AI updates, September–October 2026: models, prices, shutdowns

The five weeks from 1 September to 6 October 2026 were one of the busiest stretches of the year for anyone running AI in production. OpenAI, Anthropic and Google each shipped new top models; xAI, DeepSeek and Alibaba updated theirs; and TypeSafe released a model that does not generate text at all. The same weeks ended two widely used APIs, scheduled a long list of shutdowns, and put dates on price rises.

This roundup sticks to what each vendor published on its own pages, read on 6 October. It leaves out benchmark claims and concentrates on what changes a budget, a deployment or a line of code.

OpenAI: GPT-6 in three tiers, then GPT-6.1 Sol

OpenAI released GPT-6 Astra on 3 September as its most capable model, at $10 per million input tokens and $50 per million output¹². Requests with more than 272K input tokens are billed at twice the input rate and 1.5 times the output rate for the whole request². Astra arrived with constraints that matter to existing code: it does not accept none reasoning effort, custom temperature or top_p, or logprobs, and tool calling requires the Responses API¹. AWS made Astra generally available on Amazon Bedrock on 8 September³.

On 22 September OpenAI added GPT-6 Sol at $2 input, $0.20 cached input and $10 output, and GPT-6 Luna at $0.10, $0.01 and $0.50¹. A week later, on 29 September, GPT-6.1 Sol followed at the same $2 and $10, with cached input at $0.10, alongside an "ultrafast" service tier for Astra on the Responses API¹.

The platform changed around the models too. The Agents API entered public beta on 10 September; project API keys can now carry expiry dates with an admin-enforced maximum lifetime; and organizations and projects gained controls over who may create keys¹. Sora 2 and the Videos API shut down on 24 September⁴.

Anthropic: Fable 5.1, Opus 5.5 and Sonnet 5.5

Anthropic launched Claude Fable 5.1 on 1 September at $10 and $50, with cache reads cut to $0.25 per million, a 1M-token context and 128K output⁵. Claude Opus 5.5 followed on 22 September at $4 and $20, below Opus 5's $5 and $25, and Claude Sonnet 5.5 on 28 September at $2 and $10, which the launch note calls introductory pricing⁵⁶. Anthropic also confirmed that Sonnet 5's $2 and $10 is now its standard price and that a planned increase to $3 and $15 will not happen⁶.

The launches carry breaking changes that a model-id swap will hit:

• On Fable 5.1, Opus 5.5 and Sonnet 5.5, forced tool use — tool_choice of any or tool — returns a 400 error; Anthropic's advice is auto with strict tool use⁵.
• On Opus 5.5, thinking cannot be disabled; both disabled and enabled thinking settings return a 400⁵.
• On Sonnet 5.5, turning off up-front thinking now means thinking: {"type": "between_tools"} instead of disabled, and thinking blocks are tied to the model and the conversation⁵.

Anthropic also made cache diagnostics generally available on 23 September, and resumed billing on 24 September for refusals that arrive before any output in three categories⁵. On 30 September it announced the retirement of Claude Sonnet 4.5 on 30 November 2026, with Sonnet 5.5 as the replacement⁵. Claude Haiku 4.5 is listed for retirement no sooner than 15 October 2026⁷.

Google: Gemini 3.8 Flash and a dated price rise

Google made Gemini 3.8 Flash generally available with a 1M-token context and 64K output⁸. It prices 3.8, 3.7 and 3.6 Flash identically: $0.75 input and $3.75 output through 31 December 2026, then $1.50 and $7.50 from 1 January 2027, with context caching doubling on the same day⁹. Any forecast that runs past December should use the 2027 rate.

Migration is not a model-id swap here either. Google's notes say to remove the temperature, top_p and top_k sampling parameters, to replace thinking_budget with the thinking_level setting (low, medium by default, or high), and that the minimal level returns an error on 3.8 Flash⁸. Google lists 2 October as the shutdown date of the original Nano Banana image model, gemini-2.5-flash-image, and 22 October as the earliest shutdown of the Veo 3.1 preview ids¹⁰⁹.

xAI, DeepSeek, Alibaba and Mistral

xAI released Grok 4.7 in September¹¹: $2 input, $0.50 cached and $6 output below 200K prompt tokens, $4 and $12 above, with a 500K-token context and reasoning effort from low to xhigh¹².

