On 2026-03-06 a coordinated wave shipped — 725 verified tests passing, zero failures

A single rollout day landed GPT-5.4 across the ecosystem in 9 merged PRs spanning 6 repos, backed by test counts that sum to 725 with no reported failures.

So where did all that change actually go? Six repos.

Provider code, routing, defaults, and docs — all moved at once

The rollout spanned six named repos, so a single model change touched every layer that has to agree on which model to run.

First, provider-openai had to get the model's behavior right.

A new ModelCapabilities architecture fixed the context window and added reasoning levels

provider-openai's PR #27 corrected the context window from a hardcoded 400K to GPT-5.4's real 272K and added xhigh and none reasoning-effort levels — not just the name, the behavior.

But getting a model right this way is fragile.

A reasoning-capable model with default effort None can silently lose reasoning tokens

When default_reasoning_effort is None, the include=[reasoning.encrypted_content] param is still required — miss it and reasoning tokens vanish silently. A regression fix (#28) caught it the next day.

Discipline like this is why the next drop can be made easy.

A version-aware glob makes the resolver auto-pick the highest model — no matrix edit needed

The OpenAI matrices switched to a gpt-[0-9].[0-9] glob whose version-aware sort picks the highest available version automatically, so the next model drop needs no routing-matrix edit.

Then GPT-5.5 landed and put the design to the test.

When GPT-5.5 landed it was a near-mechanical swap — 248 routing-hook tests passed, no changes needed

The glob proved itself on the very next bump: commit #17 lifted the mid- and heavy-tier OpenAI models to gpt-5.5 with 248 routing-hook tests passing and no changes needed.

So what does this mean for you as a user?

Pull latest and get the right model at the right reasoning level — no config, and each drop gets easier

GPT-5.4 is the default across the ecosystem — foundation's provider bundles all set default_model: gpt-5.4 and the matrices resolve OpenAI roles automatically. Copilot stays pinned by design, since its provider has no list_models fallback.