The previous flagship, now half the price of GPT-5.5. Still the right pick for most production code and agent loops.
With GPT-5.5 shipping in April 2026, GPT-5.4 quietly became the better default for most developers. The pricing did not change — $2.50 input and $15 output per 1M tokens — but the relative value moved a lot. GPT-5.5 doubled the rate without doubling the win on most workloads, so the cost-per-correct-answer calculation often still favors GPT-5.4.
OpenAI positions GPT-5.4 as the coding and agentic specialist. In practice that means strong tool-calling discipline, fewer rambling answers, and predictable structured-output behavior. For anything that loops — code agents, multi-step workflows, browser automation — GPT-5.4 is rarely the wrong choice.
GPT-5.4 is multimodal (vision + text + audio in / text + audio out depending on the endpoint), supports the same tool-use surface as GPT-5.5, and ships in three sizes: GPT-5.4, GPT-5.4 mini ($0.25 / $2), and GPT-5.4 nano ($0.05 / $0.40). The full version stays the right pick for code generation, refactors, and any agent that needs to plan more than three steps ahead.
It trails GPT-5.5 on the very longest contexts (272K vs 1M) and on the hardest open-ended reasoning. It beats GPT-5.5 on price-per-token by roughly 2x.
| Model | Input / 1M | Output / 1M | Context |
|---|---|---|---|
| GPT-5.4 | $2.50 | $15 | 272K |
| GPT-5.4 mini | $0.25 | $2 | 272K |
| GPT-5.5 | $5 | $30 | 1M |
| Claude Sonnet 4.6 | $3 | $15 | 200K (1M β) |
Sonnet 4.6 is the closest cross-family alternative — slightly more expensive on input, identical on output, comparable quality on most tasks. Pick GPT-5.4 when you live in the OpenAI tool ecosystem; pick Sonnet 4.6 when you want Anthropic's coding style and longer-context beta.