What early testers are actually doing with GPT-6 Astra

Trả lời trực tiếp

Nine verified launch-week quotes on GPT-6 Astra: a code-review win, harness-dependent benchmarks, uneven computer use, and mixed verdicts on price. Hướng dẫn này dành cho các đội sản phẩm và nền tảng đang so sánh chất lượng model, chi phí, chính sách định tuyến, và rủi ro ra mắt.

Launch week, in quotes you can check

Reaction to GPT-6 Astra splits by what each tester actually ran. The catalog frame first: openai/gpt-6-astra lists at $10 input / $50 output per million tokens (USD) with $1/$2 cache hits for prompts up to 272,000 tokens and $20/$75 above that, a 1,050,000-token context window, 8,192 max output tokens (2,048 default), Chat Completions and Responses API support, and Thinking, Streaming, Computer Use, and Function Calling capabilities. Every quote below is verbatim, attributed, dated, and linked to a live source, grouped by test subject: code review, benchmarks, computer use, and price. The verdicts disagree, and the disagreement is the finding.

Code review: the clearest published win

CodeRabbit's launch-week evaluation (2026-09-04) is the most concrete published claim in Astra's favor: "Astra caught approximately 4% more labeled bugs through actionable findings than GPT-5.6 Sol, and 22% more than Opus 5" (https://www.coderabbit.ai/blog/gpt-6-astra-code-review-evaluation). Their write-up also reports the gap widening on cross-file reviews: 20% over GPT-5.6 Sol and 33% over Opus 5. Hands-on notes point the same way. Claire Vo, writing in Lenny's Newsletter (2026-09-03): "I got early access to GPT-6 Astra: when I say this model broke through tasks I couldn't crack with 5.6 Sol or Fable, I mean it specifically" (https://www.lennysnewsletter.com/p/gpt-6-astra-is-a-banger-heres-everything). One vendor evaluation plus one practitioner anecdote is a signal, not proof — both test coding, and neither has tested your repository.

Benchmarks depend on the harness

ARC Prize (Greg Kamradt, 2026-09-03) published two Astra results on ARC-AGI-3 Semi-Private, and the spread is the story: "With the Provider Adapter harness, Astra (high) scores 99.9% for $19K." Under the Standard harness the same post reports 62.7% for $26K (https://arcprize.org/blog/astra), and ARC states plainly that it is not claiming AGI. Readers unpacked the gap anyway. On the thread covering the post, Hacker News user NiekvdMaas (2026-09-03) wrote: "First chart: from score 61 (GPT-5.6 Sol) to drumroll 61 (GPT-6 Astra)" (https://news.ycombinator.com/item?id=49556147), and user emp17344's response to the ARC numbers was: "99.9% with the right harness? Ok, we're at AGI then." (https://news.ycombinator.com/item?id=49555691). Artificial Analysis (2026-09-03) added a capability datapoint, reporting "a significant decrease in hallucination rate from 92% to 51% at max effort" (https://artificialanalysis.ai/articles/benchmarking-gpt-6-astra). Read the harness label before the score; every number in this section belongs to the outfit that produced it.

Computer use: early and uneven

OpenAI's published OSWorld evaluation lists Astra at 72.6% — the vendor's own number on the vendor's own benchmark, not an independent measurement and not a NextModel run. The early hands-on record is rougher. Hacker News user Topfi (2026-09-05): "GPT-6 Astra (on Light) stops a lot and often in unintuitive ways that I haven't seen in a while" (https://news.ycombinator.com/item?id=49571621). The catalog lists Computer Use among Astra's capabilities, so the launch-week pairing is: high vendor score, uneven practitioner experience. Teams building on it should plan for the stops — retries, checkpoints, and a human-readable log of what the agent did before it quit.

Pricing: the 272K cliff dominates the conversation

DataCamp's comparison (2026-09-05) frames the market: "GPT-6 Astra and Claude Fable 5.1 both cost $10/$50 per million tokens, but win on different scoreboards." (https://www.datacamp.com/blog/gpt-6-astra-vs-claude-fable-5-1). Under most complaints sits the tier structure: listed rates of $10/$50 per million tokens up to 272,000 prompt tokens, $20/$75 above, with cache hits at $1 and $2. Rate limits drew the sharpest single comment. Hacker News user kbrannigan (2026-09-05): "It's very expensive. After 15 message I burned through my 5 hour limits." (https://news.ycombinator.com/item?id=49571621). The API-side discipline is the same either way: stay under the 272K boundary where the work allows, cache repeated prefixes, and route routine traffic to cheaper catalog legs. Worked numbers are in our pricing explainer (/blog/gpt-6-astra-pricing-explained), and the series hub is /gpt-6.

What we cannot verify

This post quotes other people's measurements and contains none of ours. The red lines:

• Every quoted number belongs to its source: CodeRabbit, ARC Prize, Artificial Analysis, and DataCamp ran their own evaluations, and OSWorld 72.6% is OpenAI's published figure. NextModel ran none of these benchmarks.
• Hacker News commenters are pseudonymous handles. We can verify that the quoted text appears on the linked threads, not their setups, limits, or prompts.
• X threads circulating during launch week were left out on purpose: we could not verify their text against a live source, and unverified screenshots fail this site's citation bar.
• Quote text was checked against the linked pages on 2026-09-06. Links rot; if one dies, treat the quote as unverifiable.
• Catalog facts in the opening section are the NextModel listing for openai/gpt-6-astra as of publication.

FAQ