# AI API changes in October 2026: what to act on, and when

URL: https://infro.io/blog/ai-api-changes-october-2026
Author: INFRO Editorial (https://infro.io/blog/editorial)
Publisher: INFRO
Published: 2026-10-27
Last modified: 2026-10-27
Last reviewed: 2026-10-27 (next review due 2026-11-26)
Topic: Architecture and migration
Format: Analysis

> Between 6 and 31 October 2026 the changes that need action are: Anthropic's API version 2026-10-21, which new SDK majors send by default and which returns 400 for temperature, top_p and top_k; Gemini 3.7 and 3.5 Flash requests now answered by 3.8 and 3.6 Flash; Gemini's Generate Content API storing requests by default for new paid projects; Vertex AI retiring Gemini 2.5; and billing that starts on 31 October for DeepSeek's server-side tool calls. Plan this quarter for Gemini 4 Argon's rate doubling on 1 December, OpenAI's GPT-4o and GPT-4.1 shutdown on 14 January 2027, and Azure's and Moonshot's own retirements.

Key takeaways:
- Upgrading to Anthropic's new SDK majors moves a client to API version 2026-10-21, where temperature, top_p and top_k return a 400 even at their default values.
- Gemini 3.7 Flash and 3.5 Flash were not shut down: since 8 October their requests are answered by 3.8 Flash and 3.6 Flash, with no change in your code.
- New paid Gemini API projects store Generate Content requests by default since 26 October; existing projects keep their setting.
- On 31 October DeepSeek starts billing server-side Search and Fetch calls, OpenAI's Evals turn read-only and Moonshot closes its K2 models to new users.
- The same model retires on different dates on different platforms: Gemini 3.6 Flash leaves Vertex AI on 19 November, and Claude Haiku 4.5 leaves Microsoft Foundry on 15 November while Anthropic lists it as active.

In the second half of October 2026 the AI vendors changed more running code than they launched. The items that need action now are a new Claude API version that Anthropic's new SDKs send by default, Gemini 3.7 and 3.5 Flash requests that are already answered by a newer model, a Gemini storage default that flipped for new projects, previews that were switched off, and tool calls that start billing on 31 October. The items to plan this quarter are a Gemini 4 Argon rate that doubles on 1 December, OpenAI's GPT-4o and GPT-4.1 shutdown on 14 January 2027, and retirements on Azure and Moonshot that do not match the model vendors' own dates.

This roundup covers 6 to 31 October. Everything earlier — GPT-6, the Claude 5.5 models, Gemini 3.8 Flash, OpenAI's 23 October shutdowns — is in [AI updates, September–October 2026](https://infro.io/blog/ai-updates-september-october-2026) and [OpenAI's October 23 shutdowns](https://infro.io/blog/openai-model-shutdowns-october-2026), and is not repeated here. Each item below was confirmed on the vendor's own page on 27 October; the dated events come from the [AI updates timeline](https://infro.io/ai-updates).

## Act now

### Anthropic's API version 2026-10-21 arrives with the new SDKs

On 21 October Anthropic released API version `2026-10-21` and new SDK and CLI majors that send it by default: Python 2.0.0, TypeScript 1.0.0, C# 13.0.0, Go 2.0.0, Java 3.0.0, PHP 1.0.0, Ruby 2.0.0 and `ant` 2.0.0[1]. The release notes add that these majors carry breaking changes unrelated to the API version and move from the MIT licence to Apache 2.0[1].

On `2026-10-21`, three request shapes that used to work return a 400[2]:

- `temperature`, `top_p` or `top_k` in a Messages or Message Batch request, "even at the parameter's default value".
- A `document` block placed directly in an `assistant` message.
- A `text` block with a non-empty `citations` array in a `user` message.

