Skip to main content

What's New in AI (2026 Edition)

info

This content was developed with AI assistance and is regularly reviewed for accuracy.

Snapshot: November 2026. Model names, prices, and product features change monthly; this page is the maintained chronology for the whole site, and its lineup table and timeline are refreshed as releases land (last pass: November 12, 2026).

In short: every major lab replaced its lineup during 2026. The frontier in November is Claude Fable 5.2 (with Opus 5.5, Sonnet 5.5, and Haiku 5.5 beneath it), OpenAI's GPT-6 family, Google's Gemini 4 Argon and Gemini 4 Carbon, and Meta's first frontier model, Muse Wonder. Reasoning is on by default and controlled by an effort setting rather than a "think step by step" phrase; 1M-token context is standard across the Claude family; memory, connectors, and web search reach even free tiers; and the agent is no longer a product you buy but a mode inside the $20 app you already pay for ("Claude Cowork is now just Claude," ChatGPT Work, Dots). Free tiers were downgraded along the way, Sora is gone, and the Model Context Protocol now lives under the Linux Foundation with a July specification that retired much of what 2025 tutorials taught. Facts marked "reported" rest on press or aggregator coverage rather than a vendor page.

Jump to a month: Jan · Feb · Mar · Apr · May · Jun · Jul · Aug · Sep · Oct · Nov (latest). Coming back after a year away? Skip to what this means if you last used AI in 2024.

The current lineup (November 2026)​

This table is the one place on AI Maniacs that lists model names, free-tier models, context windows, and consumer prices. Other pages link here rather than repeating it. Anthropic figures come from Anthropic's own pages; everything else marked "reported" comes from press or aggregators dated within the last two weeks and should be re-checked before you quote it.

LabFlagship (Nov 2026)What the free tier getsContext windowConsumer price
Anthropic (Claude)Claude Fable 5.2 (Nov 9), with Opus 5.5, Sonnet 5.5, and Haiku 5.5 beneath itSonnet and Haiku, plus web search, memory, connectors, voice, file creation, and up to five projects1M tokens on all four models (about 555,000 words), 128K max output, knowledge cutoff June 2026 (not yet stated for Fable 5.2)Pro $20/month ($17 billed annually); Max $100; Ultra $200 (the renamed Max 20x)
OpenAI (ChatGPT)GPT-6 Astra (Sep 3), with GPT-6.1 Sol, GPT-6 Sol, GPT-6.1 Terra (Oct 26), and GPT-6 LunaGPT-6 Luna, with ads on Free and Go since February 9, in 89 countries since October 21 (reported)1M tokens reported; not confirmed against OpenAI's pagesGo $8; Plus $20; Pro $100, $200, or $500 (reported)
Google (Gemini)Gemini 4 Argon (announced Sep 30, broadly available Oct 13), above Gemini 3.1 Pro and 3.8 Flash; Gemini 4 Carbon (announced Oct 28, in the paid API and for AI Ultra since Nov 10, reported)Flash-Lite only since Oct 9 (reported); Argon needs AI Pro or Ultra, Carbon needs Ultra (reported)Argon is quoted with a 1M-token output limit; its input window was not confirmedAI Plus $4.99; AI Pro $19.99; AI Ultra $99.99 or $199.99 (reported)
Meta (Muse)Muse Wonder (Oct 12); Muse Wonder Full only for Muse Code subscribers; the small, cheap Muse Ember (Nov 9, reported) for high-volume agent stepsThe Meta AI assistant and the Muse app's free tier run on the lighter Muse Spark (reported)About 1M tokens (single source, reported)Muse app Power $20, Maximum $100, US only (reported)
xAI (Grok)Grok 4.7 (Sep 21, reported)Not researched for this snapshot500K tokens (reported)Not researched for this snapshot
Microsoft (Copilot)Runs on partner and in-house MAI models; the current consumer model was not confirmedA free Copilot in Windows, Edge, and the webNot publishedMicrosoft 365 Premium about $19.99/month after Copilot Pro ended Aug 1 (reported)

