What happened
Real-time voice stopped being a demo
On 15 September gemini-3.8-live and its extended-thinking variant reached general availability in the Gemini API, with continuous two-way speech, native acoustic processing and a claimed 97.7% accuracy on audio understanding. A week later, on the 22nd and 23rd, the synthesis models shipped: gemini-3.8-flash-tts for expressive narration and voice design, and gemini-3.8-flash-lite-tts at six dollars per million audio tokens — around 0.9 cents a minute — covering 130 languages with automatic detection, cloning from a 30-second sample, SynthID watermarks, two-speaker dialogue and inline stage directions. With PyTorch 2.14 on the input side and Whisper or Gemini 3.5 Transcribe Live handling transcription, the period leaves full voice pipelines documented under 800 ms, with asynchronous tool calls mid-conversation.
Why it matters: a cent a minute changes what you can put in front of a client. An agent answering a distributor's phone twelve hours a day stops being a number you have to defend and becomes a minor line next to the wage of whoever used to answer. What does not change is everything else: the script, what happens when a caller asks for something nobody planned for, and who the call gets handed to. Price is no longer the excuse for not building it, so the project is now won or lost on design.
Gemini API release notes · Gemini 3.8 Flash TTS analysis · Gemini 3.8 Live and real-time operation · PyTorch announcements
Three new models, and one cuts the price of the expensive part
On 17 September Google shipped antigravity-preview-09-2026 in the Gemini API, aimed at long-horizon autonomous software engineering and the orchestration of complex multi-agent systems with tool execution. On the 22nd OpenAI made gpt-6-sol and gpt-6-luna generally available for fast inference and lightweight reasoning, and Anthropic introduced claude-opus-5.5, which the announcement says matches Fable 5.1 at four dollars in and twenty out per million tokens — 40% less operating cost. On the same 3.8 series, Google's Enterprise Agent Platform added private VPC connectivity templates for agents consuming models and enterprise tools from inside a closed network.
Why it matters: last period's pattern was models measured on whether they finish a task. This one brings the other half — the same result for less. Forty percent off the model doing the heavy lifting is the difference between running a process once a day and running it every time an order arrives. And VPC connectivity is the requirement that shows up the moment a client has an ERP that never touches the internet: worth noting the template now exists, because until now that conversation ended in a hand-rolled proxy.
Gemini API release notes · OpenAI release notes · Opus 5.5 · Gemini Enterprise Agent Platform
MCP no longer needs glue code
Google Cloud API Gateway gained the ability to turn corporate REST APIs into Model Context Protocol tools without writing an adapter layer, and a managed MCP server for AlloyDB and PostgreSQL was introduced, letting agents query relational databases under a standard schema. On the LangChain side, every agent deployed to LangSmith Cloud now automatically exposes an MCP endpoint any compatible client can consume, and the new Python release added delegated LangSmith access for sandboxes, stop_reason capture in agentic runs and experimental typed validation. The migration from AgentExecutor to create_agent on LangGraph with MCPAdapter has a support deadline of December 2026. In parallel, Home Assistant 2026.9, Smarsh and Cloudflare adopted the standard under a stateless architecture.
Why it matters: until now, exposing an in-house system to an agent meant writing an MCP server, then maintaining and versioning it. If the gateway generates that from the REST API already in place, the work drops from development to configuration — and that is what makes wiring up a client's billing system something other than a separate project. The flip side is the December date: any agent still running on AgentExecutor has an expiry on it, and that migration is better done while it is an afternoon rather than an emergency.
REST APIs to MCP tools with API Gateway · AlloyDB MCP server · LangSmith Cloud changelog · LangChain release policy
Supabase opened an exit pipe and shut several doors
Between 21 and 23 September Supabase launched Pipelines, formerly Database Replication, which streams changes in real time from PostgreSQL into BigQuery and other analytical stores. On the 18th came Health Check Advisors, for auditing connection bottlenecks, memory usage and vacuum statistics, alongside managed versioning for extensions such as pgvector, which removes the mismatch between a local container and the cloud. On the 23rd the logs.all endpoint was fully retired in favour of the ClickHouse-based analytics stack, and strict public-schema isolation becomes mandatory in October, with data access mediated by FastAPI microservices and Pydantic v2 validation.
