 70B, pretrained and instruction-tuned, available for broad use.
For AIOS builders: Open models reached quality good enough for many everyday business tasks, so self-hosted assistants became realistic.
AI raters: Claude 72 · ChatGPT 73 · Jev 71 (Jev 55% sure) · pushes: owner control, local
API responses and function calls can be constrained to exactly match a developer-supplied JSON Schema.
For AIOS builders: Made model output safe to feed into automations, databases and other code without manual cleanup.
AI raters: Claude 66 · ChatGPT 83 · Jev 71 (Jev 49% sure) · pushes: cloud
Cache and reuse long prompt prefixes to cut latency by up to 80% and costs by up to 90% in the Claude API.
For AIOS builders: Made it cheap to send the same large rule set, memory and documents on every call.
AI raters: Claude 64 · ChatGPT 74 · Jev 70 (Jev 56% sure) · pushes: cloud
11B and 90B vision models, 1B and 3B text models for edge and mobile, and Llama Stack distributions.
For AIOS builders: Open models could read images and documents, and small models could run on laptops and phones for private tasks.
AI raters: Claude 66 · ChatGPT 77 · Jev 69 (Jev 68% sure) · pushes: owner control, local
Qwen2.5 general models 0.5B to 72B plus Coder and Math variants; most sizes under Apache 2.0.
For AIOS builders: A strong, permissive open family in every size; became a common base for local assistants and fine-tunes.
AI raters: Claude 68 · ChatGPT 77 · Jev 67 (Jev 87% sure) · pushes: owner control, local
Three models at different price and speed levels, with vision input; Opus led several benchmarks at launch.
For AIOS builders: First time Claude matched or beat GPT-4, giving builders a real choice for daily work.
AI raters: Claude 62 · ChatGPT 78 · Jev 68 (Jev 73% sure) · pushes: cloud
Pro and Team users group chats with uploaded documents and per-project instructions, and share work with teammates.
For AIOS builders: A ready-made container for business knowledge and rules per client or workflow.
AI raters: Claude 66 · ChatGPT 74 · Jev 67 (Jev 85% sure) · pushes: platform hold, cloud
Faster mid-tier model that beat Claude 3 Opus, plus Artifacts, a side window for generated code and documents.
For AIOS builders: Made Claude a common choice for coding and drafting, with working outputs shown next to the chat.
AI raters: Claude 64 · ChatGPT 79 · Jev 67 (Jev 77% sure) · pushes: cloud
Open-weights Mixture-of-Experts model with efficient attention, offered on DeepSeek's API at very low pay-as-you-go prices.
For AIOS builders: Showed strong models could be served far cheaper and set off a round of API price cuts among Chinese providers.
AI raters: Claude 56 · ChatGPT 72 · Jev 67 (Jev 87% sure) · pushes: owner control
Autonomous coding agent in early access that plans and executes engineering tasks; 13.86% on SWE-bench vs 1.96% prior best.
For AIOS builders: Made the autonomous coding agent a product category and set the pattern of agents working in their own sandbox.
AI raters: Claude 55 · ChatGPT 70 · Jev 67 (Jev 43% sure) · pushes: platform hold, cloud
Lightweight open models in 2B and 7B sizes, built from Gemini research, for laptops, workstations and cloud.
For AIOS builders: Added a credible open-weight option for private or local AI setups.
AI raters: Claude 58 · ChatGPT 73 · Jev 66 (Jev 95% sure) · pushes: owner control, local
FLUX.1 text-to-image suite: [pro] by API, [dev] open weights for non-commercial use, [schnell] under Apache 2.0.
For AIOS builders: High-quality image generation that could run locally or by API, useful for marketing and content workflows.
AI raters: Claude 52 · ChatGPT 63 · Jev 63 (Jev 68% sure) · pushes: owner control, local
Users create Gems, custom versions of Gemini with their own instructions, on Advanced, Business and Enterprise.
For AIOS builders: Google's counterpart to GPTs: reusable assistants with stored rules for repeat tasks.
AI raters: Claude 52 · ChatGPT 63 · Jev 62 (Jev 69% sure) · pushes: platform hold, cloud
ChatGPT remembers details from past chats, with controls to view, delete or turn memory off; started as a small test.
For AIOS builders: Brought persistent memory into a mainstream assistant, a core piece of an AI business brain.