Three notes on the numbers. First, word counts: Anthropic states that 1M tokens is roughly 555,000 words (or 2.5 million characters) on the tokenizer introduced with Claude Opus 4.7, and that models before it fit about 750,000 words; an earlier version of this page used the older figure, and 200K tokens is roughly 150,000 words. Second, Google's "1M" for Gemini 4 Argon is an output limit in Google's own announcement, not a confirmed input window, so do not compare it directly with Claude's 1M input context. Third, API prices for the Claude family per million input and output tokens are Fable 5.2 $8/$40 (down from Fable 5.1's $10/$50), Opus 5.5 $4/$20, Sonnet 5.5 $2/$10, and Haiku 5.5 $0.10/$0.50 for prompts up to 100K tokens ($0.50/$2.50 above that); a 900K-token request is billed at the same per-token rate as a 9K one on every model except Haiku's price step. Independent scoring is less tidy than vendor tables: Artificial Analysis placed Opus 5.5 first and Sonnet 5.5 second on its Intelligence Index around October 10, while a reviewer noted that Argon's published benchmark scores reflect its High mode rather than the subscriber default (reported). Treat every vendor benchmark as marketing until a third party reproduces it.

Retired in 2026​

The turnover was as much about removals as launches. Anthropic retired Claude 3.7 Sonnet and 3.5 Haiku (February 19), Claude 3 Haiku (April 20), Claude Opus 4 and Sonnet 4 (June 15), and Opus 4.1 (August 5), and deprecated Sonnet 4.5 on September 30 ahead of a November 30 retirement; no Claude 3 model has been usable on the API since April. OpenAI shut down the Sora video app (announced March 24, offline April 26, API removed September 24) and the Atlas browser (announced July 9, decommissioned August 9), and in August stopped individuals creating or publishing Custom GPTs, replacing them with Workspace Agents for Business and Enterprise plans. Google closed Project Mariner on May 4 (reported). Microsoft ended Copilot Pro on August 1 (reported). Anthropic retired the per-user Claude Code Slack bot on October 5 in favour of Claude Tag. If a 2024 or 2025 tutorial names any of these, read it as history.

2026 month by month​

Every entry carries its date. Items marked "reported" rest on press or aggregator coverage; the rest are confirmed against a vendor or standards-body page.

January​

The year opened with agents arriving on the desktop and the first signs that free tiers would be monetised.

  • Jan 9: Anthropic published "Demystifying evals for AI agents," which set the vocabulary (task, trial, grader, pass@k versus pass^k) that our Evals for Agents page uses.
  • Jan 12: Claude Cowork launched as a research preview: a desktop agent that works inside a folder you grant it, first for Max subscribers and then, reportedly, for Pro (reported).
  • Jan 16: OpenAI said it would test ads for logged-in Free and Go users in the US (reported).
  • January: Google introduced the Universal Commerce Protocol for agent-driven checkout (reported); Duolingo made its AI grammar explanations free to all users (reported); the observability vendor Langfuse was acquired by ClickHouse (reported).

February​

February brought the ads, the first model retirements, and a demonstration of what parallel agents can build.

  • Feb 5: Anthropic described sixteen parallel Claude instances building a roughly 100,000-line C compiler over two weeks for about $20,000 in API cost, coordinated by git lock files with no orchestrator agent. The same day Kuaishou launched Kling 3.0 globally, followed about three days later by ByteDance's Seedance 2.0 (both reported).
  • Feb 9: Ads began appearing for ChatGPT Free and Go users in the US; paid tiers stay ad-free (reported).
  • Feb 12: Anthropic said Claude Code had passed a $2.5 billion annualised run rate, more than doubling since January 1 (company statement, via press).
  • Feb 19: Claude 3.7 Sonnet and Claude 3.5 Haiku were retired from the API, the first of four 2026 retirement waves.
  • Feb 24–26: Cowork admins got private plugin marketplaces (Feb 24); Google released Nano Banana 2, its Flash-speed image model, free and rate-limited inside the Gemini app (Feb 26, reported). Also in February: GitHub launched Agent HQ to route tasks across Copilot, Claude, Codex, and custom agents (reported), and Cowork gained Google Drive, Gmail, Docusign, and FactSet connectors for enterprises (reported).

March​

March was the month a flagship consumer product disappeared.