The sampling change is the one most code will hit, because many wrappers send `temperature` on every call. Anthropic's guide says models that already accept only the defaults, Claude 4.7 and later, sample as before when you drop the field; on models that accept other values, such as Sonnet 4.6, a request that lowered `temperature` loses that effect and "no other setting replaces it"[2]. Anthropic's deprecations page now lists the three parameters as removed in this version[3].

**What to do.** Treat an Anthropic SDK upgrade as a code change, not a dependency bump. Before you upgrade, search for every place that sets sampling parameters, including retry wrappers and framework defaults, and remove them or keep that caller on `2023-06-01` by pinning the SDK major. If you call the API without an SDK, nothing changes until you set the header yourself. The full migration, field by field, is in [Anthropic API version 2026-10-21: what breaks](https://infro.io/blog/anthropic-api-version-2026-10-21).

### Gemini Flash requests are already served by another model

On 8 October Google deprecated `gemini-3.7-flash` and `gemini-3.5-flash` on the Gemini API. Requests to 3.7 Flash are "automatically routed" to `gemini-3.8-flash`, and requests to 3.5 Flash to `gemini-3.6-flash`[4]. Google's deprecation page lists no shutdown date for either[5].

Nothing fails, which is the problem: an evaluation, a prompt or a cost estimate tuned to 3.7 Flash now runs on a different model. Re-run your evaluation on the model actually answering, and check which model id the response reports.

Vertex AI keeps its own schedule. There, Gemini 3.6 Flash retires on 19 November 2026 and 3.7 Flash on 28 January 2027, both with 3.8 Flash as the replacement, and the Gemini 2.5 Pro, Flash and Flash-Lite models retired on 20 October[6]. Google's page notes that retirement dates may be extended but will not move earlier[6].

### Gemini's Generate Content API stores requests by default on new paid projects

Since 26 October, Generate Content API storage is on by default (`store=true`) for new paid-tier projects, matching the Interactions API; existing projects keep their current setting[4]. To opt out, set `store=false` per request or turn off Generate Content logging in the Logs and Datasets settings[4].

This matters to anyone whose data-handling commitments assume Gemini requests are not kept. A team that creates a new paid project for a new product inherits the new default. Decide per project, write the choice down, and check it in whatever review creates cloud projects.

### Preview ids that stopped answering

Three sets of ids ended this month, and a caller you missed shows up as errors rather than a warning:

- **Veo 3.1 previews.** `veo-3.1-generate-preview`, `veo-3.1-fast-generate-preview`, `veo-3.1-lite-generate-preview` and `gemini-omni-flash-preview` had an earliest shutdown of 22 October; Google's deprecation page names `gemini-omni-1.1-flash` as the replacement[5].
- **Deep Research Pro preview.** `deep-research-pro-preview-12-2025` shut down on 23 October; Google points `agent` in `interactions.create` calls to `deep-research-preview-04-2026` or `deep-research-max-preview-04-2026`[4].
- **Older OpenAI ids on Perplexity.** At 00:00 UTC on 24 October, Perplexity's Agent and Orchestrator APIs stopped accepting `openai/gpt-5.4`, its mini and nano, `gpt-5.2`, `gpt-5.1`, `gpt-5` and `gpt-5-mini`, with no one-to-one replacement named; Perplexity asks you to update direct model choices and fallback chains[7].

The last one is the reminder that an aggregator retires a model on its own date, which can come before the model vendor's.

### Three changes on 31 October

- **DeepSeek tool billing.** Search and Fetch calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free for DeepSeek API users until 16:00 UTC on 31 October, then billed at $0.75 per 1,000 calls for each endpoint[8]. Agent loops that call search freely will show a new cost line.
- **OpenAI Evals turn read-only.** Existing evals become read-only on 31 October, ahead of the Evals dashboard and API shutting down on 30 November[9]. Export what you need while you can still change it.
- **Kimi K2 closes to new users.** From 31 October Moonshot's K2 models are no longer available to newly registered users, and the whole K2 series is planned for retirement by the end of December 2026[10]. A new environment or a new account cannot be set up on K2 after that date.

