Short answer: GPT-5 Mini (gpt-5-mini) and GPT-5.4 Mini (gpt-5.4-mini) are distinct model IDs. OpenAI marks GPT-5 Mini as deprecated and recommends GPT-5.6 Terra as a starting point for most new low-latency, high-volume workloads. GPT-5.4 Mini is a separate candidate for coding, computer use, and subagent workflows. Check the current model page for the model and API surface you plan to use, then evaluate it on representative tasks before committing to production. (developers.openai.com)
Which GPT-5 mini model should you evaluate?
| Model | What OpenAI’s API documentation says | Selection starting point |
|---|---|---|
GPT-5 Mini (gpt-5-mini) |
Marked deprecated. The model page lists text and image input, a 400,000-token context window, and up to 128,000 output tokens. | Consider it only if you have a specific compatibility or migration reason. Check the deprecation notice and your account’s access before relying on it in production. |
GPT-5.4 Mini (gpt-5.4-mini) |
A separate model ID, described as a small model for coding, computer use, and subagents. Its page lists text and image input, a 400,000-token context window, and up to 128,000 output tokens. | Shortlist it for relevant coding, image-input, or tool-using tasks, then evaluate it on representative work. |
| GPT-5.6 Terra | GPT-5 Mini’s model page recommends starting with GPT-5.6 Terra for most new low-latency, high-volume workloads. | Consider it as an alternative for a new service in that category; it is not another name for GPT-5 Mini. Check its own model page before selecting it. |
The identifiers, status, inputs, and limits above are listed on OpenAI’s model pages. A dated announcement or third-party catalog does not establish current access for your account. (developers.openai.com)
How do GPT-5 Mini and GPT-5.4 Mini compare?
In its March 17, 2026 announcement, OpenAI said GPT-5.4 Mini improves over GPT-5 Mini across coding, reasoning, multimodal understanding, and tool use, while running more than twice as fast. The announcement also reports the benchmark results below. These are OpenAI-reported evaluation results, not a guarantee of quality or latency for your application. (openai.com)
| Evaluation | GPT-5.4 Mini | GPT-5 Mini |
|---|---|---|
| SWE-Bench Pro (Public) | 54.4% | 45.7% |
| Terminal-Bench 2.0 | 60.0% | 38.2% |
| OSWorld-Verified | 72.1% | 42.0% |
For these reported comparisons, GPT-5.4 Mini was evaluated at xhigh reasoning effort and GPT-5 Mini at its highest available effort, high. Benchmark results depend on the evaluation setup; use them to inform a shortlist, then test your prompts, tools, and acceptance criteria. (openai.com)
Match a candidate to the work
- Coding assistant or code-editing agent: Evaluate GPT-5.4 Mini. OpenAI highlights targeted edits, codebase navigation, front-end generation, and debugging loops. Measure task success and the human correction required, not just response speed. (
openai.com) - Computer use or tool-using workflow: GPT-5.4 Mini is a candidate. OpenAI lists computer use among the tools supported when using the Responses API. Confirm the specific tool and endpoint in the current documentation for your integration. (
developers.openai.com) - Text-and-image input: OpenAI’s model pages list image input for both models. That identifies a supported input modality; it does not guarantee accuracy on your image task. Test representative examples, including ambiguous ones. (
developers.openai.com) - New low-latency, high-volume service: Don’t choose GPT-5 Mini solely because its listed token rates are lower. OpenAI marks it deprecated and recommends GPT-5.6 Terra as a starting point for most new workloads in this category. Compare eligible candidates on the same task set, including retries, tools, and human review. (
developers.openai.com)
API identity, limits, and pricing
OpenAI lists gpt-5-mini and gpt-5.4-mini separately. The GPT-5.4 Mini page lists both the Responses API route (/v1/responses) and Chat Completions (/v1/chat/completions); it lists computer use as a Responses API tool. For GPT-5 Mini, confirm the currently supported endpoint on its model page and in the API reference rather than inferring compatibility from another model’s listing. OpenAI recommends the Responses API for new projects. (developers.openai.com)
Both model pages list a 400,000-token context window and 128,000 maximum output tokens. Rate limits depend on your API usage tier, so check the limits that apply to your account before estimating production capacity. (developers.openai.com)
| OpenAI API model | Uncached input, per 1 million tokens | Cached input, per 1 million tokens | Output, per 1 million tokens |
|---|---|---|---|
GPT-5 Mini (gpt-5-mini) |
$0.25 | $0.025 | $2.00 |
GPT-5.4 Mini (gpt-5.4-mini) |
$0.75 | $0.075 | $4.50 |
These are the USD text-token rates listed on OpenAI’s model pages, per 1 million tokens. Tool-specific charges may also apply. OpenAI’s GPT-5.4 Mini page says regional-processing endpoints have a 10% uplift for that model. Check current pricing and applicable conditions before budgeting. GPT-5 Mini’s lower listed rates do not change its deprecated status. (developers.openai.com)
Worked estimate across four requests: Suppose each request uses 250,000 uncached input tokens and generates 50,000 output tokens. Each call stays within the listed context and output limits; together, the four calls use 1 million input tokens and 200,000 output tokens. At the listed rates, the token charges total $0.65 for GPT-5 Mini (1 × $0.25 + 0.2 × $2.00) or $1.65 for GPT-5.4 Mini (1 × $0.75 + 0.2 × $4.50). This arithmetic excludes tool charges and assumes no cached input; it is not a workload forecast.
Production API checks before you deploy
- Confirm model ID and status. The API pages list
gpt-5-miniandgpt-5.4-minias separate IDs, and GPT-5 Mini is marked deprecated. Verify account access and migration implications before relying on it. - Confirm endpoint and request details. Use the current API reference for the selected model to verify the endpoint, request fields, authentication, and response handling. Follow the migration guide if moving from Chat Completions. Don’t treat a third-party configuration snippet as a complete OpenAI API request.
- Verify features for the selected model and endpoint. Check the model page for each modality, parameter, or tool you intend to use. Don’t assume support on one model or API surface carries over to another.
- Check account-specific limits and costs. Confirm your usage-tier limits and account for token use, retries, and any tool charges when sizing the service.
- Run a representative evaluation. Test the same cases across shortlisted models. Record task success, output quality, latency, token use, and failure or retry rate. For tool-using workflows, evaluate the full tool loop, not just one model response.
Bottom line
For a new production build, don’t default to GPT-5 Mini without a specific reason: OpenAI marks it deprecated and recommends GPT-5.6 Terra as a starting point for most new low-latency, high-volume workloads. Evaluate GPT-5.4 Mini when coding, computer use, image input, or subagents fit your requirements. Confirm the current API details and test the real workload before choosing. (developers.openai.com)