by OpenAI
GPT-5.6 Sol medium
$1.36per solved task
Solves 62.3 % of tasks · May 15 – Jul 1, 2026
For OpenCode · Aider · Cline · Hermes · OpenClaw · updated weekly
The cheapest model that's good enough.
Which model should power your coding agent? The cheapest per solved task often solves too few of them, so we measure what one solved GitHub issue really costs and recommend the cheapest model that solves at least 80 % as many tasks as the best one.
Good enough →
Cheapest
MiMo V2.5 Pro
$0.223 per solved task
Solves 46.5 %, below good enough
Our pick
GPT-5.6 Sol [medium]
$1.36 per solved task
Solves 62.3 %, good enough
Best score
Fable 5 [high]
$6.82 per solved task
+2.2 pts, within the margin of error, for 5× the price
Your next 1,000 agent tasks
$3,550 saved with GPT-5.6 Sol
Fable 5's 22 extra fixes cost $161 each, within the margin of error (vs $1.36 per fix with our pick).
Not a fluke: since Aug 2025, the top score has cost a median 3.3× more per solved task than good enough (≥ 2× in 10 of 11 months).
Ranking
Sorted by cost per solved task. Models solving less than 80 % (51.6 %) as many tasks as the best one (or failing the 80 % SWE-bench Verified gate) are greyed out and never recommended. Beaten by X: X solves at least as many tasks for less per solved task.
| Model | Solved rate | Verified | Cost / problem | Cost / solved task | Copy ID |
|---|---|---|---|---|---|
| MiMo V2.5 Pro Below threshold | 46.5 % ± 0.5 | 48.0 % | $0.104 | $0.223 | |
| GPT-5.6 Luna [medium] Beaten by MiMo V2.5 Pro | 43.6 % ± 1.5 | 55.0 % | $0.108 | $0.248 | |
| DeepSeek-V4 Pro [high] Beaten by MiMo V2.5 Pro | 40.2 % ± 1.3 | 52.0 % | $0.149 | $0.371 | |
| Qwen3.6-35B-A3B Beaten by MiMo V2.5 Pro | 24.7 % ± 0.8 | 39.5 % | $0.269 | $1.09 | |
| GPT-5.6 Sol [medium] Our pick | 62.3 % ± 1.8 | 68.2 % | $0.85 | $1.36 | |
| Qwen3.6-27B Beaten by MiMo V2.5 Pro | 31.2 % ± 1.7 | 38.0 % | $0.619 | $1.99 | |
| MiniMax M3 Beaten by GPT-5.6 Sol | 47.2 % ± 1.1 | 45.0 % | $0.955 | $2.02 | |
| GLM-5.2 [high] | 62.9 % ± 1.2 | 60.5 % | $1.40 | $2.23 | |
| Grok 4.5 [high] | 63.8 % ± 0.6 | 64.0 % | $1.47 | $2.31 | |
| Sonnet 5 [high] Beaten by GPT-5.6 Sol | 56.8 % ± 0.9 | 65.4 % | $1.43 | $2.51 | |
| Opus 5 [high] Beaten by Grok 4.5 | 63.4 % ± 1.4 | 60.0 % | $3.47 | $5.46 | |
| Qwen3.5-35B-A3B Beaten by MiMo V2.5 Pro | 17.1 % ± 1.1 | 28.0 % | $0.995 | $5.81 | |
| Fable 5 [high] | 64.5 % ± 1.4 | 64.8 % | $4.40 | $6.82 |
Indicative pick
Gemini-3-Flash Gemini CLI (ACP)
$0.427per solved task
Solves 77.2 % on SWE-bench · $0.33 per attempt
3.5× less per solved task than the top score (OpenHands · claude-fable-5, 95.8 %)
Winning Stacks
Sorted by cost per solved task. Stacks solving less than 80 % (76.6 %) of the best score are folded below.
A different benchmark (SWE-bench) and period than By Model: scores and costs are not comparable between the two tabs.
The bar is set by the top score (95.8 %); the next is 83.8 %.
