Best ChatGPT for coding: the short answer
For coding questions in regular ChatGPT, start with GPT-6 at Medium thinking if it’s available to you; try High or Extra High for harder reasoning tasks when those options appear. For hands-on work across a software project, Codex with GPT-6.1 Sol is a practical starting point if your plan and workspace provide access. Consider Astra for especially demanding work—but in Codex, compare it against your included usage allowance, not API token prices.
These are workload-based starting points, not a claim that one model wins every coding task. ChatGPT’s model name and thinking setting are not necessarily the same as a selectable model in Codex: OpenAI says GPT-6 Sol powers several GPT-6 thinking levels in regular Chat, while GPT-6.1 Sol is a model for Work and Codex and is not available in regular Chat conversations. OpenAI describes GPT-6.1 Sol for complex coding and GPT-6 Astra as its most capable model for demanding work. (ChatGPT models and thinking levels)
Choose a coding model by task
| What you’re doing | Where to start | When to choose differently |
|---|---|---|
| Asking about a snippet, explaining an error, or drafting a small function | GPT-6 in ChatGPT at Medium thinking, if available | Try High or Extra High for more demanding questions if your model menu offers them. These are thinking settings, not separate model names. (ChatGPT models and thinking levels) |
| Implementing a feature or debugging across project files | Codex with GPT-6.1 Sol, if available in your plan and workspace | Consider Astra for a task where stronger capability may justify drawing more on your Work and Codex allowance. Compare outcomes rather than API rates. (ChatGPT Work and Codex) |
| Generating code in your own application through an API | GPT-6 Astra (gpt-6-astra) is OpenAI’s suggested starting point for complex coding |
Compare GPT-6.1 Sol (gpt-6.1-sol) for complex work at lower standard token rates; consider GPT-6 Luna (gpt-6-luna) for focused, high-volume tasks if it passes your tests. (OpenAI API model catalogue) |
Codex is OpenAI’s software-development experience; Chat is for conversational assistance. Their available models, tools, and usage rules differ. A model’s API context capacity also does not mean it automatically has access to your repository. (ChatGPT Work and Codex)
ChatGPT model vs. custom GPT vs. API model
These labels refer to different things:
- ChatGPT model and thinking setting: In regular Chat, you may select GPT-6 and a thinking level. OpenAI says GPT-6 Sol powers Instant, Medium, High, and Extra High; Pro uses GPT-6 Pro, powered by GPT-6 Astra. The options shown depend on your plan and workspace. (ChatGPT models and thinking levels)
- Codex: A software-development experience with its own model picker and usage rules. GPT-6.1 Sol and GPT-6 Astra are among the models offered in Work and Codex, but availability depends on plan, workspace settings, and rollout. OpenAI says GPT-6.1 Sol is not available in regular Chat conversations. (ChatGPT Work and Codex)
- API model: A model your application calls by an API model ID, such as
gpt-6-astraorgpt-6.1-sol. API token charges are separate from ChatGPT subscription usage. (GPT-6.1 Sol API model page) - Custom GPT: A configured GPT for a particular purpose. It is not, by itself, a separate answer to which underlying ChatGPT model or API model is best for coding.