AI raters: Claude 62 · ChatGPT 70 · Jev 58 (Jev 43% sure) · pushes: platform hold, cloud
The AI Act entered into force, with obligations phased in: prohibitions first, general-purpose AI rules after 12 months.
For AIOS builders: Set the legal frame for AI use in EU businesses; builders began tracking risk levels, transparency and literacy duties.
AI raters: Claude 58 · ChatGPT 73 · Jev 61 (Jev 31% sure)
ChatGPT can search the web and answer with links to sources, starting with Plus and Team users.
For AIOS builders: Removed the need to paste current information in by hand; research tasks could use fresh data.
AI raters: Claude 53 · ChatGPT 68 · Jev 60 (Jev 53% sure) · pushes: platform hold, cloud
NotebookLM generates a two-host audio discussion from the documents in a notebook.
For AIOS builders: A new way to review and share knowledge; it drew wide attention to NotebookLM as a research tool.
AI raters: Claude 53 · ChatGPT 54 · Jev 55 (Jev 65% sure) · pushes: platform hold, cloud
Bard renamed Gemini, a new mobile app, and Gemini Advanced with Ultra 1.0 in the $19.99 Google One AI Premium plan.
For AIOS builders: Bundled Google's top model into a consumer plan, with Gemini in Gmail and Docs promised next.
AI raters: Claude 50 · ChatGPT 53 · Jev 58 (Jev 37% sure) · pushes: platform hold, cloud
Llama 2 pretrained and chat models released free for research and commercial use, available on Azure, AWS and Hugging Face.
For AIOS builders: First widely usable open model for business: teams could self-host or fine-tune a capable assistant with no per-token fee.
AI raters: Claude 84 · ChatGPT 85 · Jev 78 (Jev 60% sure) · pushes: owner control, local
Models return structured JSON calls to developer-defined functions; also 16K context and price cuts.
For AIOS builders: Became the basic pattern for connecting a model to your own tools, databases and APIs.
AI raters: Claude 78 · ChatGPT 89 · Jev 82 (Jev 88% sure) · pushes: cloud
gpt-3.5-turbo, the ChatGPT model, at $0.002 per 1K tokens (10x cheaper than earlier GPT-3.5), plus a Whisper speech-to-text API.
For AIOS builders: Made it affordable to build custom assistants and transcription into your own tools, scripts and workflows.
AI raters: Claude 68 · ChatGPT 88 · Jev 81 (Jev 36% sure) · pushes: cloud
Large multimodal model with much stronger reasoning and exam-level results; in ChatGPT Plus and via an API waitlist.
For AIOS builders: First model reliable enough for complex instructions, long documents and multi-step business tasks.
AI raters: Claude 72 · ChatGPT 87 · Jev 79 (Jev 73% sure) · pushes: platform hold, cloud
Custom GPTs with instructions, files and actions; GPT-4 Turbo with 128K context and lower prices; Assistants API.
For AIOS builders: Packaged rules, knowledge and tools into reusable assistants, an early form of an AI business setup.
AI raters: Claude 72 · ChatGPT 86 · Jev 78 (Jev 61% sure) · pushes: platform hold, cloud
Sparse Mixture-of-Experts model under Apache 2.0, matching or beating GPT-3.5 on most benchmarks, 6x faster than Llama 2 70B.
For AIOS builders: First open model at GPT-3.5 level that was cheap to run; made self-hosted business assistants credible.
AI raters: Claude 72 · ChatGPT 79 · Jev 68 (Jev 79% sure) · pushes: owner control, local
Managed access to models from AI21, Anthropic, Cohere, Meta, Stability AI and Amazon, including Llama 2 as a managed API.
For AIOS builders: Companies on AWS could use several model vendors inside existing cloud contracts, security and data controls.
AI raters: Claude 58 · ChatGPT 71 · Jev 74 (Jev 40% sure) · pushes: cloud
Context grew from 9K to 100K tokens, enough to read long reports or a short book in one prompt.
For AIOS builders: Made it possible to load entire manuals, contracts and knowledge files into a single conversation.
AI raters: Claude 60 · ChatGPT 80 · Jev 71 (Jev 56% sure) · pushes: cloud
7.3B model under Apache 2.0 that beats Llama 2 13B on all reported benchmarks.
For AIOS builders: A small, fully permissive model that runs on a laptop; a common base for local assistants and fine-tunes.
