Short answer: OpenAI reported that GPT-5 scored 74.9% on SWE-bench Verified and 88% on Aider Polyglot at its August 7, 2025 launch. These are results for GPT-5—not GPT-5.2. They are useful reference points, but they come from different tests and do not prove how well the model will perform on your repository or workflow. More importantly for a production decision, OpenAI’s current API documentation marks GPT-5 as deprecated and recommends GPT-6 Astra instead. (openai.com)
GPT-5 coding benchmark results
| Benchmark | Reported GPT-5 result | What the benchmark tests | Important qualification |
|---|---|---|---|
| SWE-bench Verified | 74.9% | Resolving real GitHub software issues by producing patches that pass tests. Verified is a human-reviewed subset of 500 SWE-bench tasks. | OpenAI’s announcement reports the score but the captured announcement does not specify the exact agent/scaffold setup behind it. Different leaderboard agent setups are not automatically comparable. (openai.com) |
| Aider Polyglot | 88.0% | Completing 225 challenging Exercism coding exercises across C++, Go, Java, JavaScript, Python, and Rust, using Aider’s code-editing workflow. | Aider’s leaderboard labels the result GPT-5 (high) and reports 91.6% correct edit format. This is a score from a particular benchmark configuration, not a general measure of production engineering quality. (aider.chat) |
OpenAI published the two scores together as GPT-5’s launch results. They measure different things: SWE-bench Verified focuses on resolving existing software issues, while Aider Polyglot tests coding exercises and successful code editing across several languages. Don’t average the percentages or read them as a single “coding score.” (openai.com)
How to interpret the scores
- Treat each number as benchmark-specific. Aider explains that its “percent completed correctly” measures the share of coding tasks successfully completed, including solving the exercise and editing code to implement the solution. It also tracks whether the model followed the requested edit format—a related but separate measure. (
aider.chat) - Keep model versions and settings separate. The GPT-5 results above should not be replaced with GPT-5.2 results. Likewise, a result for GPT-5 at high reasoning effort is not equivalent to a result at a different effort setting. The Aider leaderboard, for example, displays GPT-5 high at 88.0% and medium at 86.7%. (
aider.chat) - Do not infer a universal winner. A benchmark score does not establish performance on your languages, framework versions, private codebase, review standards, or tool setup. The supplied results also do not provide a same-conditions comparison against every current alternative.
For context, a separate article reports 55.6% for GPT-5.2 Thinking on SWE-Bench Pro, compared with 52.0% for Claude Opus 4.5 and 43.3% for Gemini 3 Pro. That is a different model version and benchmark variant; it is not a GPT-5 result and should not be mixed with the GPT-5 scores above. (openai.com)
Is GPT-5 a sensible production API choice now?
Not as a new default choice based on these benchmark results alone. As of November 12, 2026, OpenAI’s API model page labels GPT-5 deprecated, identifies the dated snapshot as gpt-5-2025-08-07, and recommends GPT-6 Astra as the latest model. A benchmark score can remain useful historical evidence even when the model is no longer the recommended production target; it does not show that the same model is currently available for a new deployment. (developers.openai.com)
For a coding-model decision, use the result as a starting point, then:
- Check the intended API model’s status and exact identifier in the provider’s current documentation before building against it.
- Choose representative tasks from your own work: for example, bug fixes, tests, refactors, dependency upgrades, and multi-file changes in the repositories and languages you actually use.
- Keep the comparison controlled: same task set, tools, context supplied, retry rules, reasoning settings, and acceptance tests for each candidate.
- Measure outcomes that matter to the team: tests passed, correctness after review, regressions, human editing time, latency, and total API cost. The benchmark figures do not establish those results for your workload.
- Have engineers review generated changes before they are merged or deployed. Benchmark performance is not a substitute for code review or security checks.
The captured research does not establish a current production API setup for a non-deprecated GPT-5 model, nor does it provide workload-specific GPT-5 results. If GPT-5 itself is required for a legacy comparison, first confirm access and lifecycle details with the provider; for a new production selection, evaluate a currently supported model instead.
Bottom line
GPT-5’s reported launch scores were 74.9% on SWE-bench Verified and 88% on Aider Polyglot. They are meaningful, version-specific benchmark results—not proof of universal coding leadership or production fit. Since OpenAI now marks GPT-5 deprecated in its API documentation, teams choosing a model for a new production integration should verify the supported model first and run a controlled evaluation on their own coding tasks. (openai.com)