On August 8, 2025, OpenAI released GPT-5. The core announcements: a major API price cut, a new Agentic Coding capability, and no multimodal upgrade in this release. Everything below is launch-day information; verify against OpenAI’s official docs before relying on it.

Pricing: a significant API price cut

OpenAI stated the GPT-5 API price is roughly 1/15 of Claude Opus 4.1 (announcement figure — check the official pricing page for exact numbers). For developers whose costs scale with API usage, this is the most direct change in this release.

Agentic Coding: from autocomplete to fixing a whole repo

Agentic Coding is the centerpiece of this release:

  • Given a vague request (e.g. “fix the bug in this project with a thousand files”), the model reads the codebase, understands context, and locates the problem on its own.
  • When its code fails on the first run, the model reads the error logs and iterates until it passes.
  • Live demo: a 400-line aircraft aerodynamics simulation, completed in about 2 minutes.

A useful analogy: earlier models were like an intern who needs you to clean up after them; Agentic Coding behaves more like a senior engineer who reads the codebase and verifies their own results.

Other launch-day numbers (to verify)

  • Hallucination rate: announced as roughly 45% lower than GPT-4o (wait for independent evaluation).
  • Context window: 400K (announcement figure).
  • No complete evaluation methodology was shown on stage; benchmark with your own data before committing.

Limitation: no multimodal update

GPT-5 does not support audio or video input/output in this release, and no image generation. Compared with multimodal-generation products of the same period (e.g. Google Genie 3), this release is clearly focused on text and coding.

Summary

For the developer ecosystem, this is a clearly positioned release: lower price plus stronger coding ability, directly affecting API application costs and the ceiling of coding agents. For those expecting an “all-purpose multimodal model,” it falls short. Both judgments are based on launch-day information; follow official docs and independent evaluations.

Follow the Rexai Programming WeChat account for Rust / AI technical content.