What happened
Three frontier models in two weeks, all measured against the same thing
On 1 September Anthropic released Claude Fable 5.1 and Mythos 5.1, with a 1M-token context window, 128K output tokens and built-in adaptive reasoning, evaluated on OSWorld 2.0. In the same period OpenAI shipped GPT-6 Astra, aimed at multi-step reasoning, browsing, autonomous computer use and technical incident resolution, scoring 88% on SRE-Bench. On 2 September Google made gemini-3.8-flash generally available, the most capable model in the Flash line, at the same pricing as 3.7.
Why it matters: all three are measured against the same thing, and it is not conversation. OSWorld and SRE-Bench test whether a model can drive a machine and close an incident without being walked through it. That is a change of category: until recently you picked a model by how it wrote or what a million tokens cost; now you pick it by whether it finishes the job. And 3.8 Flash arriving at 3.7 pricing matters more than it looks once the volume comes from an automated process rather than a person typing.
Anthropic — Claude Fable 5.1 and Mythos 5.1 · OpenAI — GPT-6 Astra · Gemini API changelog
Gemini now watches video the way you search, not the way you sit through it
Agentic Video Understanding went into production in September for Gemini Flash 3.7 and 3.5 Lite: instead of processing the whole track, the model navigates the timeline and requests transcripts, specific frames or audio on demand through the Interactions and GenerateContent APIs. The stated reduction reaches 88% of token consumption on long videos compared with static processing.
Why it matters: video was the format that cost kept out of reach. Warehouse cameras, recorded training sessions, meetings you need decisions out of. Processing a full hour to answer one narrow question never fit a small budget, and the alternative was trimming by hand before sending. Searching instead of reading everything is what makes that request viable.
Gemini API changelog · Google AI for Developers
MCP stopped being an adapter you write by hand
LangChain 1.4, released 3 September, introduced the langchain.mcp namespace and the MCPAdapter class: any MCP server, remote over HTTP or local via FastMCP, plugs in as a native LangChain and LangGraph tool with no glue code. LangSmith Cloud added distributed tracing with OpenTelemetry for multi-agent applications. On the protocol side, Microsoft and Cloudflare are consolidating stateless MCP 2.0 for connecting agents to databases and APIs without session management.
Why it matters: glue code between a framework and a protocol is work nobody bills for and that breaks on every release. An adapter shipped inside the framework means adding a tool to an agent is a config line again. Together with OpenTelemetry, traceability stops being a LangSmith-specific thing and becomes something you read with the same tooling already watching the rest of the backend.
LangChain changelog · LangSmith Cloud changelog · MCP roadmap
A small fix in datasette-mcp that solves a large problem
On 1 September Simon Willison released datasette-mcp 0.2, which exposes a native MCP server over SQLite and Datasette databases. The central change: execute_sql now returns arrays of objects instead of positional arrays.
Why it matters: with a positional array the model has to remember which column was third, and when it gets that wrong it does not fail, it invents. This is the kind of error that never reaches a log or throws an exception: it shows up in a report with a number in the wrong column. Returning named objects eliminates that entire class of hallucination, and it is a design decision worth copying in any in-house tool that hands rows to a model.
Simon Willison — datasette-mcp 0.2
The defaults closed
Supabase changed the Data API default: new tables in the public schema are no longer exposed unless explicitly enabled. Google, in parallel, is moving toward authentication keys tied to Service Accounts on the Gemini Enterprise Agent Platform, and on 2 September updated its model lifecycle documentation, retiring legacy endpoints.
Why it matters: both changes point the same way, and they land in the same fortnight as the models that operate on their own. An open default is one problem when a person is querying; it is another thing when the caller is an agent allowed to chain requests. The downside is familiar: a project that worked because the table exposed itself breaks when the default changes, and that check belongs before the next migration, not after.
Supabase changelog · Gemini Enterprise Agent Platform · Gemini API deprecations
The execution layer got faster
PyTorch 2.14.0 reached general availability on 2 September, with the wider announcement on the 9th: GIL-free execution on Python 3.13 and 3.14, accelerated scaled_dot_product_attention, and AOT compilation improvements for audio and vision models. Transformers v5.14.1 integrated SDPA in an optimised path, with inference speedups of up to 40%. On the audio side, Gemini 3.5 Transcribe Live added continuous WebSocket transcription with word-level timestamps and multi-speaker detection, and Lyria 3.5 entered public preview on 3 September.
Why it matters: free-threaded stopped being experimental, which was the condition for using it on anything serving real requests. A 40% inference gain and word-level timestamps sound like lab details, but they are exactly what decides whether transcribing a shop's calls fits on a rented machine or has to be outsourced by the minute.
PyTorch announcements · Transformers on PyPI
Smaller embeddings, more predictable pgvector
The period's analyses point to quantised embeddings truncated to 512 or 768 dimensions, Matryoshka-style, cutting HNSW index RAM on Postgres by up to 50% with no meaningful retrieval degradation. On the infrastructure side, CREATE EXTENSION pgvector now automatically adopts the project-managed version, so a local Docker environment and the cloud land on the same version, and the Supabase CLI reached v2.117.x on 5 September with Envoy settled as the default self-hosted gateway. In image generation and understanding there was nothing new: FLUX 2 and Qwen2.5-VL in FP8 and NVFP4 remain the floor for running on your own hardware.
Why it matters: this chains directly onto last edition's text-embedding-004 deprecation. If a reindex is happening anyway, it is the moment to choose a dimension rather than inherit the default: half the index RAM is the difference between a small instance and one you have to justify every month. And pgvector's version no longer depending on the environment removes the most irritating class of bug there is, the one that only appears in production.
Embedding cost comparison 2026 · Supabase CLI releases · Qwen2.5-VL
What changed since the last period
The previous edition closed on auditability: managed authentication, end-to-end traces, regression tests. Those were the tools for watching what an agent does. This period delivered the agents that make the watching worth it. GPT-6 Astra, Fable 5.1 and 3.8 Flash are not pitched as better conversationalists: they are pitched on driving a computer and closing an incident.
What is interesting is that both arrived almost together. Supabase closing automatic table exposure and Google moving to Service Account-bound keys are not reactions to these launches; they shipped in parallel. But the order worked out well: first it became auditable, then it became safe to let it operate. That is the reverse of this industry's usual sequence.
What I am watching
- Whether the operating benchmarks hold outside the benchmark. 88% on SRE-Bench and record OSWorld 2.0 scores come from prepared environments. The real test is a system with dirty data, expired credentials and a case nobody documented. Worth measuring with a task of your own before widening an agent's scope in production.
- What breaks under the new defaults. The Supabase Data API change is silent until a table that used to expose itself stops answering. Worth auditing what depends on that behaviour before the next migration.
- Whether the reindex can be done once. Between last period's embedding deprecation and the evidence that truncating to 512 or 768 dimensions does not hurt retrieval, there is a single window to do the heavy work properly. Doing it twice is the expensive option.
Worth keeping in mind
Much of what was announced here is capability claimed by whoever sells it, measured on benchmarks they also chose. The numbers in this edition are the vendors' published figures, not results of my own. This records what shipped, not what has been proven.