History

1. 22 October 2026

   Claude Opus 5.5: thinking on two Claude Code releases

   Sprint tickets: every run on Claude Code 2.1.308 thought less than every run on 2.1.284, 19% less in the typical run. Bug hunt: the two releases overlapped. Same hidden-test scores in the typical run. One bug-hunt run on 2.1.308 fixed 19 of 20 bugs.

2. 21 October 2026

   Claude Opus 5.5, rechecked after posts said it got worse

   Same test scores as before. It thought about as much as on 29 September, when the account extras were on; on 30 September, with them off as in the recheck, it had thought 2 to 3 times as much. One bug-hunt run added tests at the end of an existing test file, which the gate counts as editing it.

3. 16 October 2026

   Muse Wonder on the quick suite

   A typical tickets-test run cost an estimated $0.748. All 81 of its runs passed.

4. 13 October 2026

   Gemini 4 Argon on its release day

   Typical explain-task runs in the three-model batch: GPT-6.1 Sol 10.6 s, Gemini 4 Argon 11.4 s, Claude Opus 5.5 15.9 s.

5. 12 October 2026

   GPT-6.1 Sol, rechecked after posts said it got better

   Same test scores as before. Every measure on real app work overlapped its earlier range.

6. 7 October 2026

   Claude Haiku 5.5 vs GPT-6 Luna

   Typical explain-task runs: Haiku 5.5 7.3 s and GPT-6 Luna 7.6 s at medium; 15.2 s and 9.0 s at xhigh.

7. 29 September to 6 October 2026

   GPT-6.1 Sol, timed day by day

   Every GPT-6.1 Sol run from 5 October's retest to 6 October streamed more text chunks per second than every run in the eight earlier batches. Five of the twelve runs from 5 October's retest to 6 October finished inside launch day's 10.7 to 12.8 s; one finished sooner and six took longer.

Thinking is the reasoning tokens each tool reports. Hidden tests are tests the model never sees. Typical means the middle run, or the midpoint of the two middle runs. A batch is one set of runs made together. Account extras are what a subscription login loads by default: the plugins, skills and connectors in Claude Code, the connected apps and plugins in Codex CLI. Streamed text chunks per second are a stand-in for tokens per second. A run passes its gate when every test passes, its own new ones included, and no existing test file was changed, renamed or deleted; new test files are fine.

The launch-day runs of Claude Sonnet 5.5 on 28 September and GPT-6.1 Sol on 29 September are in the Quick suite page's method.