What Changed for Quality Teams: November 2026
The monthly briefing: what changed between 20 October and 19 November 2026, who it affects, what to do, and the source. Nothing here is a prediction.
The change most likely to slip past a test suite this month raises no error at all. From 1 December, Google's Gemini API quietly sends requests for two retired models to newer ones, so the system under test changes and every check still passes. It lands in the same fortnight as other model retirements, the shutdown of OpenAI's hosted Evals, and an Anthropic API version that rejects the temperature setting many graders pin. If your AI quality evidence rests on a fixed model, a hosted eval tool and a fixed temperature, all three move between 30 November and 14 January. Re-baseline before the dates, not after.
Models you test against retire, and one family re-routes silently
What changed. Anthropic retires claude-sonnet-4-5-20250929 on the Claude API on 30 November 2026 (on Amazon Bedrock, 8 April 2027). OpenAI removes the gpt-5 family and o3 snapshots on 11 December 2026, and the dated GPT-4o and GPT-4.1 snapshots on 14 January 2027; fine-tuned models follow their base model. From 1 December 2026, Gemini API requests to gemini-2.5-pro route automatically to gemini-3.8-flash, and gemini-2.5-flash to gemini-3.5-flash-lite, billed at the new rates. Google lists these as earliest possible dates.
Who it affects. Any feature, agent, test oracle or grader pinned to these IDs. The Gemini re-route is the dangerous one: a suite that only checks for errors will pass while the behaviour changes.
What to do. Search code and configuration for the IDs this week, and run your regression and evaluation suite on each replacement before its date. For Gemini, pin the replacement yourself and re-baseline instead of letting the re-route switch for you.
Sources: Anthropic, Amazon Bedrock, OpenAI, Google
OpenAI Evals and reusable prompts shut down on 30 November
What changed. OpenAI's hosted Evals are read-only. The Evals dashboard, API and graders shut down on 30 November 2026, along with the v1/prompts API and reusable prompt objects. OpenAI's suggested replacement for Evals is Promptfoo.
Who it affects. Teams whose regression tests, graders, eval datasets or production prompts live in OpenAI's hosted objects rather than their own repository.
What to do. Export eval definitions, datasets, graders and past results now, and move prompts into version control beside the code that uses them. Re-baseline on the new harness before using it to judge the model migrations above, so the ruler and the thing measured don't change in the same week.
Source: OpenAI deprecations
Anthropic's new API version rejects temperature settings
What changed. In Anthropic's API version 2026-10-21, "a request that sends temperature, top_p, or top_k returns a 400 error, even at the parameter's default value." SDKs use the API version current when their major version was released, so the new major SDK versions move you onto it. Requests on the previous version, 2023-06-01, still accept the parameters.
Who it affects. Teams whose dependency bot will offer the SDK major upgrade, especially evaluation harnesses and LLM-as-judge setups that pin a low temperature for repeatable scores.
What to do. Remove the three parameters before merging the upgrade. Then measure grader score variance on a fixed set of cases, before and after; if it grew, sample more than once or widen pass thresholds.
GitHub blocks pull_request_target in public repositories from 2 December
What changed. From 2 December 2026, GitHub Actions blocks pull_request_target events by default in public repositories on GitHub Free, Team and Enterprise Cloud, unless an Actions event policy allows them. The last brownout ran on 16 November. Private and internal repositories and GitHub Enterprise Server are not affected.
Who it affects. Public projects whose CI runs on pull requests from forks.
What to do. Use GitHub's Insights to see which runs the rule would block. Allow only the named workflow files that need the event, not the event in general.
Source: GitHub changelog, 2 November
EU deadlines on 2 and 9 December: AI marking, new bans, product liability
What changed. Under Regulation (EU) 2026/1744, generative AI systems placed on the market before 2 August 2026 must meet the AI Act's Article 50(2) duty to mark output as AI-generated, machine-readably, by 2 December 2026. Two new prohibitions apply the same day, covering generated non-consensual intimate imagery and child sexual abuse material. Member States must transpose the revised Product Liability Directive, (EU) 2024/2853, by 9 December 2026: software, including AI, becomes a product, and courts can order disclosure and presume a defect. On 12 November, EUR-Lex listed national measures from five Member States, including Germany's, published 27 October.
Who it affects. Providers of generative systems on the EU market, and anyone selling software to EU customers.
What to do. Test that marking survives your own export, resize and compression paths, and red-team generators for the prohibited outputs, keeping the results. Keep release, test and patch records as if a court will read them.
The EU's first harmonised AI standard is a quality management standard
What changed. On 14 October 2026, just before this window, the Commission published the reference of EN 18286:2026, Artificial intelligence – Quality management system for EU AI Act regulatory purposes, in the Official Journal (Implementing Decision (EU) 2026/2274). It is the first harmonised standard under the AI Act. Under Article 40, a high-risk system that conforms to it is presumed to meet the requirements it covers, and only those.
Who it affects. Providers of high-risk AI systems, and the quality teams who will run their quality management system.
What to do. Gap-assess your quality management system against EN 18286 while the high-risk obligations are still a year or more away (2 December 2027 for Annex III, 2 August 2028 for Annex I). Record which requirements it covers and which you must evidence another way.
