---
format: "aidr-story-markdown/v1"
id: "407e97e8876a4efbf640874a67d5e3842eae29ff876d7e86059bffc34e96f20a"
canonical_url: "https://aidr.today/407e97e8?lang=en"
title: "OpenAI's Astra Scores 77.3% on Browser Use Benchmark v2 Versus Opus 5's 50.5%"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-09-06T14:38:07.000Z"
category: "Models"
topics: ["openai","astra","benchmark","browser-use","agent","inference"]
source_urls: ["https://huggingnews.com/ai/openais-astra-scores-773percent-on-browser-use-benchmark-v2-versus-opus-2385c4b2","https://x.com/SebastienBubeck/status/2096489923033588176","https://x.com/npew/status/2096595548753350807","https://x.com/RhysSullivan/status/2096540986596048907","https://x.com/teortaxesTex/status/2096494062769434644","https://x.com/scaling01/status/2096589364855722031","https://x.com/htihle/status/2096581452582265071","https://x.com/koltregaskes/status/2096516000892194974"]
summary: "Early user tests showed OpenAI's GPT-6 Astra handling browser, desktop and visual tasks directly from prompts. On Browser Use Benchmark v2, Astra medium scored 77.3% versus Opus 5's 50.5%, and Astra high scored 92.9% on WeirdML, matching Fable 5.1 max for the top score. Separate demonstrations showed Astra rebuilding an apartment display and adding a telescope feed in about 20 minutes, and other users said it solved logic puzzles without code. Some users also said Astra consumed tokens quickly, with one Pro x20 subscriber saying a day of Astra Light use nearly exhausted the plan's weekly usage."
---

# OpenAI's Astra Scores 77\.3% on Browser Use Benchmark v2 Versus Opus 5's 50\.5%

> [Open the canonical story](<https://aidr.today/407e97e8?lang=en>)

**Published:** 2026-09-06T14:38:07.000Z
**Category:** Models
**Topics:** openai, astra, benchmark, browser\-use, agent, inference

## Summary

Early user tests showed OpenAI's GPT\-6 Astra handling browser, desktop and visual tasks directly from prompts\. On Browser Use Benchmark v2, Astra medium scored 77\.3% versus Opus 5's 50\.5%, and Astra high scored 92\.9% on WeirdML, matching Fable 5\.1 max for the top score\. Separate demonstrations showed Astra rebuilding an apartment display and adding a telescope feed in about 20 minutes, and other users said it solved logic puzzles without code\. Some users also said Astra consumed tokens quickly, with one Pro x20 subscriber saying a day of Astra Light use nearly exhausted the plan's weekly usage\.

## Sources

- [Story source](<https://huggingnews.com/ai/openais-astra-scores-773percent-on-browser-use-benchmark-v2-versus-opus-2385c4b2>)
- [Story source](<https://x.com/SebastienBubeck/status/2096489923033588176>)
- [Supporting source](<https://x.com/npew/status/2096595548753350807>)
- [Supporting source](<https://x.com/RhysSullivan/status/2096540986596048907>)
- [Supporting source](<https://x.com/teortaxesTex/status/2096494062769434644>)
- [Supporting source](<https://x.com/scaling01/status/2096589364855722031>)
- [Supporting source](<https://x.com/htihle/status/2096581452582265071>)
- [Supporting source](<https://x.com/koltregaskes/status/2096516000892194974>)