What the change log recorded in the first quarter of 2026

Nine entries went into the change log for the first quarter of 2026, across OpenAI, Anthropic and Google. Five carry the action rerun, two carry re-prompt, one carries re-pin and one carries migrate. This post is the narrative version of that list, because a table of forty entries does not show you the shape.

Every count here is a count of our own log. It is not a measurement of any model, and nothing below says a model got better or worse.

The one that needed no deploy

Start with the re-pin, because it is the one that happens without you.

On 21 January 2026 Google changed the latest aliases to point to Gemini 3 models. Anyone calling an alias was served something new with no code change, which is behaviour Google documents: a Gemini "latest" alias points to the most recent release of a model variation and is hot-swapped with every new release of that variation.[Google] Google's own documentation also says stable models usually don't change and that most production apps should use a specific stable model.[Google]

One entry out of nine, and the only one in the quarter that could reach production without anybody approving it. That ratio is why the rerun protocol keeps a weekly scheduled run rather than waiting for announcements.

The two that broke caller code

Re-prompt entries are the ones where the request contract moved.

On 29 January 2026 Anthropic took structured outputs out of beta on the Claude API for Claude Sonnet 4.5, Claude Opus 4.5 and Claude Haiku 4.5, and the notes state that the output_format parameter moved to output_config.format, while structured outputs remain in public beta on Amazon Bedrock and Microsoft Foundry. A renamed parameter is a deploy, and the partner-platform difference is the part teams miss: the same feature on two platforms is now two code paths.

On 5 February 2026 Anthropic launched claude-opus-4-6, with notes stating that adaptive thinking is recommended and manual thinking is deprecated, that an effort parameter with levels low, medium, high, xhigh and max launched, that a 1M token context window is available in beta with long context pricing above 200k input tokens, and that the media limit rose from 100 to 600 images or PDF pages with the 1M context. If you had hand-tuned a thinking budget, that tuning is now expressed differently, and the cases you tuned it on are the cases to rerun.

The five that were an invitation, not an emergency

Most announcements are a new model you could adopt. The action is a comparison run against what you serve today, and nothing else until the run comes back.

• 3 February 2026, OpenAI optimized its inference stack and states that gpt-5.2 and gpt-5.2-codex now run about 40 percent faster with unchanged weights. Faster changes timeout and retry behaviour, so the rerun here is as much about your timing logic as your scores.
• 17 February 2026, Anthropic launched claude-sonnet-4-6 with notes stating that web search and web fetch are free when used with code execution, and that several tools are available without a beta header. Removing a header can change which tool path a request takes.
• 19 February 2026, Google released gemini-3.1-pro-preview and a gemini-3.1-pro-preview-customtools variant. We keep the custom tools variant in a separate run so tool behaviour differences stay visible.
• 5 March 2026, OpenAI released GPT-5.4 with a 1M token context window, tool search, computer use and compaction, alongside gpt-5.4-pro. Compaction changes what the model still has in front of it, which is a retrieval (the step that fetches documents or records and puts them into the prompt) question as much as a model one.
• 13 March 2026, Anthropic took the 1M token context window out of beta for Claude Opus 4.6 and Claude Sonnet 4.6, at standard pricing, with the dedicated 1M rate limits removed and the media limit raised to 600 images or PDF pages per request.

Read as a group, four of those five touch context, tools or timing rather than answer quality. That is the useful pattern in the quarter: the changes that reach a working feature are mostly about plumbing.

The one with a deadline

On 9 March 2026 Google shut down Gemini 3 Pro Preview. Migrate is the only action with a date attached, which is why it is the only one that belongs in a calendar rather than a backlog.

The warning you get depends on whose platform you are on, and each provider publishes its own commitment. OpenAI says at least 6 months[src] for a generally available model, at least 3 months for specialized variants, and that preview models may be retired with much shorter notice, such as 2 weeks, with no minimum committed.[OpenAI] Anthropic describes its model lifecycle in four states, active, legacy, deprecated and retired, and says requests to retired models will fail.[Anthropic] Three companies, three separate commitments, no shared norm.

What the count does not include

An announcement is a floor, not a ceiling. Future AGI writes that OpenAI, Anthropic, Google and most frontier providers update model weights without a major version bump, which is Future AGI's characterisation of the market rather than a provider statement.[Future AGI] We publish no entry without a provider source, so an unannounced change is not in the nine, by construction.

That is the honest limit of a change log, and it is also why the log is not the product. The log tells you what to run. The run tells you what moved, read against thresholds you agreed, which is covered in EV-THRESH and reported monthly under EV-REPORT. If you want the per-provider slices, they are at OpenAI, Anthropic and Google.