GPT-5 Arrives, and the Number Worth Reading Is the Price List
A router in the chat product, three fixed sizes in the API, and benchmarks with footnotes.
OpenAI released GPT-5 on 7 August 2025 and made it the default model in ChatGPT. Free users got it the same day. Plus subscribers, at 20 dollars a month, got higher usage limits; Pro subscribers, at 200 dollars a month, got unlimited access plus a separate GPT-5 Pro variant. Team, Enterprise and Edu accounts were told they would get it as their default the following week.
The structural change is that GPT-5 in ChatGPT is not a single model. A real-time router decides whether a question gets a quick answer or a longer round of reasoning, and it makes that call per message rather than leaving it to the user. Free users who exhaust their GPT-5 allowance are moved to GPT-5 mini rather than cut off.
The API is a different product, and for anyone building software it is the one that matters. There GPT-5 ships in three sizes — gpt-5, gpt-5-mini and gpt-5-nano — chosen explicitly, with no router in the way. Published prices per million tokens are 1.25 dollars in and 10 dollars out for gpt-5, 0.25 and 2 for gpt-5-mini, and 0.05 and 0.40 for gpt-5-nano. Every size accepts up to 272,000 input tokens and emits up to 128,000 reasoning and output tokens, for 400,000 in total.
The numbers, and the footnotes attached to them
OpenAI led with coding and long-running agent work. On SWE-bench Verified, a suite built from real GitHub issues, GPT-5 scored 74.9 percent against o3's 69.1, while using 22 percent fewer tokens and 45 percent fewer tool calls. On Aider polyglot, a code-editing evaluation, it reached 88 percent, which OpenAI describes as cutting the error rate by roughly a third relative to o3. On tau2-bench telecom, a multi-turn tool-calling test, it scored 96.7 percent. Among early testers quoted by OpenAI, Cursor described it as the smartest model its team has used and Windsurf reported it as state of the art on their internal evaluations.
The footnotes matter. OpenAI's own note on the SWE-bench figure says 23 of the 500 problems were left out because they did not run reliably on its infrastructure, so 74.9 percent is a score over 477 tasks. The margin over Anthropic's Claude Opus 4.1, which TechCrunch put at 74.5 percent, is narrower than that omission. And the same tau-bench family that produced 96.7 on telecom has two sections where GPT-5 did not lead: TechCrunch reported 63.5 percent on airline, behind o3's 64.8, and 81.1 percent on retail, behind Claude Opus 4.1's 82.4. Sam Altman called GPT-5 "the best model in the world". TechCrunch's own assessment was more measured, describing it as comparable to the other frontier models across several areas rather than clearly ahead of them.
What this means if you are building something
The pricing deserves more attention than the benchmark. The non-reasoning model that ChatGPT uses is available in the API as gpt-5-chat-latest at the same 1.25 and 10 dollars as gpt-5 itself, so reasoning carries no per-token premium within this family. Combine that with the reported reduction in tokens and tool calls per task and the cost of an agentic run falls further than the sticker price implies. Before defaulting to the largest size, test gpt-5-mini at 0.25 and 2 dollars on the routine work — extraction, classification, drafting — where most client budgets actually go.
Do not build a client product on the consumer interface. Inside ChatGPT the routing decision belongs to OpenAI, not to the user and not to you, and access arrived on different dates for different plans: free accounts on day one, Team, Enterprise and Edu a week later. If your product depends on a particular model being reachable through a chat product you do not control, that is not a dependency you can hold. Pin a model name in the API instead.
Run your own evaluation before you accept any of this. The figures above were produced by the company selling the model, on suites with documented omissions. Twenty real tasks from your own codebase, scored by hand, will tell you more in an afternoon than the full benchmark table. Two new API controls make that cheap to try: reasoning effort now accepts a minimal setting for faster answers, and a verbosity parameter takes low, medium or high.
Our reading, as of 11 August 2025: the coding and tool-calling gains are worth re-running your own tests over, and the mini tier changes the arithmetic on features that were previously too expensive to ship. The consumer product and the API are on different footing, though. Build on the one with version numbers.
Sources
