Model page

GPT-5

Frontier reasoning depth with best-in-class reliability.

About GPT-5

GPT-5 is OpenAI's unified flagship — a single model that scales its reasoning effort up or down to match the task, rather than making you pick a separate variant. The result shows up most clearly in hard technical work: a perfect AIME 2025 score with tools, 74.9% on SWE-bench Verified, and a 1.6% error rate on medical benchmark HealthBench make it one of the more capable models available for code, math, and domain-specific research. Its 400,000-token context window handles large codebases and lengthy documents without truncation. Users consistently praise the step-up in accuracy and the meaningful reduction in hallucinations over GPT-4o. The honest caveat: GPT-5 trades warmth for precision. Early adopters widely noted that responses are shorter, cooler, and noticeably less conversational than its predecessor — a real shift if personality and back-and-forth rapport matter to your workflow. Latency is also substantial; extended reasoning produces a time-to-first-token around 68 seconds, which rules it out for anything requiring snappy replies.

Best for

  • Advanced coding and software engineering tasks requiring multi-step reasoning
  • Mathematical problem-solving at competition and graduate level
  • Medical and scientific research where accuracy outweighs response speed
  • Visual analysis of complex diagrams, charts, and technical screenshots
  • Long-form technical writing, documentation, and deep document analysis

Specs & capabilities

How GPT-5 stacks up — intelligence, speed, context, and modalities.

Capability

Intelligence

High

Capability

Speed

Medium

Capability

Context window

400,000 tokens

Capability

Knowledge cutoff

September 30, 2024

Frequently asked questions

What does GPT-5 cost?

Via the OpenAI API, GPT-5 is priced at $1.25 per million input tokens and $10.00 per million output tokens.

How large is GPT-5's context window?

The base model supports 400,000 tokens of context, making it well-suited for large codebases, long documents, or extended research sessions.

Why does GPT-5 sometimes take a long time to respond?

GPT-5 can spend extra time on extended reasoning for complex tasks. This boosts accuracy significantly but pushes time-to-first-token to around 68 seconds in high-effort mode — well above the industry median of roughly 2–3 seconds. If you need fast replies, it is not the right fit.

How does GPT-5 compare to GPT-4o?

GPT-5 is considerably stronger on benchmarks — particularly math, coding, and factual accuracy — but users widely report it feels colder and less witty than GPT-4o. It is the better choice for precision-heavy work; GPT-4o remains preferred for conversational or creative tasks.

Is GPT-5 a single model?

Yes. On the API, GPT-5 is one model with an adjustable reasoning-effort setting — you control how much it thinks. The automatic switching people sometimes describe is a feature of the ChatGPT app's product layer, not the GPT-5 model itself.

What is GPT-5's knowledge cutoff?

GPT-5's training data has a cutoff of September 30, 2024. Events or publications after that date are outside its knowledge unless you provide context directly.

Related models