AI news, hand-picked and dated
14 entries, newest first. Each one says what happened, why it changes a decision you might actually make about a tool, and links to the announcement from the vendor that made it.
October 2026
Requests to Gemini 3.7 Flash now go to 3.8 Flash without asking
The Gemini API changelog entry for 8 October deprecates gemini-3.7-flash and states that all requests to it are automatically routed to gemini-3.8-flash. The old model name still works; it just no longer means what it says.
Silent routing is the most invisible kind of model change: an integration that pinned 3.7 Flash keeps running, keeps being billed, and quietly answers differently from this week. That is fine on the vendor's terms and terrible for yours if you ever compared two tools on 3.7's behaviour. For subscribers it means the same thing as the September gate on 2.5 — the model named in a tool's copy resolves to whatever the vendor's routing table points at now, which is worth asking about when an answer suddenly reads worse.
Claude Haiku 5.5: a 1M-token context window on the cheap tier
Anthropic launched Claude Haiku 5.5 on 7 October, the third model of the 5.5 family it said would follow Opus 5.5 "in the coming weeks". Its platform docs list a 1M-token context window, up to 128K tokens of output, adaptive thinking, and pricing from $0.10 per million input tokens and $0.50 per million output, moving to higher rates once prompts pass 100K tokens. It is available through the API and the cloud marketplaces Anthropic lists at launch. The page calls it the "cheapest, fastest, and most capable small model we've ever released" — a vendor claim, not something we have tested.
Haiku-class models are what ride inside the free and low-cost tiers of a hundred subscription apps, so this price line is what a wrapper's marginal request actually costs next month. The practical change is long documents: a million tokens on the cheap tier means feeding an entire contract pile to a summariser stops being an enterprise feature. If a note-taking or document tool on your card has been promising that for a year and charging premium rates for it, check which model it names now.
OpenAI put GPT-6 in ChatGPT with "Intelligent UI" — details we could not read
OpenAI's news index lists an announcement titled "GPT-6 and Intelligent UI for everyone", describing GPT-6 in ChatGPT answering with fully interactive user interfaces rather than plain text. We are repeating nothing else from it: the announcement page blocks automated fetching, so plans, tiers and rollout dates sit beyond what we could verify from the vendor's own page, and this desk does not fill that gap from third-party summaries.
The one claim we can read on the vendor's own page — answers arriving as interactive interfaces — is the interesting part for anyone paying for a subscription. ChatGPT generating widgets instead of text competes with the small tools you may pay for separately: form-builders, dashboards, one-off trackers. Log in and run a task you normally do twice; if the answer is now an interface you can use, the value of the second tool is the first question. What your specific tier got is not something we can tell you from what OpenAI published retrievably — check the in-app changelog.
An image model went GA and its predecessor was deprecated in the same changelog entry
Google's 6 October changelog entry marks Gemini Nano Banana 2.1 as generally available and, in the same breath, records that the gemini-3.1-flash-image model has been retired. Same product family, same day, no transition period worth mentioning.
Image generation is the layer where model swaps are most visible to end users, because style changes. A tool that advertises "the best image model" is advertising a moving part, and when the vendor retires it the tool either upgrades quietly or quietly gets worse. Ask which one happened last time you noticed a change in output.
Gemini 2.5 gets a hard date: 16 October
Google Cloud's Vertex AI release notes record, in an update dated 2 October, that the retirement dates for Gemini 2.5 Pro, 2.5 Flash and 2.5 Flash-Lite have been set to 16 October 2026. This completes what the API changelog started on 18 September, when access to the 2.5 line was restricted to recently-active accounts.
The gate became a wall, with a date you can put in a calendar. Anything still pinned to a 2.5 model — your own automations, a smaller wrapper that shipped its integration last year — has days, not months, to move or find out what its vendor's auto-migration does to output quality. A tool that cannot name the model it replaced 2.5 with is telling you the swap was not a decision.
September 2026
Claude Opus 5.5: the top tier got 40% cheaper and the subscriber limits went up
Anthropic introduced Claude Opus 5.5 on 22 September as the first model of the 5.5 family, priced at $4 per million input and $20 per million output — 40% below Opus 5 — and claims output arrives more than 30% faster. Alongside the model, the five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans were increased.
For anyone on a subscription rather than the API, the limit change is the story. Price cuts on a model you do not buy per token are irrelevant; the wall you actually hit is the rolling usage window, and it moved in your favour. This is also the clearest example this month of why our profiles separate "what the API costs" from "what your plan lets you do".
OpenAI widened the GPT-6 line with cheaper Sol and Luna
OpenAI announced GPT-6 Sol and GPT-6 Luna on 22 September, positioned as a cheaper pair inside the GPT-6 family, and followed with GPT-6.1 Sol on 29 September. We are deliberately not quoting per-token figures here: OpenAI's announcement pages block automated fetching, so anything we repeated would come from a third-party summary rather than the vendor, and this site does not publish numbers it has not read on the page that owns them.