## Plan this quarter

**Gemini 4 Argon's rate doubles on 1 December.** Gemini 4 Argon reached general availability on 13 October[4]. It is not on the free tier, and its paid rates are $2 input and $10 output per million tokens through 30 November, then $4 and $20 from 1 December, with output including thinking tokens[11]. A budget for December onward should use the second rate; [budgeting for launch and introductory prices](https://infro.io/blog/ai-model-launch-pricing-budgets) works through how.

**OpenAI's GPT-4o and GPT-4.1 family shuts down on 14 January 2027.** On 14 October OpenAI announced that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini shut down on 14 January 2027, with `gpt-6-luna` as the replacement; fine-tuned versions follow their base models' schedule[9]. These are some of the most widely embedded ids in production code, which makes this the largest migration of the quarter. A dated plan for it is in [the GPT-4o and GPT-4.1 shutdown guide](https://infro.io/blog/gpt-4o-gpt-4-1-shutdown-january-2027); the general method is in the [deprecation playbook](https://infro.io/blog/model-deprecation-migration-playbook).

**Gemini preview speech and Live models end on 17 November.** `gemini-3.1-flash-tts-preview`, `gemini-2.5-flash-preview-tts`, `gemini-2.5-pro-preview-tts`, `gemini-3.1-flash-live-preview` and `gemini-2.5-flash-native-audio-preview-12-2025` have an earliest shutdown of 17 November; the replacements are `gemini-3.8-flash-tts` or `-flash-lite-tts`, and `gemini-3.8-live`[5].

**Azure retires on its own dates.** Microsoft's schedule lists `claude-haiku-4-5` (both hosting versions) and `codex-mini` for 15 November 2026, `mistral-medium-2505` and `Ministral-3B` for 9 December, and `mistral-small-2503`, deprecated on 9 October, for 15 February 2027[12]. Anthropic, by contrast, still lists Claude Haiku 4.5 as active on its own platforms[3]. If you run the same model on two platforms, keep two dates.

**Anthropic user profiles for platforms.** On 20 October Anthropic launched user profiles and the `/v1/user_profiles` endpoints, so that an enforcement action can apply to one end customer instead of the whole organization; on 21 October profiles gained `enforcement_status` and `enforcement_actions`[1]. A product that serves many customers from one Anthropic organization should plan to send a profile per customer: it narrows the blast radius of one customer's misuse. [Anthropic user profiles](https://infro.io/blog/anthropic-user-profiles-end-customer-attribution) covers the endpoints.

## Watch: releases and platform changes

None of these require action, but each changes what a team may choose next:

- **New models.** Claude Haiku 5.5 on 7 October (1M-token context, 128K output, and it rejects `temperature`, `top_p`, `top_k` and prefill)[1]; Nano Banana 2.1 generally available on 6 October, with `gemini-3.1-flash-image` deprecated and no date set[4]; GPT-6.1 Terra on 26 October[13]; Mistral Large 4 in public preview on 6 October and Voxtral 3 speech models in public preview on 20 October[14]. The open-weight releases of the month are compared in [October's open-weight wave](https://infro.io/blog/open-weight-frontier-october-2026).
- **Where models run.** Anthropic launched Claude Models on Google Cloud (Anthropic managed) on 22 October[1]; what that changes against Vertex AI, Bedrock and the direct APIs is in [Claude and OpenAI models through Google Cloud](https://infro.io/blog/claude-openai-models-through-google-cloud).
- **OpenAI platform.** The Decisions API became generally available on `gpt-6-luna` on 22 October, with up to 400 questions per request; 12-hour prompt caching arrived on 20 October with cache writes at 2.5 times the input price ([when the write pays off](https://infro.io/blog/openai-12-hour-prompt-cache)); and since 21 October organizations can restrict which service tiers each project may use[13].
- **Anthropic platform.** Fast mode for Claude Opus 5.5 left research preview on 15 October, OpenTelemetry export entered beta on 14 October, and Sonnet 5.5 cache reads dropped from $0.20 to $0.10 per million tokens on 7 October[1].

