Short answer: GPT-5 had notable coding results when OpenAI introduced it, but OpenAI’s API model reference marks it deprecated. For complex coding work through the API, one documented candidate to evaluate is gpt-6-astra. Its model page describes it for complex coding and lists tools including hosted shell, code interpreter, and apply patch when using the Responses API. That makes it a concrete candidate to test—not proof that it will outperform another model on your codebase. In ChatGPT, the available options and access conditions differ from the API. (developers.openai.com)
What does “ChatGPT 5 coding” mean?
It can refer to two different ways of working:
- Coding in ChatGPT: You use the ChatGPT interface and choose among the models and thinking levels available to your plan and workspace.
- Coding through the OpenAI API: Your application sends requests to a specific API model ID, such as
gpt-6-astra. A ChatGPT model label is not an API model ID.
That distinction matters: the API model page lists gpt-6-astra as the model ID and supports it on both the Responses API and Chat Completions. The page lists function calling as supported, and lists tools such as hosted shell, code interpreter, and apply patch for use with the Responses API. (developers.openai.com)
How good was GPT-5 at coding?
In its GPT-5 launch post, OpenAI reported scores of 74.9% on SWE-bench Verified and 88% on Aider Polyglot. These are historical, vendor-reported results—not a current ranking or a prediction of how GPT-5 or another model will perform on your codebase. The supplied material does not provide enough evaluation setup to support a detailed comparison with current models.
A secondary article also described GPT-5 as useful for front-end generation, multi-file scaffolding, and repository debugging. Its captured text included no test details or benchmark evidence, so those task-specific descriptions should be treated as unverified claims, not independent proof. GPT-5’s historical results do not override the API reference’s deprecation status.
Which model should you choose for coding now?
| Your situation | Documented option to shortlist | Relevant conditions |
|---|---|---|
| Complex coding through the API, especially work involving multiple steps or repository tools | gpt-6-astra |
OpenAI describes it for complex coding. The model page lists Responses API and Chat Completions endpoints; hosted shell, code interpreter, and apply patch are listed among its Responses API tools. Evaluate it on your own tasks rather than assuming a benchmark or product description proves fit. (developers.openai.com) |
| Coding in ChatGPT, if GPT-6 is available for your account | GPT-6 | OpenAI’s Help Center says GPT-6 is rolling out to eligible Plus, Pro, Business, and Enterprise users. If it does not appear in your model menu, it may not yet be available for your account. Availability and thinking levels depend on plan and workspace settings. (help.openai.com) |
| You specifically want the Astra-powered Pro option in ChatGPT | GPT-6 Pro | The Help Center says it is available on Pro $100, Pro $200, Business, and Enterprise plans; Enterprise access also depends on workspace permissions. Plus includes GPT-6 Astra in ChatGPT Work and Codex. These are ChatGPT access routes, not API model IDs. (help.openai.com) |
| You have an existing GPT-5 API integration | Review its deprecation status and shortlist a currently documented option, such as gpt-6-astra, for a bounded migration evaluation. |
Confirm the API model ID, tools, and endpoint you will use in the model documentation, then test against your own workflow before switching. |
These are workload-based shortlisting suggestions, not claims that one option has been tested or shown to outperform another on your tasks. ChatGPT access and API access are separate; do not assume that a ChatGPT subscription provides API usage or that a ChatGPT model label is a usable API ID.
How to evaluate a coding model before adopting it
A published benchmark cannot establish whether a model fits your codebase, tools, and review process. Use a small, repeatable evaluation:
- Choose representative tasks. Include a bug fix, a change spanning multiple files, a test-writing or debugging task, and—if relevant—a small UI feature.
- Keep the setup consistent. Give each candidate the same task description, repository state, tests, tool access, and time or token budget. Record the exact model ID and settings so you can repeat the comparison.
- Check the work. Track whether tests pass, whether the change meets requirements, whether it introduces regressions, and how much review or rework it needs.
- Check operational fit. For API use, confirm the endpoint and tools you need are supported. Consult current documentation for price and limits, and measure your own latency and usage rather than assuming a benchmark predicts them.
- Keep a human review gate. Review generated code and run relevant tests before merging or deploying. This is a proposed engineering practice, not a reported test result for a particular model.
For example, if a model proposes a bug fix but leaves failing tests or needs extensive rewriting, that is a weak result for your workflow—even if its benchmark scores look strong. This evaluation method is a practical proposal, not a test performed for this article.
Bottom line
GPT-5 had notable coding benchmark results at launch, but OpenAI’s API reference marks it deprecated. For complex API-based coding work, gpt-6-astra is a documented candidate: its model page lists coding use, function calling, and Responses API tools including hosted shell, code interpreter, and apply patch. In ChatGPT, GPT-6 and GPT-6 Pro have different access conditions; Plus also includes Astra in ChatGPT Work and Codex. Shortlist based on where you work, then evaluate the candidate on representative tasks with tests and human review. (developers.openai.com)
Sources
GPT-6 Astra Model — OpenAI API
GPT-6 and other models in ChatGPT — OpenAI Help Center