  • Mar 9: a16z's sixth Top 100 consumer AI apps list put ChatGPT at roughly 900 million weekly users, named Claude and Gemini the fastest-growing paid subscriptions, and admitted CapCut, Canva, Notion, and Grammarly as "AI-enhanced" apps (reported).
  • Mar 17–20: Midjourney V8 entered alpha, the biggest version jump of its year (Mar 17); NotebookLM added Cinematic Video Overviews and editable slide revisions, turning the study notebook into a content factory (Mar 20).
  • Mar 24: OpenAI announced the shutdown of Sora, its video app, less than six months after launch (reported). The same day OpenAI scaled back in-chat Instant Checkout in favour of product discovery (reported), and Anthropic published its harness-design results: a full planner, generator, and evaluator harness produced a playable game in six hours and $200 where a solo twenty-minute run did not.
  • Mar 25–26: Claude Code shipped auto mode, a classifier that replaces most permission prompts; Anthropic disclosed that users had been approving 93% of manual prompts (Mar 25). Suno v5.5 added voice capture and custom models (Mar 26).
  • March: Mistral Small 4 shipped under Apache 2.0, Perplexity's Comet browser reached iOS, the A2A protocol reached version 1.0, and by the first week of the month more than twenty-five agent products supported the Agent Skills open standard (all reported).

April​

April combined the year's most-cited safety incident with MCP's first governance milestones.

  • Apr 2–3: The first MCP Dev Summit under the Linux Foundation was held in New York (reported); by April the Agentic AI Foundation reportedly had more than 170 member organisations.
  • Apr 8–9: Anthropic published its Managed Agents architecture, decoupling the "brain" that calls the model from the "hands" that run code (Apr 8); Claude Cowork reached general availability with enterprise deployment controls (Apr 9; the GA date is reported, the enterprise features are on Anthropic's page).
  • Apr 13: The Stanford AI Index 2026 reported that capability "is not plateauing," that the US–China model gap had effectively closed, and that hallucination rates on a knowledge-versus-belief test across twenty-six top models ranged from 22% to 94% (figures via secondary summaries).
  • Apr 20–21: Claude 3 Haiku was retired (Apr 20); Google launched API-grade Deep Research and Deep Research Max on Gemini 3.1 Pro (Apr 21).
  • Late April: A Cursor coding agent running Claude Opus 4.6 deleted PocketOS's production database and backups in about nine seconds, using an over-privileged key it found in an unrelated file; no prompt injection was involved (reported Apr 28). Also in April: GPT-5.5 and GPT Image 2 shipped (both reported), Mistral Medium 3.5 arrived on April 30 (reported), and Midjourney V8.1 entered alpha on April 14.

May​

May was Google's month, and the month Microsoft put a price on agent identity.

  • May 1: Microsoft Agent 365 became generally available at $15 per user per month, licensed per user rather than per agent.
  • May 4–5: Google shut down Project Mariner, its standalone browsing agent, folding the technology into Gemini and Chrome auto-browse (May 4, reported). Microsoft's Work Trend Index 2026 found that the skills workers cite most are quality control of AI output (50%) and critical thinking (46%), and that active agents in Microsoft 365 grew fifteen-fold year over year (May 5).
  • May 19–20: Google I/O announced Gemini Spark, an always-on personal agent running in Google Cloud for Ultra subscribers; "Gemini Intelligence," an agentic layer across Android; Gemini 3.5 Flash; Antigravity 2.0; and Universal Cart, which went live with major US retailers at Google Marketing Live on May 20 (all reported).
  • May 25: Anthropic's "How we contain Claude" post set out the containment vocabulary (blast radius, approval fatigue, egress controls) and conceded that "model-layer protection can never be 100% effective."
  • May: Granola shipped Briefs, which prepare you for a meeting from past notes, Gmail, and the web (reported); SAP's investment reportedly valued n8n at $5.2 billion; Meta's Ray-Ban Display opened to third-party apps (May 15, reported); Claude Opus 4.8 shipped around May 28 (date inferred from Anthropic's retirement schedule).

June​

June brought Apple's reset, the Mythos line, and the retirement of the Claude 4 generation.