## The dates in one table

| Date | Vendor | Change | Bucket |
| --- | --- | --- | --- |
| 8 Oct 2026 | Google | gemini-3.7-flash and gemini-3.5-flash routed to 3.8 and 3.6 Flash | Act now |
| 20 Oct 2026 | Google Cloud | Gemini 2.5 Pro, Flash, Flash-Lite retired on Vertex AI | Act now |
| 21 Oct 2026 | Anthropic | API version 2026-10-21; new SDK majors send it | Act now |
| 22–24 Oct 2026 | Google, Perplexity | Veo 3.1 previews, Deep Research Pro preview, older OpenAI ids on Perplexity end | Act now |
| 26 Oct 2026 | Google | store=true by default for new paid Gemini projects | Act now |
| 31 Oct 2026 | DeepSeek, OpenAI, Moonshot | Tool-call billing starts; Evals read-only; K2 closed to new users | Act now |
| 15 Nov 2026 | Microsoft | claude-haiku-4-5 and codex-mini retire on Azure | Plan |
| 17 Nov 2026 | Google | Gemini preview TTS and Live ids end (earliest) | Plan |
| 19 Nov 2026 | Google Cloud | Gemini 3.6 Flash retires on Vertex AI | Plan |
| 1 Dec 2026 | Google | Gemini 4 Argon moves from $2/$10 to $4/$20 | Plan |
| 9 Dec 2026 | Microsoft | mistral-medium-2505 and Ministral-3B retire on Azure | Plan |
| 31 Dec 2026 | Moonshot | K2 series planned retirement | Plan |
| 14 Jan 2027 | OpenAI | GPT-4o, GPT-4o mini, GPT-4.1, GPT-4.1 mini shut down | Plan |
| 28 Jan 2027 | Google Cloud | Gemini 3.7 Flash retires on Vertex AI | Plan |
| 15 Feb 2027 | Microsoft | mistral-small-2503 retires on Azure | Plan |

Table: Dated changes from 6 October 2026 onward, from each vendor's page read on 2026-10-27. Gemini preview dates are Google's earliest possible dates.

## What to ask the team

For a CTO or a finance lead, the month reduces to four questions:

1. **Who owns SDK upgrades?** The Anthropic change ships inside a routine dependency bump. Someone has to read release notes for every AI SDK before it merges.
2. **Do we know which model answered?** Auto-routing means the model you asked for and the model that answered can differ. Logging the response's model id, per request, is the only way to see it.
3. **Is each platform's date on our calendar?** The same model has different end dates on the vendor's API, Vertex AI, Azure and an aggregator. A single "retirement date" per model is wrong.
4. **Which rates in our forecast change before the money is spent?** Argon in December, and earlier-announced changes such as Gemini Flash in January, should already be in next quarter's numbers.

## Tracking a month like this with INFRO

INFRO is a control plane that sits between your applications and the provider accounts you already hold; each provider bills you directly. It helps with the parts of this list that show up in traffic through it, and not with the rest.

**Who still sends a model.** `GET /v1/usage` groups requests and estimated provider cost by model, project and key, and request history filters by model, so you can find every caller of `gemini-3.7-flash` or a GPT-4o snapshot before its date. Traffic that goes straight to a provider does not appear.

**Retirements and notices.** INFRO records vendor deprecations and shutdown dates for catalog models, raises a `model.retiring` alert for an organization that sent a retiring model in the last 30 days, naming the keys and projects that called it (emailed to the billing contact by default, and to any channel your alert rules name), and from the shutdown date refuses the model before any provider call, naming the replacement. [Model allowlists](https://infro.io/docs/org/model-allowlists) block an id earlier if you want callers to fail in staging first.