DeepSeek released V4.1-Flash on 10 September, served under the deepseek-flash name, with new rates and off-peak prices at half the peak rate; the older deepseek-v4-flash names now route to it¹³. Its pages disagree about V4 Pro: the announcement says V4 Pro requests move to V4.1-Flash from 14 September until V4.1-Pro launches¹³, while the change log says DeepSeek will keep serving V4 Pro after that date with billing unchanged¹⁴. Check which model your responses report.

Alibaba moved the unpinned qwen3.8-max alias to the new qwen3.8-max-0902 snapshot on 5 September, with pricing unchanged¹⁵. A request that named the alias got a different model with no code change — the case for pinning snapshots in production.

Mistral gave one day's notice on 29 September that OCR 4.0 would retire on 30 September, recommending OCR 4.1 at the same price, and set 31 October for retiring GLM 5.2 on its platform in favour of GLM 5.3¹⁶.

Perplexity ends Sonar Chat Completions

Perplexity ended Sonar Chat Completions support on 27 September. Synchronous and streaming requests keep working because Perplexity is reformulating them as Agent API requests, rolling out by model; asynchronous Sonar requests are no longer supported and have to move to the Agent API's background mode¹⁷. A request that keeps working while its semantics change underneath it is worth re-testing, not just leaving alone.

TypeSafe Jev: a model that does not write

TypeSafe announced Jev on 15 September as its first "System One" model: it answers typed questions about content — yes/no, one of a set of options, a score on a rubric — with probabilities rather than text, and bills $0.042 per million input tokens with output free¹⁸. It is a different kind of model from everything else in this list, built for the classification, routing and checking work that otherwise runs on chat models. What Jev is, its use cases and the cost arithmetic cover it in depth.

What the month says about running AI in production

Upgrades are code changes. Every major vendor shipped a new version that rejects parameters the previous one accepted: sampling parameters on GPT-6 Astra and Gemini 3.8 Flash, disabled thinking and forced tool use on the new Claude models¹⁵⁸. A model id in a configuration file is not a safe place to change models without a test run.

Prices fell at the top and the bottom, with dates attached. Opus 5.5 lists below Opus 5, and GPT-6 Luna lists at $0.10 input⁶¹. But Gemini Flash's rate doubles in January and Sonnet 5.5 launched at introductory pricing⁹⁵. Budget on the rate that will apply when the money is spent, not the one on the launch post.

Notice periods are shrinking at the edges. Most retirements came with months of notice, but Mistral's OCR 4.0 had one day, and aliases such as qwen3.8-max moved on their own¹⁶¹⁵. Teams that learn about a retirement from failing requests find out on the vendor's schedule, not their own.

APIs are diverging. Tool calling on GPT-6 Astra needs the Responses API, and Perplexity's Sonar now runs as Agent API calls¹¹⁷. The OpenAI-compatible Chat Completions shape remains the common denominator, but it no longer reaches every feature of every top model.

Deadlines for the next two quarters

The largest single event is OpenAI's 23 October shutdown of older snapshots, among them gpt-3.5-turbo-0125, gpt-4-turbo, gpt-4.1-nano, gpt-4o-2024-05-13, the dated o1, o3-mini and o4-mini snapshots, and gpt-image-1⁴. On 30 November OpenAI also shuts down the reusable prompts API, Agent Builder and the Evals platform, and on 1 December gpt-image-1.5 and gpt-image-1-mini⁴. OpenAI deprecated GPT-5.4 nano, GPT-5.1 and GPT-5.3-Codex on 1 October, with shutdown on 1 April 2027⁴.

Regulation has dates too. The European Commission states that the AI Act's rules for high-risk systems in areas such as employment, education and critical infrastructure apply from 2 December 2027¹⁹. Teams using AI to screen candidates or score applicants should be designing for that now.

Handling a month like this with INFRO

INFRO's model catalog carries each vendor's list rate for the models above as reference data for cost estimation, with a page per model, such as GPT-6 Luna and Claude Sonnet 5.5. Each provider still bills its own account directly; a negotiated rate entered for your organization replaces the list rate in every estimate.

Three gateway features map onto the checklist. Usage grouped by model shows whether any traffic still uses a retiring id, and how much: GET /v1/usage?group_by=model returns requests and estimated cost per model across every application that calls through the gateway. Routes turn a model change into a configuration change: a weighted route can send 5% of traffic to the replacement, and the old model stays as a fallback target until it is removed, as the deprecation playbook describes. Model allowlists stop applications from adopting a model the organization has not evaluated, whatever was announced this week.