  • Jun 2–3: Project Glasswing expanded to about 150 organisations in more than fifteen countries (Jun 2); the NSA and CISA jointly published MCP security design guidance (Jun 2, reported); Microsoft Build refreshed Copilot and previewed a GitHub Copilot desktop app (reported); Cognition rebranded Windsurf as Devin Desktop (Jun 2, reported).
  • Jun 8: At WWDC Apple unveiled a rebuilt Siri with Google's Gemini handling complex requests, shipping with iOS 27 on iPhone 15 Pro and newer (reported).
  • Jun 9–12: Anthropic released Claude Fable 5 and, to vetted partners only, Claude Mythos 5 (Jun 9). Access to both was suspended on June 12 pending US government export-control approval and restored for US organisations on July 1. Lovable reportedly passed $500 million in annualised revenue the same week.
  • Jun 15–17: Claude Opus 4 and Sonnet 4 were retired (Jun 15); ChatGPT rebuilt scheduled tasks for Plus, Pro, Business, and Enterprise, replacing Pulse (Jun 17).
  • Late June: Claude Sonnet 5 shipped around June 30 (date inferred from the retirement schedule); Microsoft AI reportedly released seven in-house models including MAI-Image-2.5 (reported, unverified); and the term "loop engineering," credited to Addy Osmani, was in circulation by the end of the month (reported).

July​

July rewrote the plumbing: a new OpenAI flagship, the end of the Atlas browser, and the largest MCP revision since launch.

  • Jul 7–9: Cowork reached web and mobile beta (Jul 7, reported). OpenAI released GPT-5.6, launched ChatGPT Work as its agent mode on paid plans, and announced that the Atlas browser would shut down on August 9 (Jul 9; Work from OpenAI's release notes, the rest reported).
  • Jul 23–24: ChatGPT Health opened to US users, connecting Apple Health and medical records (Jul 23). Anthropic released Claude Opus 5 at $5/$25 per million tokens with an adjustable effort setting (Jul 24). Midjourney V8.2 became the default the same day.
  • Jul 28: MCP specification revision 2026-07-28: a stateless core (no handshake, no session header), Multi Round-Trip Requests replacing server-initiated elicitation, hardened OAuth with Client ID Metadata Documents, a formal extensions framework including Tasks, and a twelve-month deprecation policy. The legacy HTTP+SSE transport, Roots, and Sampling were deprecated. Tier 1 SDKs were seeing close to half a billion downloads a month. Our Build and Secure an MCP Server page is written to this revision.
  • July: Moonshot released the Kimi K3 open-weight model (reported; weight-release dates conflict across sources).

August​

August was a month of consolidation: browsers into apps, memory across products, and Custom GPTs into Workspace Agents.

  • Aug 1–5: Microsoft ended Copilot Pro and moved consumer Copilot features into Microsoft 365 Premium at about $19.99 a month; a free Copilot remains in Windows, Edge, and the web (Aug 1, reported). Claude Opus 4.1 was retired (Aug 5).
  • Aug 9–12: ChatGPT Atlas was decommissioned, with browsing moving into the ChatGPT desktop app and a Chrome extension (Aug 9, reported). The Claude in Chrome side panel became a full Cowork session, saved to history and resumable from desktop, web, and mobile (Aug 12, reported).
  • Aug 22–26: The MCP maintainers published their post-spec roadmap (Aug 22). Anthropic merged the memory systems of chat and Cowork so Claude remembers across both (Aug 25, reported), and announced a built-in browser in the Cowork side panel (Aug 26). McKinsey's State of AI 2026 found 40% of billion-dollar companies scaling agents while smaller firms stayed flat at 22% (Aug 25, reported).
  • August: Individuals could no longer create or publish Custom GPTs; OpenAI replaced them with Workspace Agents for Business and Enterprise plans. Zenity's zero-click "PleaseFix" demonstrations at Black Hat hijacked Claude in Chrome, Gemini in Chrome, Comet, Atlas, and Copilot Edge (reported).
  • August: Alibaba's Qwen 3.8 shipped with a 27B dense model under Apache 2.0 (reported); SemiAnalysis reported Google had shelved Gemini 3.5 Pro (reported, contested by Google at the time); Google's A2A protocol reportedly joined the Agentic AI Foundation (Aug 17, single source).

September​

September was the densest release month of the year: new flagships from four labs in four weeks.