**Silent model changes.** [Traffic changes](https://infro.io/docs/observability/traffic-changes) raise a finding when a connection starts being answered by a model id it had not reported before, which is how an auto-routed request looks from outside, provided the provider's response reports the new id.

What INFRO does not do here: it does not upgrade or pin your Anthropic SDK, so the parameter review above is still yours to run. It does not change a Gemini project's storage setting; that is set in your Google project. And a vendor's announcement is not in INFRO's catalog until INFRO adds it, so the vendor page stays the source of truth for a date.

## Questions

### Does upgrading the Anthropic SDK break my requests?

It can. The new majors released on 21 October (Python 2.0.0, TypeScript 1.0.0, Go 2.0.0, Java 3.0.0 and others) send anthropic-version 2026-10-21, which returns a 400 if a request sends temperature, top_p or top_k, puts a document block in an assistant message, or puts citations on a user text block. Remove those fields first, or keep the request on 2023-06-01 by staying on your current SDK major.

### Is Gemini 3.7 Flash shut down?

Not on the Gemini API. Google deprecated it on 8 October and routes its requests to gemini-3.8-flash; gemini-3.5-flash goes to gemini-3.6-flash. Google's deprecation page lists no shutdown date for either. On Vertex AI the dates are different: Gemini 3.6 Flash retires on 19 November 2026 and Gemini 3.7 Flash on 28 January 2027.

### When do OpenAI's GPT-4o and GPT-4.1 models stop working?

OpenAI announced on 14 October that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini shut down on 14 January 2027, with gpt-6-luna as the replacement. Fine-tuned versions follow their base models. On Azure OpenAI, Microsoft lists most of these versions for 14 April 2027.

### What changes for DeepSeek users on 31 October?

From 16:00 UTC on 31 October, Search and Fetch calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are billed. DeepSeek lists $0.75 per 1,000 calls for each endpoint; until then they were free for DeepSeek API users.

### Where can I see every dated change in one place?

INFRO's AI updates timeline at infro.io/ai-updates lists releases, deprecations, shutdowns and price changes with the vendor page each one came from. Always confirm a date on the vendor's page before you plan around it.

## Sources

1. Claude Platform release notes — Anthropic. https://platform.claude.com/docs/en/release-notes/overview (accessed 2026-10-27)
2. Migrating to API version 2026-10-21 — Anthropic. https://platform.claude.com/docs/en/api/migration-2026-10-21 (accessed 2026-10-27)
3. Model deprecations — Anthropic. https://platform.claude.com/docs/en/about-claude/model-deprecations (accessed 2026-10-27)
4. Gemini API release notes — Google. https://ai.google.dev/gemini-api/docs/changelog (accessed 2026-10-27)
5. Gemini deprecations — Google. https://ai.google.dev/gemini-api/docs/deprecations (accessed 2026-10-27)
6. Model versions and lifecycle — Google Cloud. https://docs.cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions (accessed 2026-10-27)
7. Changelog — Perplexity. https://docs.perplexity.ai/changelog (accessed 2026-10-27)
8. DeepSeek-V4.1-Flash-1017 — DeepSeek. https://api-docs.deepseek.com/news/news261017/ (accessed 2026-10-27)
9. Deprecations — OpenAI. https://developers.openai.com/api/docs/deprecations (accessed 2026-10-27)
10. Chat model pricing — Moonshot AI. https://platform.kimi.ai/docs/pricing/chat (accessed 2026-10-27)
11. Gemini Developer API pricing — Google. https://ai.google.dev/gemini-api/docs/pricing (accessed 2026-10-27)
12. Model retirement schedule — Microsoft. https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/model-retirement-schedule (accessed 2026-10-27)
13. Changelog — OpenAI. https://developers.openai.com/api/docs/changelog (accessed 2026-10-27)
14. Changelog — Mistral AI. https://docs.mistral.ai/resources/changelogs (accessed 2026-10-27)