Breaking

OpenAI's GPT-6 Astra Edges Claude Opus 5 in Published Agent Benchmarks, With Cost and Safety Caveats

September 16, 2026 at 23:46 EDT

Foundation Models
AI Agents
Software Dev & Coding

OpenAI has introduced GPT-6 Astra as its most capable agentic model yet, and the company’s published benchmark table puts it ahead of Anthropic’s Claude Opus 5 on several shared computer-use, professional-work and coding tests. The comparison is not a clean knockout: independent score aggregators still show overlapping uncertainty ranges, and Claude Opus 5 keeps a lower standard API price.

September 2026 - OpenAI

Astra Pushes the Agent Model Race

GPT-6 Astra leads Claude Opus 5 on several published computer-use and coding benchmarks, while Opus 5 keeps a clear price advantage and remains competitive in independent lanes.

59.3%

Astra on Agents' Last Exam

55.5%

Claude Opus 5 on the same test

$10/$50

Astra input/output per MTok

$5/$25

Opus 5 input/output per MTok

Published agent benchmark edge

Agents' Last Exam compares complex professional tasks in real software. The columns are scaled to the reported percentages.

59.3%

GPT-6 Astra

55.5%

Claude Opus 5

The caveat is oversight

Astra reached OpenAI's Critical cybersecurity threshold and is more capable at computer use, but its written reasoning became harder to monitor in adversarial tests. Extra checks can pause or stop legitimate work.

Capability

Computer use, coding, research, cyber

Risk control

Monitoring, refusals, trusted access

Why teams care

Higher published scores on several agentic work tests
Large context for tool-heavy research and coding
Launch users report clearer coding artifacts

Why teams hesitate

Opus 5 standard token pricing is half as high
Independent intervals overlap in overall scores
Safety checks can interrupt legitimate work

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.