GPT-5.4 Guide: 1M Context, Computer Use and the 272K Cliff

Before GPT-5.4, choosing an OpenAI model for engineering work meant choosing between two lines: the general GPT models for reasoning and writing, and the Codex models for agentic coding. GPT-5.4, released on 5 March 2026, was the release that folded the coding gains of GPT-5.3-Codex back into the general flagship, and added native computer use and a far larger context window on top.

That makes it a sensible default for a lot of work. It also comes with one pricing rule that will surprise anyone who takes the 1M window at face value.

One model where there used to be two

OpenAI's reported results show why the merge was possible. GPT-5.4 scores 57.7% on SWE-Bench Pro, a little above GPT-5.3-Codex's 56.8%, so the general model no longer gives up coding quality to the specialist. On GDPval, OpenAI's test of professional work across 44 occupations, it matched or beat industry professionals in 83% of comparisons, up from 70.9% for GPT-5.2.

The API profile reflects the same ambition. OpenAI lists support for function calling, web search, file search, tool search, image generation, code interpreter, hosted shell, apply_patch, skills, computer use and MCP in the Responses API, and the model also works through Chat Completions. Knowledge cutoff is 31 August 2025. The current snapshot is gpt-5.4-2026-03-05.

There is no published parameter count or architecture, and no open weights.

The 1M window has a price cliff at 272K

OpenAI documents a context window of 1,050,000 tokens and up to 128,000 output tokens. Standard pricing is $2.50 per million input tokens, $0.25 for cached input and $15 per million output, and OpenRouter shows the same base rates at the time of writing.

The catch is in the long-context rule. Once a prompt exceeds 272,000 input tokens, OpenAI bills 2x input and 1.5x output for the full session, and that applies to standard, batch and flex processing alike. It is not a surcharge on the tokens above the line; it reprices the whole request.

A worked example makes the size of the step clear. A 270,000-token prompt with a 5,000-token answer costs about $0.68 of input and $0.08 of output, roughly $0.75. Add 30,000 tokens to cross the line and the same answer costs about $1.50 of input and $0.11 of output, roughly $1.61. Eleven per cent more context more than doubled the bill.

The practical rule is to treat 272K as the working budget and the rest of the window as an expensive overflow. Retrieval, compaction and trimming tool output are worth more on this model than on one with flat long-context pricing. The token cost of long context covers when filling a large window is worth paying for.

Tool search changes what large tool sets cost

Agents connected to several MCP servers often send tens of thousands of tokens of tool definitions on every call, whether or not those tools are used. GPT-5.4 added tool search, which lets the model fetch tool definitions on demand instead of receiving all of them up front.

OpenAI reported a 47% reduction in total token usage in internal testing across 250 tasks and 36 MCP servers. Your saving will depend on how many tools you expose and how many a typical task touches, but the direction is reliable: the larger the catalogue, the more tool search returns. It also interacts with the 272K rule above, because definitions that are never loaded never push a prompt over the line. The token cost of tool calls covers how to measure definition overhead before and after.

Computer use, and what 75% means

GPT-5.4 reads screenshots and issues keyboard and mouse actions natively. OpenAI reports 75.0% on OSWorld-Verified, above the 72.4% human baseline that benchmark cites and up from 47.3% for GPT-5.2.

Read that number carefully. OSWorld tasks are short, well-defined desktop jobs in a controlled environment. Beating the human baseline there means the model can be trusted with routine GUI steps under supervision, not that it can run an unattended desktop session against production systems. Every action it takes through a screen is an action you cannot easily diff or roll back, so put approval gates in front of anything that writes.

Reasoning effort starts at none

GPT-5.4 supports five effort levels: none, which is the default, then low, medium, high and xhigh. Leaving it at the default gives you a fast, non-reasoning model; you have to ask for depth.

That default is easy to miss when comparing against published benchmarks, which are usually run at high or xhigh. If your evaluation runs at none and the vendor chart ran at xhigh, you are not testing the same thing. Pick the effort level per task type: none for formatting and extraction, medium for routine code, high or xhigh for debugging and design.

Choosing within the 5.4 family

GPT-5.4 is the middle of a family rather than the only option.

• GPT-5.4 mini scores 54.4% on SWE-Bench Pro and 72.1% on OSWorld-Verified at $0.75 input and $4.50 output. For subagents and high-volume coding it is often the better buy; see the GPT-5.4 mini guide.
• GPT-5.4 nano is for classification, extraction and ranking at $0.20 input and $1.25 output, and is weak at computer use.
• GPT-5.3-Codex remains a reasonable choice for shell-heavy agent loops that fit within its 272K input cap; the GPT-5.3-Codex guide compares the two.
• GPT-5.4 Pro spends far more compute per request and is priced for occasional, very hard questions.

OpenRouter also lists openai/gpt-5.4:batch, the same model on an asynchronous endpoint at half price: $1.25 input and $7.50 output per million tokens at the time of writing. If a job can wait for results, such as overnight evaluation runs or bulk document processing, it is the cheapest way to use the full model.

The short version: use GPT-5.4 as the general default, keep prompts under 272K on purpose, turn on tool search once you expose more than a handful of tools, and set reasoning effort explicitly rather than trusting the default.