SAN FRANCISCO — OpenAI is in the final stages of preparing GPT-5 for public release, and internal benchmarks reviewed by GlanceFeed suggest the next-generation model represents a qualitative leap over GPT-4 that even the company's own researchers find surprising. GPT-5 is natively multimodal, processing text, images, audio, and video as first-class inputs rather than bolted-on additions. It demonstrates "agentic" capabilities — the ability to plan multi-step tasks, use external tools, and verify its own outputs against trusted sources. On a comprehensive science benchmark, the model reportedly passes at the PhD level in physics, chemistry, and biology.

OpenAI has been unusually transparent about the safety measures surrounding GPT-5. The model has undergone six months of red-teaming by external experts, including adversarial testing for misuse in cybersecurity, biological weapon design, and mass influence operations. A new "Constitutional Alignment" layer — inspired by Anthropic's research — encodes a set of behavioral principles directly into the model's training objective.

Pricing and release remain guarded, but enterprise partners have been notified to expect availability by September 2026. A free tier with rate limits and a "GPT-5 Pro" subscription at $30/month have been discussed internally. The most significant unknown is whether GPT-5 will be open-weight: OpenAI leadership is reportedly divided, with CTO Mira Murati advocating for limited open release under a responsible use license.