Top AI Models Solve 64% of Real Software Bugs on First Try, 81% With Five Attempts

The headline of SWE-rebench this week says Fable 5 leads the field at 64.5% on the first try. Two columns over, GLM-5.2 [high] reaches 81% with five attempts. The gap between those two numbers is the story.

SWE-rebench, a public benchmark of real GitHub bug fixes, has 111 problems pulled from 65 repositories across Go, Java, Python, Rust, and TypeScript. The current table has outgrown the source title, with roughly 17 evaluated entries alongside a long N/A tail rather than the 13-models-and-4-agents framing that still circulates on Hacker News. The top of the resolved-rate column reads Fable 5 [high] at 64.5%, Grok 4.5 [high] at 63.8%, Claude Opus 5 [high] at 63.4%, GLM-5.2 [high] at 62.9%, and GPT-5.6 Sol [medium] at 62.3%. That column measures the share of problems a model solves in a single attempt under the operator's setup, and it is the number most public coverage of the leaderboard will print.

The Pass@5 column, the share of problems a model solves when given up to five attempts, tells a different story. [GLM-5.2 [high] reaches 81.1% at Pass@5](https://swe-rebench.com), a gap of 18.2 points over its own resolved rate. Fable 5 [high], Grok 4.5 [high], and Opus 5 [high] post comparable headroom. The top of the leaderboard is not five models running away with the field; it is five models each capable of recovering roughly one in five of the bugs they miss on the first try. For a developer routing a coding agent through a build pipeline, that is the difference between a model that fails closed and a model that recovers under retries.

The cost column reshapes the ranking again. The cheapest model in the current window, MiMo V2.5 Pro, solves 46.5% of problems for $0.10 each. Fable 5 [high] at 64.5% resolved costs $4.40 per problem, a 44x spread. The cheapest coding agent on the table, Cursor, is $0.41 per problem at 51.7% resolved. Junie, Claude Code, and Codex sit between 58% and 62% resolved with cost-per-problem in the middle of the spread. A team picking on resolved rate alone overpays for the top of the column; a team picking on cost alone over-buys failure.

Token efficiency widens the variance further. [GPT-5.6 Luna [medium] uses 395,522 tokens per problem](https://swe-rebench.com) at 43.6% resolved. M3, a smaller entry near the bottom of the table, burns 13,869,459 tokens per problem at 47.2% resolved. That is a 35x spread in compute for a 3.6-point gap in success. Two models can hit similar success rates while one is being charged like a long-context research run and the other is being charged like a code-completion call.

The four agent entries tell a second story. Junie 61.8%, Claude Code 60.4%, Codex 58.0%, and Cursor 51.7% sit just under the strongest raw models and well above the long tail. The gap between Claude Code (the agent, 7th in the resolved-rate column) and Claude Opus 5 (the model it wraps, 3rd) is the kind of detail a buyer comparing vendors will want to see, and the Hacker News thread on the leaderboard is full of practitioners asking the same question.

Three caveats belong to the body, not the lede. The leaderboard is operator-published with vendor-submitted runs and no independent reruns in the public record, so the deltas are reported, not adjudicated. SWE-rebench's own header flags "potential contamination," "external system," and "beyond eval range" for several rows. The current window also has a long list of N/A entries: every Claude Opus 4.x, every Sonnet 3.5/4/4.5/4.6, the DeepSeek R1/V3/V3.1/V3.2/V4 Flash family, and Devstral 2, Small-2, and Small-2505. The top of the column is provisional. The methodology and decontamination pipeline are described in the SWE-rebench paper; the lineage to SWE-bench is documented in the foundational paper.

For a team choosing a coding assistant in mid-2026, the resolved-rate column is the least interesting of the three. The Pass@5 column is the reliability question: how much of the work the model misses on the first try can a retry loop recover, and what does that retry cost? The leaderboard's three columns now sit side by side: 62.9% to 81.1% on the same model, $0.10 to $4.40 across the field, 395K to 13.9M tokens on the same problem set. A buyer who reads only the first column is reading a third of the table.