California: chatbot disclosure and a route to a human from 1 January
What changed. AB 1609 (Chapter 733, signed 28 September 2026) bars businesses with more than $500 million in national annual revenue from presenting a customer-service chatbot as human, and requires a clear "not human" disclosure where a reasonable person could be misled. Customers must be able to ask for a human during business hours, with a good-faith effort to connect within 15 minutes. Public prosecutors enforce it: up to $5,000 for a first violation and $10,000 for each after. With no start date in the bill, California's default applies: 1 January 2027.
Who it affects. Large businesses serving Californians through chat or voice agents, and the teams that test them.
What to do. Add test cases for the disclosure and for "I want a person" in every channel, in the phrasings customers use, and measure time to a human.
Source: AB 1609 chaptered text
Coming up
Date | What | Source |
30 November 2026 | Claude Sonnet 4.5 retires on the Claude API; OpenAI Evals and reusable prompts shut down | Anthropic, OpenAI |
1 December 2026 | Gemini 2.5 Pro and Flash requests re-route to newer models | |
2 December 2026 | GitHub blocks pull_request_target in public repositories; EU AI Act marking deadline and new prohibitions | GitHub, Regulation Tracker |
8 December 2026 | Dependabot alerts off on repositories inactive for a year; ARM32 hosted runners end | GitHub |
9 December 2026 | EU Product Liability Directive transposition deadline (software is a product) | Regulation Tracker |
11 December 2026 | OpenAI removes the GPT-5 and o3 snapshots | OpenAI |
14 December 2026 | GitHub Copilot deprecates Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6 and Kimi K3 | GitHub |
18 December 2026 | Comments close on NIST's draft adversarial machine learning taxonomy, which focuses on attacks on agentic systems; GitHub Copilot deprecates MAI-Code-1.1-Flash | NIST, GitHub |
1 January 2027 | California AB 1609 applies; Colorado SB 26-189 duties begin, though a court order currently bars enforcement | Regulation Tracker |
8 January 2027 | Amazon Bedrock: Claude Opus 4.1 end of life | AWS |
11 January 2027 | OpenAI ends service_tier: fast and priority for GPT-5.2, GPT-5.4 and GPT-5.4 mini with EU data residency | OpenAI |
14 January 2027 | OpenAI removes the GPT-4o and GPT-4.1 snapshots | OpenAI |
New on ShiftQuality
Testing Eventual Consistency: What "Eventually" Has to Mean: why a test that waits for convergence can pass for a year while users act inside the lag, and how to test the staleness budget instead.
Where Quality Capability Lives: placing quality practices in stream teams, platforms or enabling engagements, judged by what keeps working when the specialists move on.
The Honest Glossary: the engineering definitions of fifteen AI terms, with a test to run on the next pitch deck.
Career Growth for Engineers: a ten-part series, now complete, from understanding levels to handling rejection.
Sources
Anthropic, Model deprecations, Claude documentation, accessed 19 November 2026. https://platform.claude.com/docs/en/about-claude/model-deprecations
Anthropic, Migrating to API version 2026-10-21, Claude documentation, accessed 19 November 2026. https://platform.claude.com/docs/en/api/migration-2026-10-21
Amazon Web Services, Model lifecycle (Legacy), Amazon Bedrock User Guide, accessed 19 November 2026. https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle-legacy.html
OpenAI, Deprecations, API documentation, accessed 19 November 2026. https://developers.openai.com/api/docs/deprecations
Google, Gemini deprecations, accessed 19 November 2026. https://aistudio.google.com/docs/deprecations
GitHub, pull_request_target blocked by default December 2, changelog, 2 November 2026. https://github.blog/changelog/2026-11-02-pull_request_target-blocked-by-default-december-2
GitHub, Dependabot alerts will pause on inactive repositories, changelog, 10 November 2026. https://github.blog/changelog/2026-11-10-dependabot-alerts-will-pause-on-inactive-repositories
GitHub, Actions for ARM32 Linux & Windows retirement, changelog, 12 November 2026. https://github.blog/changelog/2026-11-12-actions-for-arm32-linux-windows-retirement
GitHub, Upcoming deprecation of selected GitHub Copilot models in December, changelog, 13 November 2026. https://github.blog/changelog/2026-11-13-upcoming-deprecation-of-selected-github-copilot-models-in-december
GitHub, Upcoming deprecation of MAI-Code-1.1-Flash in December, changelog, 17 November 2026. https://github.blog/changelog/2026-11-17-upcoming-deprecation-of-mai-code-1-1-flash-in-december
European Union, Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal. https://eur-lex.europa.eu/eli/reg/2026/1744/oj
European Union, Directive (EU) 2024/2853 on liability for defective products, national transposition measures, EUR-Lex, accessed 12 November 2026. https://eur-lex.europa.eu/legal-content/EN/NIM/?uri=CELEX:32024L2853
European Commission, Implementing Decision (EU) 2026/2274 of 12 October 2026 on the harmonised standards for high-risk AI systems, Official Journal, 14 October 2026. https://eur-lex.europa.eu/eli/dec_impl/2026/2274/oj
California Legislature, AB 1609 (Chapter 733, Statutes of 2026), chaptered text. https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260AB1609
NIST, Adversarial Machine Learning Taxonomy and Terminology, SP 2400-102e2026 initial public draft, 9 November 2026; comments close 18 December 2026. https://csrc.nist.gov/News/2026/adversarial-machine-learning-taxonomy-2026-draft


