SWE Atlas: Test Writing leaderboard
- Claude Fable 5.1 leads
- 67.0%
- models measured
- 18
- lead over #2
- +4.8 pts
- median
- 42.0%
Leader on SWE Atlas: Test Writing · read October 22
# SWE Atlas: Test Writing, read October 22
leaders = {
"Claude Fable 5.1": 67.0%, "Claude Opus 5": 62.2%, "Claude Fable 5": 55.6%,
}
On October 22, 2026, 18 of our catalog's models had a SWE Atlas: Test Writing result, and Anthropic Claude Fable 5.1 led with 67.0%, ahead of Anthropic Claude Opus 5 at 62.2%.
SWE Atlas Test Writing tests whether a coding agent can write production-grade tests for specific behaviors in real repositories. Scale AI runs each model inside an agent harness; we show the best result per model and name the harness.
Role in the index
A coding board we show beside the Vibe Coding Index, not inside it: the index keeps the six boards it was defined on, so a score does not move every time a board is added. Read the methodology
| # | Model | Result | Evidence | Coding index | Output $/Mtok |
|---|---|---|---|---|---|
| 1 | Claude Fable 5.1Anthropic | 67.0% | 84.2 | $50.00 | |
| 2 | Claude Opus 5Anthropic | 62.2% | 83.0 | $25.00 | |
| 3 | Claude Fable 5Anthropic | 55.6% | 79.3 | $50.00 | |
| 4 | Gemini 3.8 FlashGoogle | 53.7% | 61.6 | $3.75 | |
| 5 | GPT-6 AstraOpenAI | 50.7% | 85.1 | $50.00 | |
| 6 | Claude Opus 4.8Anthropic | 49.6% | 65.0 | $25.00 | |
| 7 | GPT-5.6 SolOpenAI | 45.9% | 73.5 | $20.00 | |
| 8 | GPT-5.4OpenAI | 44.4% | 64.4 | $15.00 | |
| 9 | GPT-5.5OpenAI | 42.6% | 71.3 | $30.00 | |
| 10 | Muse Spark 1.1Meta | 41.5% | 70.3 | $4.25 | |
| 11 | GLM 5.2Z.ai | 41.5% | 61.1 | $4.40 | |
| 12 | Claude Opus 4.7Anthropic | 38.5% | 67.5 | $25.00 | |
| 13 | Claude Opus 4.6Anthropic | 36.7% | 57.2 | $25.00 | |
| 14 | Claude Sonnet 4.6Anthropic | 31.8% | 57.3 | $15.00 | |
| 15 | GLM 5Z.ai | 28.7% | 41.7 | $3.20 | |
| 16 | DeepSeek V4 Pro 0423DeepSeek | 27.1% | 38.7 | $2.60 | |
| 17 | Kimi K2.5Moonshot AI | 25.8% | 38.6 | $2.85 | |
| 18 | MiniMax M2.5MiniMax | 18.6% | 37.2 | $1.20 |
Only models in our catalog are listed, and each appears once with its best published run. 66 rows were read from the source on October 22, 2026, and rows that match no catalog model are left out.