Is there an official GPT-5.4 model card?
Yes. OpenAI published a “GPT-5.4 Thinking System Card” on March 5, 2026. It covers model data and training, safety evaluations, and preparedness assessments. The card refers to the model as gpt-5.4-thinking; the API model page documents the API model ID as gpt-5.4. Use the system card for safety context and the API model page for API features, limits, and pricing. (OpenAI Deployment Safety Hub)
Inworld’s page is a provider listing, not OpenAI’s model card; its examples use Inworld’s endpoint. The captured CometAPI fragment is an uncited third-party source, so it does not establish official OpenAI specifications.
What OpenAI documents about GPT-5.4
OpenAI’s API documentation lists a 1,050,000-token context window, 128,000 maximum output tokens, text and image input, and text output. It lists streaming, function calling, structured outputs, file search, image input, web search, and prompt caching as supported features. The page lists tools supported through the Responses API, including function calling, web search, file search, tool search, image generation, code interpreter, hosted shell, computer use, and MCP. Tool availability is documented specifically for that API; confirm the relevant tool and configuration in the current reference before relying on it. These feature listings do not guarantee success on every task or configuration. (OpenAI API model documentation)
The System Card describes training data at a high level and includes sections on baseline safety evaluations, disallowed-content evaluations with challenging prompts, production benchmarks, jailbreaks, prompt injection, vision, health, avoiding accidental data-destructive actions, and user confirmations during computer use. It also includes preparedness assessments. OpenAI says GPT-5.4 Thinking is the first general-purpose model to have implemented mitigations for High capability in Cybersecurity. That is a specific statement about cybersecurity mitigations, not a guarantee of general application safety. Review the card’s evaluation scope and use safeguards and workload-specific tests appropriate to your own application. (GPT-5.4 Thinking System Card PDF)
API model selection: limits, reasoning, and tools
The API documentation identifies the model as gpt-5.4 and lists the default snapshot gpt-5.4-2026-03-05. Supported reasoning-effort values are none (the default), low, medium, high, and xhigh. The page lists both the Responses API and Chat Completions as supported endpoints. Its named tool list applies to the Responses API. If your application requires a particular tool or reasoning setting, verify it in the reference and test it using the endpoint and configuration you intend to deploy. (OpenAI API model documentation)
A useful starting decision:
- Evaluate GPT-5.4 when your task needs complex reasoning, long context, or tool-using workflows, and the cost and latency fit your requirements.
- Compare GPT-5.4 mini or nano when speed or cost is a priority. OpenAI describes these as smaller, faster versions; treat that positioning as a reason to test them, not proof they meet your quality bar. (OpenAI: GPT-5.4 mini and nano)
API pricing and a simple estimate
OpenAI’s API model page lists these USD prices per 1 million tokens for GPT-5.4: (OpenAI API model documentation)
| Token type | Listed price |
|---|---|
| Input | $2.50 |
| Cached input | $0.25 |
| Output | $15.00 |
Long-context example, standard pricing, without regional processing: for 900,000 uncached input tokens and 10,000 output tokens, the documented long-context condition applies because input exceeds 272,000 tokens. Input is charged at 2× the listed rate ($5.00 per million, or $4.50 for this example); output is charged at 1.5× the listed rate ($22.50 per million, or $0.225 for 10,000 tokens). The estimated total is $4.725, or about $4.73, before any other applicable charges. OpenAI states that the 2× input and 1.5× output rates apply for the full session for standard, batch, and flex pricing when prompts exceed 272,000 input tokens. (OpenAI API model documentation)
The listed input and requested output total 910,000 tokens, nominally below the stated 1,050,000-token context window. This arithmetic alone does not establish that a particular request will fit: leave headroom for any reasoning-token budget and other tokens counted by the endpoint, and confirm the applicable accounting in that endpoint’s documentation.
OpenAI also states that regional-processing (data-residency) endpoints have a 10% uplift for GPT-5.4 and GPT-5.4 Pro. The example above excludes that uplift; include it when using a qualifying regional-processing endpoint. Confirm current pricing and the conditions for your selected endpoint before budgeting. (OpenAI API model documentation)
An Inworld listing shows similar headline rates and a dash under “cache write,” but it is a third-party provider listing and does not explain the dash. For OpenAI API budgeting, use OpenAI’s documented cached-input rate and pricing conditions rather than interpreting that third-party table.
What to test before production
The model card and API reference describe published information; they do not establish that GPT-5.4 is production-ready for your specific application. Before choosing a model:
- Build a representative test set. Include typical tasks, edge cases, long inputs, ambiguous requests, and examples where incorrect output would have meaningful consequences.
- Compare candidate models on the same prompts and tools. Track task success, factual or extraction errors, tool-call failures, latency, and token use. Published evaluations provide context, but do not replace results on your workload. (OpenAI: Introducing GPT-5.4)
- Exercise the exact API path. Use the current documentation for your chosen endpoint, model ID, and tool configuration. Do not mistake an OpenAI-SDK-compatible third-party endpoint for OpenAI’s direct API. (OpenAI API model documentation)
- Model realistic costs. Use expected input and output volumes, cached-token behavior, and any applicable long-context or regional-processing conditions. Recheck the provider’s current rates before committing a budget. (OpenAI API model documentation)
- Add application safeguards. For computer use or other consequential actions, define permissions, confirmation steps, error handling, and monitoring appropriate to the risk. The System Card includes a section on user confirmations during computer use; your implementation still needs controls suited to its users and environment. (GPT-5.4 Thinking System Card PDF)
Bottom line: GPT-5.4 has an official OpenAI System Card and API documentation. Use the card for training and safety context, the API model page for technical and pricing details, and representative tests to decide whether the model’s capabilities, limits, and costs fit your production workload.
Sources
GPT-5.4 Thinking System Card — OpenAI