AI in Fall 2026: Gated Frontier Models, Computer Use and a Stateless MCP

Nov 13, 2026AI & data

A network of connected nodes above a timeline marked September, October and November

Between September 1 and mid-November 2026, Anthropic, OpenAI and Google all replaced their model lineups. Ten weeks, three new generations, and more than a dozen model names.

This post sorts that into the four trends that matter if you build with these models:

frontier releases are now gated;
price per unit of capability keeps falling;
computer use is becoming a standard product;
the agent plumbing has settled.

Dates and prices come from each vendor’s own announcements and release notes. Benchmark claims are the vendors’ own.

The releases

Anthropic: Claude 5.1, 5.5 and 5.2

Date	Model	Notes
Sep 1	Claude Fable 5.1 and Claude Mythos 5.1	Anthropic says these are the same model with different safeguards. Fable is generally available at $10/$50 per million input/output tokens. Mythos is for vetted organizations in trusted-access programs.
Sep 22	Claude Opus 5.5	The first of the 5.5 family. Anthropic says it “performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5”. $4/$20.
Sep 28	Claude Sonnet 5.5	$2/$10.
Oct 7	Claude Haiku 5.5	Anthropic says about 75% cheaper to run than Haiku 4.5.
Nov 9	Claude Fable 5.2 and Mythos 5.2	“Our most capable model for technical work”, with output up to 30% faster than 5.1 and the price cut to $8/$40.

OpenAI: GPT-6

Date	Model	Notes
Sep 3	GPT-6 Astra	Flagship, with limited organizational access at launch.
Sep 22	GPT-6 Sol and GPT-6 Luna	The everyday models: Sol at $2/$10, Luna at $0.10/$0.50.
Sep 29	GPT-6.1 Sol	OpenAI says near-Astra performance at a fifth of the price. Computer use also arrived in the Agents API, running in an OpenAI-hosted browser.
Oct 7	GPT-6 in ChatGPT	Launched with “Intelligent UI”: answers can mix text, graphics and interactive controls, and the model can start answering while still reasoning.
Oct 26	GPT-6.1 Terra	A fast model built for visual reasoning and computer use.
Oct 30	GPT-6 Rosalind	For trusted-access researchers.

Google: Gemini 4

Date	Model	Notes
Sep 30	Gemini 4 Argon announced	First released only to vetted cyber defenders through Google’s Fairwind Program.
Oct 13	Argon	Generally available in AI Studio, the API and the Gemini app. Introductory pricing of $2/$10, rising to $4/$20 after November 30.
Nov 10	Gemini 4 Carbon	Public after an October preview limited to Fairwind members. Google calls it its strongest model, with “frontier performance in cybersecurity”, and recommends it for long-horizon agentic work.
Nov 12	Gemini 4 Hydrogen	The fast, cheaper “daily driver for agentic workflows” at $1.50/$7.50. It also now powers Google Search.

Open-weight models

Xiaomi MiMo-V2.6 (Pro and Flash, September 22) is MIT-licensed. The Pro model has about a trillion total parameters, with 42B active.
Mistral Large 4 is a 1T-parameter mixture-of-experts model with about 52B active parameters. It had a public preview on October 6 and open weights followed. Check the license on the model page before you use it, because Mistral’s own pages describe it inconsistently.
Meta has published no new Llama weights in this period.

Trend 1: the most capable models ship gated first

The biggest change this fall is not a benchmark. Each of the three big labs now holds back its strongest model, or a variant of it, from general release, and the reason is the same: cybersecurity.

Anthropic ships the same model twice. Mythos has fewer restrictions and is limited to vetted organizations. Fable is generally available with stronger safeguards.
Google gave Argon, and later Carbon, to Fairwind members and to government pre-release testing before the public got them.
OpenAI launched Astra to a limited set of organizations, and Rosalind is trusted-access only.

For most developers this changes little: the generally available model is the one you can build on, and it is very good. It does mean “the best model” is now a credential question as well as a price question. Expect more security vendors to advertise access to the gated versions.

Trend 2: more capability per dollar

Prices did not fall as fast as in 2024–2025, but within each lineup they kept dropping:

Opus: Opus 5.5 costs $4/$20, down from $5/$25 for Opus 5, and Anthropic says it matches the previous flagship on most work.
Fable: Fable 5.2 is 20% cheaper than 5.1 and faster.
Small models: GPT-6 Luna and Haiku 5.5 now sit at around ten cents per million input tokens.
Gemini: watch the introductory pricing. Argon and Carbon are $2/$10 until November 30 and $4/$20 after that, so a cost estimate made this month will be half the real bill.

In practice, re-check which model each part of your app uses. A classification step that needed a mid-tier model a year ago probably runs on Haiku- or Luna-class models now. An agent that needed the flagship might run on Opus 5.5 or Sonnet 5.5.

Trend 3: computer use became a product

Models operating a computer through screenshots, clicks and typing were a demo in 2024. This fall the vendors started selling it as a standard feature:

OpenAI: GPT-6.1 Terra is built for it, and the Agents API runs computer use in a browser OpenAI hosts.
Anthropic: Claude’s apps and Claude Code drive browsers, both the user’s own Chrome and a built-in browser in the desktop app.
Google: Hydrogen is described as an agentic daily driver, and Carbon is aimed at long-horizon agent tasks.

The interesting engineering is now around the model: sandboxes, which sites an agent may act on, and what it may do there. Prompt injection from the pages an agent reads is still the main risk. The most useful pattern I have seen is the same as for human admins: least privilege, plus an explicit confirmation step before anything irreversible, like sending, paying or deleting.

Trend 4: the plumbing settled

The Model Context Protocol, the standard way to give a model tools and data, had its biggest revision on July 28, 2026. It was the first since the protocol moved to the Linux Foundation’s Agentic AI Foundation in December 2025:

It is stateless. The initialize handshake and the Mcp-Session-Id header are gone. Each request carries what the server needs, so an MCP server can sit behind an ordinary load balancer like any other HTTP API.
Roots, Sampling and Logging are deprecated, under a formal deprecation policy.
Clients have moved over. Through the fall, clients such as Claude Code made 2026-07-28 their default protocol version for remote connectors.

Combined with the computer-use push, this means the “agent stack” stopped changing every month. Building an MCP server today is a safe investment; in 2025 it was a bet.

What I would do with this

Re-run your evals on the new mid-tier models before paying for a flagship. The gap between Opus 5.5 or Sonnet 5.5 and the top model is small for most product work.
Budget for Gemini’s December price change if you adopted Argon or Carbon in October or November.
Move MCP servers to the stateless spec if you run any remotely. It removes session state and is much easier to deploy.
Treat computer use like a new employee with a browser. Give it a narrow scope and a test account, and log what it does.

Other posts this week go deeper on two of these threads: the Claude Code features that shipped alongside these models, and agent systems for investment research.

More from this week

This is part of a series of posts from November 8 to 14, 2026: