September was the busiest month for AI releases I can remember, and most of the coverage treated it as a scoreboard: who shipped the biggest model, whose benchmark went up. That is the least useful way to read it. If you build on these systems, three things happened that matter more than any single launch.
The short version
- The frontier got cheaper, not just bigger. Anthropic's Opus 5.5 claims near-Fable performance at 40% lower cost than Opus 5; Sonnet 5.5 followed six days later. Model upgrades now arrive quarterly and each one is an API migration.
- Agents became the product. OpenAI's Dots run always-on with their own cloud computer and browser. Meta's Muse reached macOS. Google shipped a production agent kit for Kotlin. The design question moved from "can it do the task" to "what is it allowed to do, and who approves."
- The labs are measuring their own automation. Anthropic says Claude now leads about 26% of its model R&D. Whatever you think of the number, the bottleneck it implies — review and specification, not generation — is coming for every team.
The month in one list
Anthropic ships Claude Fable 5.1 and Mythos 5.1, the second generation of its Mythos-class tier. Mythos stays restricted to approved organizations; Fable carries extra safeguards for general availability.
Anthropic adds automatic permission policies to Managed Agents and a CLI for connecting to live agent sessions.
Anthropic introduces on-demand conversation compaction (beta) and a thinking-mismatch control in the API.
Anthropic discloses that Claude leads 26% of its model R&D, that about 90% of R&D happens "in collaboration" with Claude, and that roughly 30,000 agents were doing research and engineering work as of August.
Meta expands its Muse agent to macOS, with direct access to files and apps behind user opt-in and approval for sensitive actions.
Google releases Agent Development Kit for Kotlin 1.0 at feature parity with the Python and Java kits, including on-device and hybrid inference on Android.
Anthropic launches Claude Opus 5.5: 1M-token context, 128k max output, always-on adaptive thinking, and an inline-tools beta that lets tool definitions ride inside a mid-conversation system message.
Anthropic moves cache diagnostics out of beta; every response now carries a
diagnosticsfield.OpenAI makes GPT‑6 Astra the stable ChatGPT release.
Anthropic launches Claude Sonnet 5.5 on its API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Thinking can no longer be disabled; forced tool use now returns a 400.
OpenAI DevDay: more than 20 products for about 2,500 attendees, headlined by Dots — always-on agents powered by GPT‑6 Astra, each with its own cloud computer and browser and connections to 4,000+ apps.
Anthropic deprecates Claude Sonnet 4.5, retirement set for November 30, 2026.
1. The frontier got cheaper, not just bigger
The headline of Opus 5.5 was not a capability jump. Anthropic's own framing was that it performs at Claude Fable 5.1's level "on most work" while costing 40% less to run than Opus 5 at default settings, with early testers reporting clearer writing. Six days later Sonnet 5.5 arrived as the faster, cheaper complement, and two days after that Sonnet 4.5 was deprecated with a retirement date eight weeks out.
Read those three dates together and the pattern is obvious: the model tier you build on now has a half-life of about a quarter. That is not a complaint — cheaper intelligence is the whole point — but it changes what "we use model X" means. It is a dependency with a release cadence, and it needs the same treatment as any other dependency.
Two details from the Sonnet 5.5 release notes make this concrete. Thinking can no longer be turned off, and a request that forces a specific tool call now returns a 400 instead of a result. Neither is dramatic. Both will break a pipeline that was written against the previous model and upgraded by changing a string. A model upgrade is an API migration, and the only safe way through one is an evaluation suite that runs on your own tasks — not a benchmark table.
On the other side of the fence, OpenAI's GPT‑6 Astra became the stable ChatGPT model on September 25, and Google's DeepMind leadership has said Gemini 4 should land "much earlier" than year end. If you are choosing a model this quarter, you are choosing it for a quarter.
2. Agents became the product
For two years "agent" mostly meant a feature inside a chat window. September is when the major platforms shipped agents as the thing you buy.
OpenAI's Dots are the clearest example. Each one runs always-on, independent of any particular device, with its own cloud computer and browser, and can reach you across ChatGPT, text, email, and Slack. OpenAI said it envisions "teams of Dots working together on your behalf," and included the first Dot in its Pro and Business Premium plans at no extra cost. Whatever the product turns out to be in practice, the architecture is explicit: the agent has a persistent environment and acts in it between your messages.
Meta's Muse landing on macOS makes a similar move at the operating-system level, with direct access to files and applications — gated behind opt-in and an approval step for sensitive actions. Google's ADK for Kotlin 1.0 is the developer-side version of the same shift: production tooling for building agents into Android and JVM apps, including on-device inference, at parity with the Python kit. And on September 10, Anthropic quietly added automatic permission policies to Managed Agents — a way to say, ahead of time, which actions an agent may take without asking.
Here is what I take from all of that. When an agent only produced text, the hard problem was quality. When an agent has a computer, a browser, and standing access to your tools, the hard problem is authority: what it may do, with what, on whose behalf, and who signs off when the action is hard to undo. Every one of these launches shipped with an answer to that question — opt-in, approval prompts, permission policies — because the companies building them know it is the question that decides whether the product survives contact with real users.
If you build agentic features, that is the design surface to invest in now. A permission policy is a first-class artifact, the same as a prompt or a tool schema: it should be written down, versioned, reviewed, and tested. I wrote about how I structure those decisions in Human-in-the-loop is a design decision, not a disclaimer.
3. The labs are measuring their own automation
The quietest item on the list is the one I keep coming back to. On September 18 Anthropic published figures on how much of its own research Claude performs: Claude leads 26% of model R&D, up from none in February; about 90% of R&D is done in collaboration with the model; and roughly 30,000 agents were doing research and engineering work as of August. The company was careful to add that the model is "not yet working completely autonomously" and that most tasks still run under human supervision.
Set aside the question of whether 26% is the right number or what "leads" means in that sentence. The practical signal is in the shape of the work. If a frontier lab's engineering is now one part specification, one part review, and a very large middle that the model fills in, then that ratio is where the rest of us are heading, just later and with smaller numbers. The scarce skills become writing a task precisely enough that an agent can take it end to end, and reviewing the result quickly enough that thirty agents don't pile up behind one person.
I find that a more useful frame for "what should I learn" than any list of tools. Specification and review are the durable parts of the job.
Smaller changes worth knowing
- Conversation compaction is now an API primitive in beta. Long-running agents accumulate state that is not worth its weight in tokens; being able to compact on demand — and control what survives — is the first-class version of a hack most of us already wrote. More on what to preserve in Context engineering.
- Cache diagnostics are out of beta, and every response carries them. Cost visibility per request stops being a side project.
- Inline tool definitions in a mid-conversation system message mean tools can be introduced when they become relevant, rather than all up front. That is good news for the tool-sprawl problem I describe in MCP, plainly.
- Safety tiering is now visible in the product line. Mythos-class models with the full capability set go to approved organizations; the Fable variants ship with additional safeguards for everyone else. Expect that to become normal.
What I would do this quarter
- Re-run your evals against the 5.5-class models. Not a benchmark — your golden set, on your tasks, with your tools. If you do not have one, that is the first job.
- Pin model versions and read the migration notes before changing them. "Thinking cannot be disabled" and "forced tool use returns 400" are the kind of lines that turn a one-character upgrade into an incident.
- Write a permission policy for every agent you run, even a prototype. Which actions are free, which need approval, which are forbidden. Put it in the repo next to the prompts.
- Instrument cost per outcome, now that the diagnostics are free. The 40% price cut on the frontier only shows up on your bill if you know what you are paying for.
That is September. The models will be cheaper again by the time I write the October note; the questions about authority and review will not have moved. Those are the ones worth your time.