AI raters: Claude 70 · ChatGPT 77 · Jev 68 (Jev 75% sure) · pushes: owner control, local
Gemini Pro API with function calling and a free quota through the web-based Google AI Studio, plus Vertex AI.
For AIOS builders: Gave builders a free way to prototype with a frontier-class model and get an API key in minutes.
AI raters: Claude 53 · ChatGPT 70 · Jev 68 (Jev 76% sure) · pushes: platform hold, cloud
200K-token context, fewer hallucinations, system prompts for custom instructions and a beta for tool use.
For AIOS builders: System prompts made standing rules a first-class input, and tool use began on the Claude API.
AI raters: Claude 57 · ChatGPT 73 · Jev 68 (Jev 77% sure) · pushes: cloud
Initial plugin support let ChatGPT browse the web, run Python code and call third-party services.
For AIOS builders: First sign that a chat assistant could use tools and live data; the approach was later replaced by GPTs and apps.
AI raters: Claude 62 · ChatGPT 80 · Jev 67 (Jev 63% sure) · pushes: platform hold, cloud
First public Claude models, offered through a chat interface and an API, in two speed and cost tiers.
For AIOS builders: Added a second strong model vendor, known for long, careful writing and following instructions.
AI raters: Claude 54 · ChatGPT 65 · Jev 66 (Jev 92% sure) · pushes: cloud
Natively multimodal model family; a tuned Gemini Pro began powering Bard the same day.
For AIOS builders: Google entered the frontier model race, adding a third serious option for builders.
AI raters: Claude 53 · ChatGPT 64 · Jev 65 (Jev 63% sure) · pushes: platform hold, cloud
AI-powered Bing and Edge in preview: chat that searches the web, summarizes answers with citations and drafts content.
For AIOS builders: First mainstream assistant grounded in live web results, the pattern later used for research steps in AI setups.
AI raters: Claude 52 · ChatGPT 69 · Jev 64 (Jev 83% sure) · pushes: platform hold, cloud
AI assistant inside Word, Excel, Outlook, Teams and other Microsoft 365 apps, generally available to enterprises at $30 per user per month.
For AIOS builders: AI reached the office tools where most company documents and email live, with access to tenant data.
AI raters: Claude 54 · ChatGPT 73 · Jev 63 (Jev 52% sure) · pushes: platform hold, cloud
Chat-based coding assistant in VS Code and Visual Studio, included for all paid Copilot users at no extra cost.
For AIOS builders: Asking for code in plain language inside the editor became standard, lowering the bar for building automations.
AI raters: Claude 52 · ChatGPT 65 · Jev 62 (Jev 61% sure) · pushes: platform hold, cloud
Improved model with longer outputs and better coding, plus a public beta of the claude.ai chat app in the US and UK.
For AIOS builders: Gave non-developers direct access to Claude and its long-document work.
AI raters: Claude 52 · ChatGPT 62 · Jev 65 (Jev 80% sure) · pushes: platform hold, cloud
Users set preferences once and ChatGPT keeps them in mind for all future conversations.
For AIOS builders: The first built-in place for persistent rules about who you are and how answers should look.
AI raters: Claude 58 · ChatGPT 72 · Jev 60 (Jev 41% sure) · pushes: platform hold, cloud
LLaMA 7B, 13B, 33B and 65B language models released to researchers under a noncommercial license, granted case by case.
For AIOS builders: Started the open-weights wave: capable models that could run on own hardware, which led to local and private assistants.
AI raters: Claude 58 · ChatGPT 60 · Jev 62 (Jev 35% sure) · pushes: owner control, local
Plus and Enterprise users can talk to ChatGPT by voice and upload images to discuss them.
For AIOS builders: Let builders capture work by voice and feed photos and screenshots straight into the assistant.
AI raters: Claude 51 · ChatGPT 68 · Jev 56 (Jev 64% sure) · pushes: platform hold, cloud
A chat interface to a GPT-3.5 model tuned with human feedback, free to use in the browser.
For AIOS builders: Made chat the default way to work with AI and started the habit of handing writing, analysis and planning to a model.
AI raters: Claude 96 · ChatGPT 98 · Jev 95 (Jev 99% sure) · pushes: platform hold, cloud
Arete AIOS
Where we are heading
For a thousand years the wealthy managed resources: land, capital and the people who worked for them. Soon every business owner will manage an AI organization. Agents do most of the work, the AI takes over the routine, and the owner keeps the thinking and the decisions.