Why it matters: native CDC removes the nightly script that copied tables, which is the component that fails most quietly in any small installation. The other three items are maintenance, and the kind that gives no warning: logs.all simply stops answering, and a dashboard querying it sits at zero without throwing an error. Same with the schema change in October. Worth auditing today what still calls those endpoints, because the cost of finding out late is paid by a client staring at an empty chart.
Supabase changelog · Supabase releases on GitHub · Supabase release notes
The embedding leaderboard moved; the technique did not
Comparisons published on 19 September put Voyage 4 and Gemini Embedding 2 at the top for API-based contextual retrieval accuracy, while BGE-M3 still leads multilingual open-source deployments. Implementation is unchanged: HNSW indexes in pgvector, Matryoshka truncation to 512 or 768 dimensions for half the RAM with no meaningful degradation, and a combination with lexical search via tsvector and websearch_to_tsquery in the same PostgreSQL instance.
Why it matters: three editions in a row describing the same architecture is a signal, and a good one. It means a decision made a month ago still holds and the reindexing work does not repeat. The model leaderboard does move, but it moves at the top: choosing between Voyage 4 and Gemini Embedding 2 shifts accuracy points, while choosing the wrong dimension shifts the server bill every month.
Best embedding models for RAG (2026) · Supabase documentation · Supabase blog
Running large models on small machines
Transformers added native support for running GGUF quantisations — llama.cpp's format — straight from Python. That makes it possible to load large language models in memory-constrained environments with minimal degradation and wire them into inference pipelines on PyTorch 2.14 GA, which settled GIL-free execution on Python 3.13 and scaled-attention acceleration. Nothing new on the vision side: Qwen2.5-VL for diagram extraction and FLUX in FP8 or NVFP4 for generation remain the floor for working on your own hardware.
Why it matters: GGUF belonged to the local-inference world and Transformers to the Python one. Having to pick one meant maintaining two paths, one for testing on a laptop and another for production. Same code on both sides removes the class of bug that only appears at deploy time, and it enables something more concrete: offering a client who will not send data outside a model running on their own server, without that meaning a separate stack to maintain.
Hugging Face blog · Transformers on PyPI · PyTorch 2.14 release blog · Qwen2.5-VL
What changed since the last period
The previous edition closed on models measured by whether they finish a task, and platforms tightening their defaults. This period adds no new capability in that direction: it adds pricing and plumbing. Voice came down to roughly a cent a minute, Opus 5.5 delivers Fable 5.1 performance 40% cheaper, and the integrations that used to be written by hand are now generated from the REST API that was already there.
It is the less eye-catching move and the more useful one. A model that operates on its own raises the question of what can be delegated; a model that operates on its own at a price that works turns it into a concrete proposal. The uncomfortable part arrives together: several of the period's items are dated removals. logs.all went on 23 September, public-schema isolation becomes mandatory in October, and AgentExecutor expires in December. What this fortnight made cheaper, it also put on a calendar.
What I am watching
- Whether voice pricing survives real volume. Six dollars per million audio tokens is the output rate; a real conversation also carries input, transcription, the reasoning model in the middle and the retries when the caller talks over the agent. Before quoting a phone agent, worth measuring one full call end to end and deriving the per-minute cost from that.
- What is still pointing at
logs.all. The endpoint went on 23 September and the failure mode is silent: a dashboard that stops updating without an error. Worth checking your own projects before a client notices, and checking what depends on the exposed public schema while you are in there, since that closes in October. - How long converting a REST API into MCP tools actually takes. The promise is no glue code. The test is an API with its own token scheme, odd pagination and errors returned with a 200, which is what the APIs of the systems you have to integrate actually look like. Worth trying it against an ugly API before promising it in a proposal.
Worth keeping in mind
The figures in this edition are the ones vendors published in their own announcements: list prices, accuracy percentages and cost improvements measured by whoever sells the model. There is no independent testing behind them. This records what shipped, not what has been verified.