OpenAI ships GPT-5 to the API as three model sizes with new format and reasoning controls

The API model is not the ChatGPT model

The most consequential line for anyone building on the platform is buried past the benchmarks: GPT-5 in the API is the reasoning model, not the router-plus-non-reasoning system that ChatGPT presents to consumers. OpenAI is explicit that gpt-5 with minimal reasoning is a distinct model from ChatGPT's non-reasoning path, and that the ChatGPT non-reasoning model is exposed separately as gpt-5-chat-latest.

This matters because it tells developers what they are actually calling. ChatGPT abstracts model selection behind a router; the API hands you the reasoning engine directly and asks you to manage the trade-offs yourself. The three sizes—gpt-5, gpt-5-mini, gpt-5-nano—exist precisely so you make those cost, latency, and quality decisions per call rather than delegating them to an opaque router.

Efficiency claims sit beside the score claims

GPT-5 scores 74.9% on SWE-bench Verified, up from o3's 69.1%, and 88% on Aider polyglot, which OpenAI frames as a one-third reduction in error rate. But the number OpenAI chose to emphasize alongside the SWE-bench figure is about consumption: relative to o3 at high reasoning effort, GPT-5 uses 22% fewer output tokens and 45% fewer tool calls to reach its score.

For agentic workloads that chain dozens of tool calls, fewer calls and fewer tokens translate directly into lower bills and lower latency. The τ2-bench telecom result—96.7%, against a published field where no model had cleared 49% two months earlier—is the headline agentic claim, but the practical read is that the same benchmark run costs less per attempt. OpenAI is arguing on unit economics, not just capability.

The honesty note is worth keeping: OpenAI omitted 23 of 500 SWE-bench problems that would not run on its infrastructure and disclosed the exact task IDs, and it swapped the MultiChallenge grader from GPT-4o to o3-mini because the default grader mis-scored responses. These are the kinds of caveats that determine whether a benchmark reproduces in your own harness.

Custom tools trade JSON reliability for plaintext grammars

The new custom tools type addresses a concrete failure mode rather than a marketing one. OpenAI notes that valid JSON requires the model to escape every quotation mark, backslash, newline, and control character, and that on long inputs—hundreds of lines of code or a five-page report—the odds of a malformed escape creep up. Custom tools let GPT-5 emit tool inputs as plaintext, constrained by a developer-supplied regex or full context-free grammar.

OpenAI reports GPT-5 scores about the same on SWE-bench Verified whether it uses custom tools or JSON tools, which is the point: the feature is not about lifting the benchmark but about removing a class of parsing failures on long outputs. Paired with the minimal reasoning_effort value and the verbosity parameter, the release is largely a set of controls that let developers dial in speed, output length, and format reliability per task.

A 400K context window and lower hallucination rates aim at correctness-sensitive agents

Every GPT-5 API model accepts up to 272,000 input tokens and emits up to 128,000 reasoning-and-output tokens, for a 400,000-token total. OpenAI paired this with long-context retrieval results—89% correct on BrowseComp Long Context at 128K–256K tokens—and a claim of roughly 80% fewer factual errors than o3 on LongFact and FactScore prompts, with the reported LongFact-Objects hallucination rate dropping from o3's 6.8% to 1.2%.

The specific implication is that OpenAI is positioning GPT-5 for agentic systems where a wrong answer propagates through subsequent tool calls—code changes, data operations, customer-service actions on mutable state. Cheaper tokens and fewer calls lower the cost of running such agents; lower hallucination rates and preamble messages between tool calls are meant to make their behavior auditable. The pricing—$1.25 per million input tokens and $10 per million output for gpt-5, down to $0.05 and $0.40 for nano—lets teams route cheap retrieval and simple steps to smaller models while reserving the full model for the steps where correctness actually costs money.