  • Sep 1–3: Claude Fable 5.1 and Mythos 5.1 (Sep 1; same $10/$50 price as Fable 5, cache reads cut to $0.25; Mythos 5.1 is the same model without Fable's cyber and biology safeguards, for verified organisations). Gemini 3.8 Flash (Sep 2, reported). GPT-6 Astra began a phased rollout as the first OpenAI model to hit its internal "Critical" cybersecurity threshold, with its most advanced capabilities restricted to vetted partners (Sep 3, reported).
  • Sep 8–16: Meta announced the Muse personal-agent app with free, $20, and $100 tiers, US only (Sep 8, reported); DeepSeek released V4.1-Flash weights (Sep 10, reported); iOS 27 shipped with the Gemini-based Siri (Sep 14, reported); Cowork and chat began merging into one Claude (Sep 16, reported).
  • Sep 21–22: xAI's Grok 4.7 at $2/$6 per million tokens with a 500K context (Sep 21, reported). Claude Opus 5.5 (Sep 22) at $4/$20: 40% cheaper to run than Opus 5, more than 30% faster, thinking that can no longer be disabled, and effort levels from Low to Max. GPT-6 Sol and GPT-6 Luna (Sep 22).
  • Sep 24–28: Sora's Videos API was removed (Sep 24, reported); Microsoft relaunched Copilot as a "super app" with Home, Code, and Autopilot tabs (Sep 25, reported); Claude Sonnet 5.5 (Sep 28) at $2/$10, more than 30% faster than Sonnet 5 at up to 30% less per task.
  • Sep 29–30: OpenAI DevDay: GPT-6.1 Sol, Dots (persistent agents for Pro tiers), ChatGPT Space, computer use in the Agents API, and a $500-a-month Pro 500 tier (the tier is reported by a single source). Google announced Gemini 4 Argon with a 1M-token output limit, first to trusted cyber defenders (Sep 30). Anthropic deprecated Claude Sonnet 4.5 (Sep 30). Court trackers counted 2,046 cases involving fabricated AI citations as of September 21, up from about 200 a year earlier (reported).

October​

October finished the lineup turnover and brought agents to the free and enterprise ends of the market.

