OpenAI unveiled GPT-6 today, marking what the company describes as its most significant leap in reasoning capability since the introduction of GPT-4. The model achieves near-human performance on standardized reasoning benchmarks and introduces a native multi-modal architecture that processes text, images, audio, and video in a unified pipeline.
Reasoning Improvements
The centerpiece of GPT-6 is its enhanced chain-of-thought reasoning. On the MATH benchmark, which tests competition-level mathematics, GPT-6 scores 94.2 percent, up from GPT-5's 76.8 percent. On GPQA, a graduate-level science reasoning test, the model reaches 83.1 percent, surpassing the average human expert baseline of 81 percent.
OpenAI attributes this leap to a new training paradigm it calls "deliberative alignment," which teaches the model to explicitly reason through safety and accuracy considerations before generating a response. The model spends additional compute during inference on complex queries, dynamically allocating thinking time proportional to problem difficulty.
Native Multimodality
Unlike previous models that bolted on vision and audio capabilities, GPT-6 was trained from scratch on a unified architecture. It can analyze a 30-minute video and answer questions about specific moments, transcribe and translate audio in real time, and generate images with consistent character and style references across a conversation.
"GPT-6 doesn't just process multiple modalities. It understands them as part of a single, coherent representation of the world," said Sam Altman, OpenAI CEO.
Context and Pricing
GPT-6 supports a 2-million-token context window, double that of GPT-5. This enables the model to process entire codebases, lengthy legal documents, or complete book series in a single prompt. Pricing is set at $15 per million input tokens and $45 per million output tokens, positioning it as a premium offering alongside the continued availability of GPT-5 at lower price points.
The model is available through the OpenAI API and ChatGPT Pro subscription immediately, with broader rollout to Plus and Team tiers scheduled over the coming weeks. OpenAI also announced partnerships with Apple and Microsoft to integrate GPT-6 into Siri and Copilot respectively.
Safety and Alignment
OpenAI released a 78-page system card detailing the model's safety evaluation, including red-team testing by over 200 external researchers. The company acknowledges that GPT-6 can generate convincing disinformation and has implemented output watermarks, though it notes that current watermarking techniques are not foolproof against determined adversaries.