Signature bake

Which coding agent to pick, based on unpublished real-world tasks.

Agents: 10; in Claude Code, Codex and pi
Tasks: 30; 20 bug fixes, 10 features
Runs: 570; 438 solved
Last updated: 9 Nov 2026

The board

Harness

Rank	Agent
1	Opus 5.5Claude Code · medium effort	53 / 6088%	$1.04	2m12s	93
2	Fable 5.21 run a taskClaude Code · medium effort	26 / 3087%	$1.78	1m51s	160
3	GPT-6 AstraCodex · low effort	50 / 6083%	$1.76	2m16s	199
3	Fable 5.1Claude Code · high effort	50 / 6083%	$3.14	3m12s	142
5	GPT-6.1 TerraCodex · low effort	48 / 6080%	$0.10	1m01s	88
5	GPT-6.1 SolCodex · low effort	48 / 6080%	$0.28	1m55s	148
7	Haiku 5.5Claude Code · medium effort	45 / 6075%	$0.046	1m32s	58
8	Sonnet 5.5Claude Code · medium effort	44 / 6073%	$0.37	43s	37
9	GLM-5.3 Flashpi · high effort	41 / 6068%	$0.078	1m49s	138
10	GPT-6 LunaCodex · medium effort	33 / 6055%	$0.025	43s	32

Click a column to sort by it.

What a fix costs

Solved against

60%70%80%90%$0.03$0.10$0.30$1$3Opus 5.5 Fable 5.2 1 run a task GPT-6 Astra Fable 5.1 GPT-6.1 Terra GPT-6.1 Sol Haiku 5.5 Sonnet 5.5 GLM-5.3 Flash GPT-6 Luna

Share of runs solved, against cost per fix on a log scale. Up and to the left is better. The line is the efficient frontier: no agent solves more for less. Click an agent to open its page.

What the board says today

Pick the cheapest agent that fits your work. Six agents solve 80-90% of tasks, with a 30x difference in prices.

A fix costs $0.10 from GPT-6.1 Terra (80%) and $3.14 from Fable 5.1 (83%).

For routine work, agents are all quite good - they solved a third of the tasks every time.

10 of the 30 tasks were solved by all 19 runs, including the agent that costs $0.025 a fix and left you the least code to review.

See every run on every task →

If you're not sure about the complexity of your work, start with Opus 5.5.

It solved the most, 88% of runs, at $1.04 a fix, and every agent that costs more solved less.

See Opus 5.5's page →

570 runs between 1 and 9 Nov 2026 · How the Signature bake is run