The Next Frontier Opens
OpenAI began granting limited beta API access to GPT-5 to enterprise customers enrolled in its research partnership program this week. The model features a 1 million token context window — sufficient to process entire codebases, legal document portfolios, or extended research archives in a single API call — alongside significantly improved performance on multi-step reasoning tasks that prior GPT models handled inconsistently.
Benchmark Numbers
On GPQA (Graduate-Level Google-Proof Q&A), GPT-5 scores 88.3% versus GPT-4o's 72.1%. On MATH-500, GPT-5 reaches 96.4%. On HumanEval coding benchmark, scores cross 98% — a first for the GPT series. These numbers position GPT-5 at or near the top of all published model evaluations on academic benchmarks, though real-world task performance has additional dimensions.
The Context Window in Practice
The 1 million token context eliminates the retrieval-augmented generation (RAG) pipeline for many enterprise use cases. Legal teams testing the beta report loading entire contract portfolios and asking cross-document questions directly. Software teams can pass entire repository codebases for architectural analysis or security audits. The engineering simplification of removing chunking and retrieval infrastructure reduces both latency and cost for many applications.
Access and Pricing
General availability is expected Q4 2026. Beta pricing is undisclosed, but participants report approximately 40% premium over GPT-4o per million tokens, which many indicate is offset by infrastructure cost savings from eliminated retrieval systems.
What It Means for the Market
GPT-5 re-establishes OpenAI's performance leadership, which had narrowed as competitors released strong models. The million-token window and reasoning improvements are genuine capability advances, not incremental tuning.