  • Oct 2–6: Anthropic committed $100 million to train 10,000 "Frontier Deployed Engineers" (Oct 2); the per-user Claude Code Slack bot was retired in favour of Claude Tag (Oct 5); Stack Overflow's 2026 survey found 66% of developers using coding assistants or agents, with Claude Code used by 66% of agent users (Oct 6, reported); Anthropic's Cyber Verification Program expanded to cover Opus 5.5 and Sonnet 5.5 (Oct 6); Claude reportedly arrived inside Google Docs, Sheets, and Slides (Oct 6, reported, unverified). Mistral opened Large 4, a roughly 1-trillion-parameter model, as an API preview, with open weights promised by the end of October and no licence named yet (Oct 6, reported).
  • Oct 7: Claude Haiku 5.5 at $0.10/$0.50 per million tokens for prompts up to 100K, the first Haiku with an effort setting and about 75% cheaper to run than Haiku 4.5. GPT-6 rolled out to regular ChatGPT, with GPT-6 Sol for paid plans and GPT-6 Luna for Free and Go, alongside "Intelligent UI" that answers with charts, forms, and buttons. Claude Max plans gained $100 or $200 a month in API credits, rolling out that week.
  • Oct 8–9: Google announced a "Gemini agent for work" spanning Gmail, Drive, Docs, Slides, Sheets, Chat, and Calendar, able to route work to Gemini or Claude models, in enterprise preview (Oct 8, reported); Claude Docs, Slides, and Design left beta while Dashboards and Motion entered it, and Anthropic updated its usage policy (Oct 8); Gemini's free tier was limited to Flash-Lite (Oct 9, reported).
  • Oct 12–13: Meta released Muse Wonder and Muse Wonder Full, its first frontier models, with Wonder at a reported $4/$20 per million tokens (Oct 12). Gemini 4 Argon became broadly available to Google AI Pro and Ultra subscribers and in the paid API at an introductory $2/$10 until November 30, then $4/$20 (Oct 13).
  • Oct 14–15: Anthropic opened OpenTelemetry trace export for agents in public beta (Oct 14); Claude Code's Code Review reached general availability for Team and Enterprise (Oct 15); Anthropic published an update on independent evaluators (Oct 15). Meta released the weights of Muse Spark Open (Oct 15): an 816B-parameter mixture-of-experts model with about 42B parameters active per token, a 1M-token context window, and text, image, video, and audio input. It is free to download under Meta's Muse Model License, which limits commercial use. Meta calls it the most capable US open-weight model; on the Artificial Analysis Intelligence Index it scores 39, behind several Chinese open models (reported).
  • Oct 16–17: GPT-6 Pro in regular ChatGPT Chat gained parallel agents: it can split one task across several agents working at the same time, and you can follow each agent's progress and open its finished work to review it. It also got faster responses and better tool and plugin use, rolling out gradually to Pro subscribers. Plus does not include GPT-6 Pro. DeepSeek updated its open-weight model as V4.1-Flash-1017 under the MIT licence, with shorter outputs and faster agent tasks, serving it in its apps and API at unchanged prices (Oct 17, reported). As of this snapshot Anthropic's pricing page reads "Claude Cowork is now just Claude," rolling out to Pro and Max, and Claude Code documents five parallelism modes (subagents, background agents, agent teams, dynamic workflows, projects) plus the /loop, /goal, and Routines primitives covered in Loops and Harnesses.
  • Oct 21: Anthropic opened a waitlist for Claude Duet, a research-preview app in which Claude, built around Opus 5.5, works as a proactive agent: it sends a morning brief, checks your connected apps every few hours, takes on tasks you assign, and shows you exactly what it is about to do and waits for approval before sending messages, making payments, signing documents, or deleting anything. It runs on its own cloud virtual machine, so your devices can be off, and comes as a standalone iOS and Android app and on claude.ai. It starts with US Pro and Max subscribers and draws on the same usage limits as the rest of Claude. The same day, Claude Security added continuous repository scanning and fix tracking in public beta for Claude Enterprise, and OpenAI expanded ChatGPT ads to 26 more countries, 89 in all, still only on Free and Go (reported).
  • Oct 22: Anthropic donated the Agent Skills standard (the SKILL.md format, its name, the agentskills.io site, and its GitHub organisation) to the Linux Foundation's Agentic AI Foundation, putting it under the same neutral governance as MCP and AGENTS.md; Anthropic says more than 70 agent platforms, including ChatGPT and Cursor, have adopted it. OpenAI's models, including the GPT-6 family, began a private preview on Google Cloud for selected enterprise customers, billed against existing Google Cloud commitments (reported).
  • Oct 26: OpenAI released GPT-6.1 Terra, a faster, cheaper mid-tier model for computer use, visual reasoning, writing, and coding, replacing GPT-5.6 Terra. It is available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and in the API as gpt-6.1-terra at $1 per million input tokens and $5 per million output, half the price of GPT-6.1 Sol. OpenAI says it completes agent tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol and makes fewer factual errors than GPT-5.6 Terra (benchmark figures are OpenAI's own). OpenAI also rates it at the Critical cybersecurity threshold under its Preparedness Framework, with the same protections it built for GPT-6 Astra. The same day, Anthropic filed an S-1 with the SEC for a proposed initial public offering.
  • Oct 28: Google announced Gemini 4 Carbon, which it calls its strongest model for cybersecurity and long-running tasks. It is not in the Gemini app or the public API yet: Google is opening it first to a subset of cyber defenders and then to more trusted partners through its Fairwind Program, and says broader access for developers, businesses, and consumers will follow with no date given. Google says it will be priced like Gemini 4 Argon, at an introductory $2/$10 per million tokens, later $4/$20.
  • Oct 29: Mistral published the open weights of Mistral Large 4 on Hugging Face, three weeks after its API preview. It is a mixture-of-experts model of about 1 trillion parameters (49 billion active per token), free to download under the Mistral Ecosystem Open-Weight License, a modified MIT licence that allows commercial use but asks companies offering it as a hosted service to get a separate commercial licence from Mistral. Running it yourself takes data-centre hardware, not a laptop.

November​

November, to the 12th, brought Anthropic's next flagship and Google's Carbon to paying customers.