It takes means, but far less than a property portfolio. What it takes most is structure, and control. That is what Arete AIOS builds: the business brain lives on the company's own computer and server, a local AI handles private data, external models do the deep work, and control stays with the owner.
Our own synthesis, in line with Hal Varian's 2011 view that what rich people have today, middle-income people get within ten years, and Sam Altman's idea that we can each have a personal AI team: Hal Varian, Foreign Policy, 2011 ↗ · Sam Altman, The Intelligence Age, 2024 ↗
Borrowed from OpenAI's internal scale, reported by Bloomberg in July 2024 (Axios summary) ↗. OpenAI calls the levels a work in progress, and critics find the jumps arbitrary. We use them as a rough map, translated into what each level means for a business. Level 5 is our goal.
AI answers in plain language.
In a business: Drafts, summaries, answers to questions.
AI thinks a hard problem through before it answers.
In a business: Analysis, research reports, checking a plan.
AI works on its own for hours, with tools and a computer.
In a business: Delegated tasks: research, files, code, web forms, with your approval.
AI helps invent new products and methods.
In a business: AI proposes improvements to your offer and your processes; you decide. Not business-ready yet.
AI does the work of an organization.
In a business: This is where Arete AIOS is heading: you run an AI organization. Agents run the repeatable operations; you set direction, approve and keep the decisions.
Borrowed from Feng, McDonald and Zhang, Levels of Autonomy for AI Agents (2025) ↗.
You drive every step; the AI answers.
Chat, 2022 to 2023You and the AI work on the same document or code together.
Projects and artifacts, 2024The AI does the task and asks you when it is unsure.
Work and coding agents, 2025The AI prepares and acts; you approve before anything leaves the company.
Where an AIOS aims for routine work, 2026The AI runs on its own; you watch results and alerts.
Only for low-risk routines with a track recordThe target for a business brain is not full autonomy. It is approver for everything that touches customers, money or people, and observer only for low-risk routines that have earned it.
After checking the evidence we see one real fork and one question that decides how far an owner can go. The fork is who holds the business brain: the memory, the skills, the agent definitions and the keys. The question is where the thinking runs: on rented cloud models or on the owner's own computers. That gives three tracks, and all of them will exist side by side.
Today the default path leads to the platforms. The standards that connect tools are opening up, but memory and hosted agents are where the platforms build their walls. A business stays in control only by choice: its brain kept as its own files, and agents it can move.
Track 1 · the platform holds the brain
Your memory, skills, agents and keys live inside ChatGPT, Claude, Gemini or Microsoft. Easiest to start, hardest to leave. This is where most businesses will go by default.
Who: OpenAI, Anthropic, Google and Microsoft; Meta and xAI want in. Specialist platforms such as Lovable, Manus or Genspark do the same for one job.
Watch for: memory you cannot export as files, agents that exist only in a vendor console, hourly fees for hosted agents.
Track 2 · the owner holds the brain
Arete AIOS today
Rules, memory and skills are the business's own files on its own computer or server, and its agents run there. The best cloud models are rented and can be swapped. You own the data and the routing; you rent the compute.
Who: Open standards (MCP, Agent Skills, AGENTS.md), open-source agents such as Hermes Agent and OpenClaw, and every vendor that supports them.
Watch for: how many tools adopt the open standards, real memory export, EU switching rights from 12 January 2027.
Track 3 · the owner holds brain and thinking
Arete AIOS Core, the goal
A computer in the office runs open models for everyday and private work. Cloud models get only the hard tasks, with personal data removed first. Medical, legal and client data stay in the building.
Who: Open-weight labs (Meta, Qwen, DeepSeek, Mistral, Google Gemma, OpenAI gpt-oss), local runtimes and desk computers with 128 GB of memory.
Watch for: how many months open models trail the frontier, what a small-office computer can run, and how much personal data local redaction still lets through.
Every rated launch, tagged by our three AI raters: does it help a business keep and move its own AI context (owner control) or keep it inside one platform (platform hold)? Does it move work toward the owner's computers (local) or a vendor's cloud? Neutral launches are left out; periods with few launches swing more.
Owner control Platform hold
Local Cloud
What to do with this