SWE-bench Leaderboard: Comparing GPT-5.2 Coding Agents

One-sentence takeaway

In SWE-bench's controlled mini-SWE-agent table, the GPT-5.2 high-reasoning entries reach 71.8%–72.8% resolved at about $0.47–$0.52 per task, but these are GPT-5.2 family/Agent harness results, not a direct score for a Chat snapshot.

Test environment

• Task set: Real GitHub issues; the original benchmark comes from 12 Python repositories, while the page also notes that expanded tasks cover 42 repositories and 9 languages.

• Agent: mini-SWE-agent; the page distinguishes between different release/harness versions.

• Output metrics: % Resolved is the percentage of task instances resolved; Avg. $ is the average cost. The page also lists model, date, and trajectory columns.

• Reproduction status: The page marks some entries as run or verified directly by the SWE-bench team.

Inputs/configuration

GPT-5.2 entries on the page:

• GPT 5.2 high, mini-SWE-agent 2.0.0, 2026-02-17: 72.80%, $0.47/task;

• GPT 5.2 high, mini-SWE-agent 1.17.2, 2025-12-11: 71.80%, $0.52/task;

• GPT 5.2, mini-SWE-agent 1.17.2, 2025-12-11: 69.00%, $0.27/task;

• GPT 5.2 Codex, mini-SWE-agent 2.0.0, 2026-02-19: 72.80%, $0.45/task.

Results data

At the time of collection, GPT 5.2 high (2.0.0 harness) and GPT 5.2 Codex both scored 72.80%; GPT 5.2 high (1.17.2) scored 71.80%, while GPT 5.2 without an explicit high designation scored 69.00%. In the same table, Claude 4.5 Sonnet high scored 71.40%, Kimi K2.5 high scored 70.80%, and DeepSeek V3.2 high scored 70.00%, providing relative reference points under the same harness.

Conclusion

GPT-5.2 is a strong coding baseline for real-issue-fixing Agent tasks, but scores are sensitive to the reasoning tier and mini-SWE-agent version. After upgrading the harness, costs and success rates must be recorded again.

Limitations

• The “GPT 5.2” in the SWE-bench table is not explicitly identified as the API gpt-5.2-chat-latest; this result cannot be treated as equivalent to a Chat Completions snapshot.

• This is an issue-resolution metric and does not cover Chat tasks such as design, communication, code review, or long-context question answering.

• The page provides average cost and resolved rate, but these are not the actual bill for each project; the complete prompt, failure trajectories, and confidence intervals are missing.

• The 2.0.0 harness from 2026-02 and the 1.17.2 harness from 2025-12 are not the same harness; cross-row comparisons must identify the version.

Reproduction steps

1. Fix the SWE-bench version, task split, model snapshot, reasoning effort, and mini-SWE-agent release.

2. Run all tasks under the same API rate limits, timeout, and patch-submission strategy; do not combine Chat and Thinking in one group.

3. Save each patch, test log, call trace, token count, elapsed time, and cost; classify results as resolved, error, or timeout.

4. Re-run with the reference models using the same harness as on the page, report the mean resolved rate and cost, and state whether the model alias is a Chat snapshot.