  • Nov 4–5: Moonshot published the open weights of Kimi K3.1 Pro on Hugging Face, two weeks after its API launch on October 23: a 2.8-trillion-parameter mixture-of-experts model with 104 billion active parameters and a 1M-token context window, under the Kimi K3.1 Pro License, with cyber-related training data deliberately removed (dates and the training-data change reported). OpenAI rolled out ChatGPT Images 2.6, which can make animated GIFs, with two new image API models, GPT Image 2.6 Flare and Sunburst (Nov 4, reported). OpenAI announced GPT-6 Cyber, a security model offered in the API only to organisations approved for its Daybreak Red trusted-access tier, at $20 per million input tokens and $100 per million output at its short-context rate; its Codex availability and the Nov 5 date are reported.
  • Nov 9: Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, again one model in two safety configurations. Fable 5.2 is available on all Claude plans that include Fable and in the API as claude-fable-5-2 at $8 per million input tokens and $40 per million output, 20% cheaper than Fable 5.1, with a 1M-token context window, outputs up to 30% faster, and an optional fast mode at double the price. Anthropic reports higher scores than Opus 5.5 and Fable 5.1 on most of its published coding, agent, and science benchmarks, and about 55% fewer blocked requests than Fable 5.1. Mythos 5.2 remains limited to organisations in Anthropic's verification programs. Pro and standard Team seats get Fable included for a launch week, through November 16. The same day, Meta released Muse Ember, its smallest and cheapest Muse model, aimed at high-volume agent steps, in Muse Code and the Meta Model API (reported).
  • Nov 10: Google opened Gemini 4 Carbon to the paid Gemini API at the same introductory $2/$10 as Argon, rising to $4/$20 after November 30, and began rolling it out to Google AI Ultra subscribers in the Gemini app, AI Mode in Search, and Antigravity; AI Pro subscribers keep Argon (reported). Anthropic launched the roadshow for its proposed initial public offering.

What this means if you last used AI in 2024​

Five pieces of 2024 advice are now wrong for the flagship apps. "The AI cannot browse" is wrong: web search is on by default, even on Claude's free tier. "It forgets between chats" is wrong: memory now spans conversations, and on Claude it spans chat and agent sessions. "It only knows up to 2023" is wrong: the current Claude models have a June 2026 cutoff, a lag of months rather than years. "Turn on reasoning mode" is wrong: on Fable 5.2 and Opus 5.5 thinking cannot be turned off, and the control you have is effort. "Split your document into chunks" is wrong for 1M-token models, though fabrication still rises with context length, so very long inputs deserve more checking, not less.

The replacement skills are small but different. Decide what to hand off and which mode fits (fast, thinking, research, or agent), then choose an effort level instead of asking for step-by-step work. Supply files, connectors, and project context rather than pasting text. Decide what memory should store, and grant connector and folder permissions deliberately, because an agent that can read your email can also be told what to do by your email. Review what an agent did, not only what it wrote: Anthropic's own measure for Opus 5.5 is "fewer hard-to-reverse actions." And keep verifying citations; the court sanctions wave of 2026 fell mostly on professionals, not beginners.

Two expectations also need resetting. A free account in November 2026 gets the small model of each family (GPT-6 Luna, Claude Sonnet or Haiku, Gemini Flash-Lite) and, on ChatGPT, ads; it is fine for learning the habits but not for judging what the frontier can do. And "the agent" is not a separate purchase: it is a mode inside the $20 subscription, which is why Agents in the Apps You Already Pay For is the first practical page in our Agents track. The habits themselves are taught in How AI Assistants Work in 2026, Decide, Supply Context, and Protect.

Where the details live​

This page records what changed and when. The pages below hold the comparisons and recommendations, and each links back here for the chronology:

For the underlying concepts (tokens, context, effort, memory, modalities, and where models still fail), read How AI Assistants Work in 2026. For the open-weight side of the year (DeepSeek V4, Qwen 3.8, Gemma 4, Kimi K3, and running a model locally), see Open-Weight Models and Running One Locally. For the quarterly state of agents, see The 2026 Agent Landscape, and for the policy side, The Global AI Regulatory Landscape.

Next steps​