GPT-5

2025-08-07 · Publisher : OpenAI · Domain : Artificial Intelligence

OpenAI launches GPT-5, a unified system that combines a fast model, a deep reasoning model and a router, with mini, nano and pro versions for the API, and it becomes the default model in ChatGPT.

What changes

  • Unified system: a fast model, GPT-5 thinking for hard problems, and a router that chooses based on the complexity and the intent of the user.
  • Replaces GPT-4o, o3, o4-mini, GPT-4.1 and GPT-4.5 for signed-in ChatGPT users.
  • Fewer hallucinations (about 45% fewer factual errors than GPT-4o with search, about 80% fewer than o3 in thinking mode) and less sycophancy (from 14.5% to under 6%).
  • In the API: gpt-5, gpt-5-mini, gpt-5-nano, with a verbosity parameter and minimal reasoning_effort, and custom tools.
  • GPT-5 pro, with more parallel compute, replaces o3-pro in ChatGPT.

OpenAI launches GPT-5, a unified system that combines a fast model, a deep reasoning model and a router, with mini, nano and pro versions for the API, and it becomes the default model in ChatGPT. ### What changes - Unified system: a fast model, GPT-5 thinking for hard problems, and a router that chooses based on the complexity and the intent of the user. - Replaces GPT-4o, o3, o4-mini, GPT-4.1 and GPT-4.5 for signed-in ChatGPT users. - Fewer hallucinations (about 45% fewer factual errors than GPT-4o with search, about 80% fewer than o3 in thinking mode) and less sycophancy (from 14.5% to under 6%). - In the API: gpt-5, gpt-5-mini, gpt-5-nano, with a verbosity parameter and minimal reasoning_effort, and custom tools. - GPT-5 pro, with more parallel compute, replaces o3-pro in ChatGPT. ### Specifications - Context 400,000 tokens (272,000 input, 128,000 output), text and image input, text output. - Cutoff: September 30, 2024 for gpt-5, May 31, 2024 for mini and nano. - API: Responses and Chat Completions; gpt-5-chat-latest for the non-reasoning ChatGPT model; gpt-5-pro in the API from October 6, 2025. ### Pricing Prices per million tokens (USD) at API launch. GPT-5 pro: 15 input and 120 output (documentation). | Model | Input | Cached input | Output | | --- | --- | --- | --- | | gpt-5 | 1.25 | 0.125 | 10 | | gpt-5-mini | 0.25 | 0.025 | 2 | | gpt-5-nano | 0.05 | 0.005 | 0.40 | ### Benchmarks Scores published by the vendor (high effort, no tools unless noted). | Benchmark | GPT-5 | mini | nano | | --- | --- | --- | --- | | AIME 2025 | 94.6% | 91.1% | 85.2% | | SWE-bench Verified (477 tasks) | 74.9% | 71.0% | 54.7% | | GPQA Diamond | 85.7% (pro: 88.4%) | 82.3% | 71.2% | | MMMU | 84.2% | 81.6% | 75.6% | | Aider polyglot | 88% | not disclosed | not disclosed | | HealthBench Hard | 46.2% | not disclosed | not disclosed | | τ2-bench telecom | 96.7% | not disclosed | not disclosed | ### Key takeaways GPT-5 simplifies model choice and becomes OpenAI's reference for agentic and coding work, with three sizes to trade off cost and latency. The SWE-bench scores cover 477 of the 500 tasks.

Links

Tags : LLM, OpenAI, GPT, GPT-5, Raisonnement, Agent