| Stack | SWE-bench | Cost / attempt | Cost / solved task |
|---|---|---|---|
| Gemini-3-Flash Gemini CLI (ACP) Indicative pick | 77.2 % | $0.33 | $0.427 |
| claude-opus-4-8 OpenHands | 83.8 % | $0.75 | $0.895 |
| claude-opus-4-6 OpenHands | 76.8 % | $0.77 | $1.00 |
| claude-opus-4-6 OpenHands (subagents) | 76.8 % | $0.85 | $1.11 |
| claude-fable-5 OpenHands | 95.8 % | $1.43 | $1.49 |
| GPT-5.5 Codex (ACP) | 78.0 % | $1.17 | $1.50 |
| claude-opus-4-7 OpenHands | 81.6 % | $1.33 | $1.63 |
| Gemini-3.1-Pro Gemini CLI (ACP) | 79.8 % | $1.40 | $1.75 |
| GPT-5.5 OpenHands | 78.2 % | $1.52 | $1.94 |
| Gemini-3.5-Flash OpenHands | 78.6 % | $1.53 | $1.95 |
| Gemini-3.1-Pro OpenHands | 76.8 % | $1.87 | $2.43 |
Show 32 stacks below the threshold (76.6 %)
| Stack | SWE-bench | Cost / attempt | Cost / solved task |
|---|---|---|---|
| DeepSeek-V4-Pro OpenHands | 73.2 % | $0.04 | $0.054 |
| MiniMax-M3 OpenHands | 76.4 % | $0.17 | $0.223 |
| DeepSeek-V3.2-Reasoner OpenHands | 71.6 % | $0.16 | $0.223 |
| MiniMax-M2.7 OpenHands | 75.6 % | $0.179 | $0.237 |
| MiniMax-M2.5 OpenHands | 72.6 % | $0.307 | $0.422 |
| GPT-5.4 Codex (ACP) | 73.8 % | $0.40 | $0.542 |
| Gemini-3-Flash OpenHands | 74.6 % | $0.42 | $0.563 |
| Kimi-K2.5 OpenHands | 68.8 % | $0.406 | $0.591 |
| Nemotron-3-Nano OpenHands | 34.2 % | $0.24 | $0.702 |
| Nemotron-3-Super OpenHands | 62.0 % | $0.466 | $0.752 |
| GLM-4.7 OpenHands | 73.4 % | $0.56 | $0.762 |
| GPT-5.4 OpenHands | 75.6 % | $0.63 | $0.833 |
| Kimi-K2.6 OpenHands | 74.6 % | $0.67 | $0.898 |
| claude-sonnet-4-5 Claude Code (ACP) | 74.4 % | $0.70 | $0.941 |
| MiniMax-M2.1 OpenHands | 68.8 % | $0.712 | $1.03 |
| GPT-5.2 OpenHands | 74.6 % | $0.86 | $1.15 |
| Trinity-Large-Thinking OpenHands | 56.8 % | $0.69 | $1.21 |
| GPT-5.2-Codex OpenHands | 73.8 % | $0.94 | $1.27 |
| Gemini-3-Pro OpenHands | 70.6 % | $0.95 | $1.35 |
| claude-sonnet-4-6 OpenHands | 74.4 % | $1.03 | $1.38 |
| GLM-5 OpenHands | 73.4 % | $1.06 | $1.44 |
| claude-opus-4-7 Claude Code (ACP) | 75.8 % | $1.15 | $1.52 |
| claude-opus-4-6 Claude Code (ACP) | 74.4 % | $1.14 | $1.53 |
| claude-sonnet-4-5 OpenHands | 74.2 % | $1.19 | $1.60 |
| Qwen3-Coder-480B OpenHands | 62.4 % | $1.26 | $2.02 |
| Qwen3-Coder-Next OpenHands | 66.6 % | $1.36 | $2.04 |
| GLM-5.1 OpenHands | 75.0 % | $1.54 | $2.05 |
| Qwen3.6-Plus OpenHands | 74.2 % | $1.54 | $2.07 |
| claude-sonnet-4-5 OpenHands (subagents) | 73.8 % | $1.54 | $2.09 |
| claude-opus-4-5 OpenHands | 76.6 % | $1.82 | $2.38 |
| Kimi-K2-Thinking OpenHands | 69.2 % | $2.00 | $2.89 |
| Qwen3.5-Flash OpenHands | 62.0 % | $2.16 | $3.49 |
Source: OpenHands Index (MIT), SWE-bench results, latest submission Jun 12, 2026. Cost per solved task = cost per attempt ÷ solved rate. Published results, not re-run here: they can be months old.
Method
Cost per solved task = cost per problem ÷ solved rate
- Source: SWE-rebench by Nebius: freshly mined GitHub issues across Python, TypeScript, Go, Java and Rust, 5 runs per model, one fixed agent scaffold for every model. Cost and solved rate come from the same runs.
- Recommendation: the cheapest cost per solved task among models reaching at least 80 % of the best solved rate on SWE-rebench.
- Cross-validation: to guard against single-benchmark overfitting, the recommended model must also reach at least 80 % of the top score on SWE-bench Verified.
- Only models, not products: agent products (Claude Code, Codex, Cursor…) are excluded.
Limits
- Your harness is not SWE-rebench's scaffold: absolute costs will differ, the ordering usually holds.
- About 111 problems per window: gaps of 1–2 points (± shown) are not significant.
- Costs reflect API prices during the evaluation window.