GPT-5.4: OpenAI reasoning model
OpenAI's workhorse of the GPT-5 family. Sits at the ×1 credit baseline alongside Claude Sonnet 4.6 and is the right default for most Codex-framework agents.
400K tokens · Text / Vision / Code · Prompt cache
Okou no longer runs GPT-5.4. This page is kept as a reference for its specs, pricing and benchmarks. For the same kind of work, use GPT 5.6 Luna.
See GPT 5.6 LunaGPT-5.4 is the workhorse of OpenAI's GPT-5 family — the model you keep running everywhere by default. Vendor-reported SWE-bench Verified at 74.9% places it in the same range as Claude Sonnet 4.6 on coding, and its tool-use accuracy is what most production Codex-framework agents are tuned against.
What is GPT-5.4?
April 2026 · Workhorse of the GPT-5 family. The recommended default for most Codex-framework agents.
GPT-5.4 is the workhorse of OpenAI's GPT-5 generation, released in April 2026 alongside the flagship GPT-5.5 and the cost-optimised GPT-5.4 Mini. OpenAI positions it as the everywhere-default for agents on the Codex framework — the model you keep running on every step unless a specific step justifies escalation to 5.5.
Architecturally GPT-5.4 shares the 400K-token context window, the reasoning_effort parameter, prompt caching and the Responses API surface with the rest of the GPT-5 family. The split versus GPT-5.5 is compute investment per token: 5.4 runs faster and cheaper, 5.5 invests more in reasoning depth. The split versus GPT-5.4 Mini is the opposite — 5.4 carries more quality for the steps that actually decide the agent run.
What's notable about GPT-5.4
Headline architecture and capability features.
GPT-5.4 uses the same architecture as the rest of the GPT-5 family: 400K-token context window, reasoning_effort parameter at four levels (minimal, low, medium, high), prompt caching where cached input bills at one-tenth the input rate, and the Responses API surface that codex CLI uses by default. Tool-use, structured outputs and computer-use are supported. Inputs are multimodal across text, vision and code.
Specs at a glance
GPT-5.4 benchmarks
Vendor-reported scores from OpenAI's GPT-5 release materials, with deltas shown against the previous OpenAI generation. Independent reviews place GPT-5.4 in the same coding-quality band as Claude Sonnet 4.6. Treat absolute percentages as directional.
GPT-5.4 pricing
Provider list price, per 1M tokens.
How GPT-5.4 behaves in practice
Observed behaviour from production agent runs.
Tool routing
Solid baseline accuracy across the standard Codex-framework tool catalogue. Where 5.5 pulls ahead is on hard edge cases (conditional tool selection, deeply nested arguments) — for the routine cases 5.4 routes correctly with significantly lower latency.
Code edits
Comparable patch quality to Claude Sonnet 4.6 on standard refactor and bug-fix workloads. Where 5.5 starts to pull ahead is on multi-file changes where the patch has to apply cleanly on the first try.
Speed
Materially faster than 5.5 — around 110 tokens/sec at medium effort per Artificial Analysis. This is part of why 5.4 stays the default for interactive chat replies and short agent loops where user-visible latency matters.
Cost efficiency
×1 credits with output behaviour in the Sonnet 4.6 quality band. For teams already on the Codex framework, this is the cost/quality sweet spot — promote to 5.5 only on steps that visibly need it.
Hallucination behaviour
Inherits the calibration improvements OpenAI shipped with the GPT-5 generation. Less prone to confident wrong answers than the GPT-4 series, especially on questions outside its training horizon.
Best agent tasks for GPT-5.4
The default agent step on the Codex framework
If your agent is already built on codex CLI or any Codex-framework integration, GPT-5.4 is the natural everywhere-default. ×1 credits, fast enough for interactive use, accurate enough for the routine tool calls that dominate most agent runs.
When to skip GPT-5.4
Skip GPT-5.4 on the hardest reasoning, computer-use or multi-file code-edit steps where 5.5 noticeably leads, and on high-volume bulk classification or pre-filter work where 5.4 Mini is four times cheaper at the vendor level.
GPT-5.4 vs other models
GPT-5.4 vs GPT-5.5
Same family, different positioning. 5.5 (×2) gives you the strongest reasoning, computer-use and first-attempt code quality; 5.4 (×1) gives you the same context window and feature set at half the credit cost and noticeably higher speed. Default to 5.4; escalate to 5.5 only on the steps that visibly need it.
GPT-5.4 vs Claude Sonnet 4.6
The two ×1 baselines, one in each ecosystem. Sonnet 4.6 runs on the Claude Code framework; GPT-5.4 runs on Codex. Pick by which framework your existing agents and tool definitions target. On raw output quality they're close enough that A/B-testing on your workload is the right call.
GPT-5.4 vs GPT-5.4 Mini
Same family, different positioning. 5.4 (×1) carries more reasoning quality per token; 5.4 Mini (×0.3) gives you a much cheaper option for bulk and pre-filter work. Use 5.4 Mini for fan-out classification and 5.4 for the steps that decide the agent run.
Frequently asked questions
What is GPT-5.4's context window?
400,000 tokens, with up to 128K tokens of output per response. The full window bills at standard rates.
Can GPT-5.4 handle images?
Yes. GPT-5.4 is multimodal. It accepts image inputs alongside text and code natively.
When should I pick GPT-5.4 over Claude Sonnet 4.6?
When your agent is already built on the Codex framework or you need the OpenAI ecosystem (tool catalogue, structured outputs, Responses API). Both sit at ×1 credits, so cost is identical and the choice comes down to framework and behaviour fit.
Does GPT-5.4 support prompt caching?
Yes. Cached input bills at $0.25 per 1M tokens — a 10× discount on the cached portion.
Alternatives
Availability of GPT-5.4 on Okou
GPT-5.4 has been removed from the Okou lineup, so it can no longer be selected in chat or in a workflow, and it is not available with your own API key. GPT 5.6 Luna covers the same cost-saving tier.
