WorkloadBenchFollow @BrenBuilds on X

GPT-6.1 Sol

GPT-6.1 Sol's own checks, newest first. Its results on the tests every model ran are on the Quick suite and Deep suite pages.

Rechecked after posts said it got better, 12 October 2026#

The same tests run again, beside its earlier runs, are on its recheck page.

Speed on the explain task, 29 September to 6 October 2026#

Its explain-task speed in eleven batches, launch day included, is on its speed page. A batch is one set of runs made together.

Launch day on the explain task, 29 September 2026

These ran before 30 September, with each CLI's defaults left on, so their numbers differ from the Quick suite's explain task table.

GPT-6.1 Sol against GPT-6 Sol and Opus 5.5#

GPT-6.1 Sol ran on 29 September 2026, the day it came out. On this task it reported no hidden thinking at all.

GPT-6.1 Sol, GPT-6 Sol and Opus 5.5 on the explain taskEach bar is the median of 2 runs, all on the same day. The GPT models ran in Codex CLI 0.159.0 at medium effort, and Opus 5.5 in Claude Code at its default medium effort.

Total tokens#

Everything the model read and wrote in the whole run

GPT-6.1 Sol: 49,053 tokens, median of 2 runsGPT-6.1 Sol49.1kGPT-6 Sol: 40,175 tokens, median of 2 runsGPT-6 Sol40.2kOpus 5.5: 46,516 tokens, median of 2 runs, mostly light runsOpus 5.546.5k*

Cost at list prices#

What the run's tokens cost at the model's published list prices, cached tokens at the cache price. GPT bars are the lower bound: Codex logs 0 cache writes on every run, and OpenAI prices cache writes up to a quarter above input

GPT-6.1 Sol: $0.0366 at list prices, median of 2 runsGPT-6.1 Sol$0.0366GPT-6 Sol: $0.0427 at list prices, median of 2 runsGPT-6 Sol$0.0427Opus 5.5: $0.150 at list prices, median of 2 runs, mostly light runsOpus 5.5$0.150*

Tokens read#

All the text sent to the model, counted each time it was sent, with Codex CLI's hidden first request

GPT-6.1 Sol: 48,786 tokens, median of 2 runsGPT-6.1 Sol48.8kGPT-6 Sol: 39,896 tokens, median of 2 runsGPT-6 Sol39.9kOpus 5.5: 44,767 tokens, median of 2 runs, mostly light runsOpus 5.544.8k*

Read from cache#

Share of what it read that came from the cache, which is cheaper. Counts Codex CLI's hidden first request

GPT-6.1 Sol: 68% of prompt tokens, median of 2 runsGPT-6.1 Sol68%GPT-6 Sol: 56% of prompt tokens, median of 2 runsGPT-6 Sol56%Opus 5.5: 70% of prompt tokens, median of 2 runs, mostly light runsOpus 5.570%*

Cache writes#

Text saved to the cache for reuse. Codex CLI always reports 0, so for Codex CLI I count what it read that did not come from the cache

GPT-6.1 Sol: 15,314 tokens, median of 2 runsGPT-6.1 Sol15.3kGPT-6 Sol: 17,752 tokens, median of 2 runsGPT-6 Sol17.8kOpus 5.5: 13,649 tokens, median of 2 runs, mostly light runsOpus 5.513.6k*

Output tokens#

Everything it wrote: the answer, its tool calls and its hidden thinking

GPT-6.1 Sol: 267 tokens, median of 2 runsGPT-6.1 Sol267GPT-6 Sol: 279 tokens, median of 2 runsGPT-6 Sol279Opus 5.5: 1,750 tokens, median of 2 runs, mostly light runsOpus 5.51.7k*

Hidden thinking#

The part of the output spent thinking, which the answer doesn't show

GPT-6.1 Sol: 0 tokens, median of 2 runsGPT-6.1 Sol0GPT-6 Sol: 80 tokens, median of 2 runsGPT-6 Sol80Opus 5.5: 502 tokens, median of 2 runs, mostly light runsOpus 5.5502*

* At least half the runs behind a starred number are light. Light runs start with fewer tokens of instructions and tools than the model's other runs, so their counts differ. What light runs are.

The table scrolls sideways on a small screen.

Same day, explain task, no skill, medians
ModelPassedTokens readFrom cacheCache writesOutputThinkingVisibleCost at list prices
GPT-6.1 SolCodex CLI 0.159.0, medium effort, set by me2 of 248,78668%15,3142670267$0.0366 to $0.0443
GPT-6 SolCodex CLI 0.159.0, medium effort, set by me2 of 239,89656%17,75227980199$0.0427 to $0.0516
Claude Opus 5.5Claude Code 2.1.284, medium effort, Claude Code's default2 of 244,767*70%*13,649*1,750*502*1,248*$0.150*