The naming is the practical lesson. A family that gains variants weekly means the model named in a tool's marketing copy is a moving target — "GPT-6" on a £20 plan and "GPT-6" on a business plan may not be the same model at all. When a subscription will not tell you which variant your tier gets, that is an answer in itself.
Xiaomi's MiMo-V2.6 arrived as open weights
Xiaomi introduced the MiMo-V2.6 series on 22 September, publishing the Pro variant under open weights; coverage at the time positioned it as the strongest open-weight model available. That last claim is a third-party judgement we have not tested, but the release itself is verifiable from the vendor's own page.
Every strong open-weight release lowers the floor for what you should expect from money spent on a closed model. It is also why "can I run this locally" moved from a privacy curiosity to a cost question — the honest answer on our local-model profiles is that the software is free and the hardware and the upkeep are not.
Grok 4.7 lands at $2 / $6 per million tokens, inside Cursor
xAI announced Grok 4.7 on 21 September, priced at $2 per million input tokens and $6 per million output, with a faster variant at double that. It is available through the API and inside third-party tools including Cursor, and the announcement lists a spread of coding and agent benchmarks — which we have not run and will not restate as fact. Selected security partners get invite-only access to red-team capability for defensive research.
The useful signal is the price line, not the leaderboard: a coding-capable model at $2/$6 puts pressure on every tool that resells model access at a flat subscription. If you are already inside Cursor, the model is a dropdown, so the honest test is your own repository on the same task, not a benchmark you cannot reproduce.
Google quietly restricted access to the Gemini 2.5 models
The Gemini API changelog entry for 18 September limits access to the 2.5 models to accounts that have actively used them recently, applies capacity limits, and points new projects at 3.8 Flash instead. The previous generation was not deleted; it was put behind a gate.
This is what model churn looks like from the inside, and most of it is invisible to whoever buys the wrapper. A lot of subscription apps advertise "powered by the latest model" while shipping whatever endpoint was cheapest when they built the integration. If a tool on your card list has not named its model and its date in a while, ask. Our evaluation guide has the five-minute version of that question.
Real-time voice models went GA across Google's stack
Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking reached general availability on 15 September, per the Gemini API changelog. A week earlier the same log recorded Gemini 3.5 Transcribe going GA, and later in the month it added 3.8 Flash TTS variants with voice design and replication tools.
Voice features in the apps you already pay for are usually built on the Live line, so this is the quiet month where "talk to it" stopped being a demo. It also matters for meeting tools: transcription and summarising keep collapsing into one product, which changes the maths if you pay for a separate note-taker. Test voice mode in the room you actually use it in — a noisy desk kills more of these features than model quality does.
Google's study tools now live under a "Gemini Notebook" name
A 15 September round-up of new study and learning features sits on a gemini-notebook product path on Google's blog, rather than the NotebookLM path that covered the same ground last year. The post itself is short on specifics in the part that loaded for us, which is its own hint about how these rollouts are being run.
Names drift, and small products get absorbed into platform brands without a migration notice. If you keep a library of sources in a note-taking AI, the practical risk is not the rename — it is the feature set moving between the free and paid tiers while you are not looking. Worth re-reading the plan you bought against the current feature page once a year.
DeepSeek ships V4.1-Flash and prices off-peak at half rate
DeepSeek released V4.1-Flash on 10 September, described in its own notes as the smallest model in the new architecture family, with native visual understanding and a 552-billion-parameter mixture-of-experts design. The release states that off-peak rates are 50% of peak rates, and claims the model's cache needs roughly a quarter of the HBM. Weights went out publicly rather than behind an API only.
Two levers most buyers ignore. First, time-of-day pricing: if you run anything scheduled — nightly summaries, batch transcription, a report that doesn't need to exist by 9am — running it off-peak is now a real discount rather than a rounding error. Second, open weights mean the same model can be run somewhere else, which is the argument for the local tools in our new Local & open models category.
Gemini 3.8 Flash goes GA, with a security-specific twin
Google took Gemini 3.8 Flash to general availability on 2 September, describing it as built for software engineering and agent work, and shipped a parallel 3.8 Flash Cyber variant aimed at defensive security tasks. The vendor's own announcement sums the pair up as "next-generation intelligence for agentic workflows and cybersecurity".
Flash-class models are the ones that end up inside consumer subscriptions, because they are cheap enough to hand to free and low-tier users. So if you use Gemini through one of Google's apps rather than the API, the model underneath your account changed this month without a dialog box. Run the same task you always run and check whether the answer got better or merely more fluent.
Keep reading
- How to evaluate an AI tool before you pay for it — the test structure behind every profile here.
- All tool profiles — each says what the tool is bad at.
- How we check a tool — what we do, and the claims we refuse to make.
- Updates and corrections — every substantive change to this site, dated.
Disclosure: this page is editorial and maintained by one person. Vendors listed here did not review it, pay for it or provide accounts for it. Outbound links are plain links, not affiliate links. This is a news summary, not a transcript — the linked announcement is the authoritative text.