On July 8, 2025, OpenAI released GPT-5, describing it as the first model that unifies the strengths of its separate reasoning (o-series) and chat (GPT-4o) families into a single system. Where o3 and o4-mini exposed reasoning as an explicit mode with a separate model ID, GPT-5 treats reasoning depth as a continuously tunable parameter within one model, eliminating the developer workflow of routing some requests to a GPT-4o endpoint and others to an o-series endpoint depending on task difficulty.

The headline technical result is a 96.7% score on GPQA Diamond, the graduate-level science evaluation, in its highest-effort configuration — a jump of roughly 3 points over o3-high. On SWE-bench Verified with the full tool scaffold, GPT-5 scores 84.2%, leading the leaderboard at launch over Claude 3.9 Sonnet and Gemini 2.7 Pro. On MMMLU it reaches 96.5% and on AIME 2025 it achieves 94.7%. OpenAI acknowledges that compute-intensive settings used for benchmark evaluations are not available through the standard API, and publishes API-accessible scores at each reasoning effort level.

The API surface extends the reasoning_effort parameter from the o-series to GPT-5, accepting 'low', 'medium', and 'high' values. At 'low', GPT-5 behaves comparably to GPT-4o in latency and output style, making it a viable drop-in replacement for latency-sensitive applications. At 'high', it engages extended chain-of-thought reasoning with token budgets in the thousands, matching the behavior of o3-high on complex tasks. OpenAI prices the model at $11 per million input tokens and $32 per million output tokens at standard effort, with reasoning tokens counting as output.

GPT-5 is multimodal from launch, accepting text and images and producing text, with audio I/O planned through the gpt-realtime endpoint later in 2025. The model also supports all Responses API capabilities at launch: stateful conversations, background mode, built-in tools (web search, file search, code interpreter, computer use), structured outputs, and prompt caching. OpenAI explicitly positions GPT-5 as the model new Responses API builds should default to, ahead of GPT-4o and the o-series, which remain available but are no longer the primary recommendation.

Anthropic, Google, and xAI all responded to the launch with benchmark comparisons emphasizing areas where their models lead — Claude 3.9 Sonnet at coding instruction following, Gemini 2.7 Pro at very long context tasks, Grok 3 on certain scientific reasoning evals — reflecting the close clustering of frontier capabilities in mid-2025. Independent evaluators including METR noted that the SWE-bench margin is within the confidence interval of repeated runs and advised against treating any single percentage point difference as definitive.

The GPT-5 launch came with updated safety documentation from OpenAI's preparedness team, including a model card disclosing uplift assessments across CBRN, cyberoffense, and persuasion categories, all rated at 'medium' or below with mitigations applied. OpenAI also published a GPT-5 system card summarizing the red-teaming program, external evaluator findings, and the deployment decisions made in response, continuing the model card format it introduced with GPT-4. Developer access began rolling out immediately on July 11, with Plus and Pro users receiving access through the chat interface in parallel.