Access and usage: GPT-6 is rolling out to eligible ChatGPT plans, and the model options shown depend on plan and workspace settings. OpenAI lists GPT-6 Pro, powered by Astra, for Pro $100, Pro $200, Business, and Enterprise plans; Enterprise access also depends on workspace model-access permissions. Plus plans include GPT-6 Astra in ChatGPT Work and Codex. GPT-6.1 Sol is rolling out for eligible paid plans, with availability depending on plan, workspace settings, and rollout access. Check your own model picker and workspace settings; these statements do not guarantee availability on every account. (ChatGPT models and thinking levels; ChatGPT Work and Codex)
For Work and Codex, model usage is governed by the applicable subscription allowance and usage rules. Purchased credits, Enterprise pay-as-you-go usage, and API usage follow separate pricing rules. So API token rates should not be used to estimate what an Astra task costs within a ChatGPT plan. (ChatGPT Work and Codex)
Choosing an OpenAI API model for production coding
OpenAI’s model catalogue suggests GPT-6 Astra for complex reasoning and coding, GPT-6.1 Sol to balance capability and cost, and GPT-6 Luna for cost-sensitive, high-volume work. These are starting points, not independent benchmark results for your application. (OpenAI API model catalogue)
| API model | Model ID | A reasonable candidate for | Standard input / output per 1M tokens |
|---|---|---|---|
| GPT-6 Astra | gpt-6-astra |
Complex coding and demanding reasoning | $10 / $50 |
| GPT-6.1 Sol | gpt-6.1-sol |
Complex coding when balancing capability and cost | $2 / $10 |
| GPT-6 Luna | gpt-6-luna |
Focused, high-volume tasks that pass your tests | $0.10 / $0.50 |
Prices are standard U.S.-dollar API token rates per million tokens, not ChatGPT subscription prices. Input and output are billed separately; tool-specific fees and other pricing conditions may apply. For API prompts above 272,000 input tokens, both Astra and Sol have higher rates applied to the full request: input and cache rates are doubled, and output rates are multiplied by 1.5. This API pricing rule does not describe subscription-covered usage in Work or Codex; check the applicable pricing details before estimating a long-context workload. (GPT-6 Astra API model page; GPT-6.1 Sol API model page; ChatGPT Work and Codex)
Check API fit before you build around a model
For Astra and Sol, OpenAI documents a 1,050,000-token context window and 128,000 maximum output. Both support function calling and structured outputs; tool calling requires the Responses API. These are API specifications—do not assume the same context or model options apply in ChatGPT or Codex. (GPT-6 Astra API model page; GPT-6.1 Sol API model page)
Function calling lets a model request actions from tools your application provides; it does not mean the model independently runs your repository’s commands or tests. Your application must provide and handle those tools. If your coding workflow needs the model to trigger a test runner, read files, or apply changes, verify that your integration implements those actions and that its outputs meet your requirements. (OpenAI function-calling guide)
Illustrative cost, not a measured test: A request using 100,000 input tokens and generating 20,000 output tokens would have standard token charges of about $2.00 for Astra, $0.40 for Sol, or $0.02 for Luna, using the listed input and output rates. This excludes tool fees, cached-input pricing, and other charges. The example is below the cited 272,000-token long-context threshold for Astra and Sol. (GPT-6 Astra API model page; GPT-6.1 Sol API model page)
How to test a coding model on your work
Don’t choose a production model based on one impressive answer. Run a small comparison you can repeat:
- Choose representative tasks: for example, fix a real bug, implement a small feature, or update several connected files.
- Hold conditions constant: give each candidate the same task description, relevant files, tools, and acceptance criteria.
- Run the code and tests: record whether the requested behavior works and whether existing tests still pass. Treat model-generated tests as useful input, not proof of correctness by themselves.
- Track the full effort: note successful completions, human corrections, retries, end-to-end time, and API token charges where applicable.
- Choose by accepted outcome: prefer the least costly option that reliably meets your requirements. Escalate to a stronger candidate only when the comparison shows an improvement worth its cost, time, or usage allowance.
This is a suggested evaluation procedure, not a claim that these models were tested here. Keep code review and automated checks in the workflow before merging generated code.
Bottom line
For conversational coding help, start with GPT-6 at Medium thinking when it is available; increase the thinking level for harder questions if your account offers it. For active project work, try Codex with GPT-6.1 Sol if your plan and workspace allow it, and consider Astra for especially demanding tasks. For production APIs, OpenAI’s catalogue points to Astra for complex coding, Sol for a capability-cost balance, and Luna for suitable high-volume work. ChatGPT, Codex, and API access and billing are distinct—confirm what your account exposes and test candidates on representative tasks before choosing. (ChatGPT Work and Codex)
Sources
GPT-6 and other models in ChatGPT — OpenAI Help Center
ChatGPT Work and Codex — OpenAI Help Center
GPT-6.1 Sol Model — OpenAI API