This website uses cookies

Read our Privacy policy and Terms of use for more information.

OpenAI released GPT-6.2 Astra on November 17, 2026, and early reviews split quickly. On paper it costs the same per token as Anthropic's Claude Fable 5.2 ($8 per million input tokens and $40 per million output tokens on both official pricing pages), but several testers report that Astra finishes the same jobs for a fraction of the total cost because it uses far fewer tokens. Fable still leads the independent index and long creative builds. The practical answer for a solo creator or small team: route high-volume work to Astra, keep creative and open-ended work on Fable, and give Astra strict constraints.

Why the reviews are so split

The first wave of reaction was cool. The research sweep behind today's episode found prominent reviewers, including Theo from T3, Dan Shipper at Every and Matt Schumer, citing steering difficulties, compressed writing, and a habit of stopping short on long tasks. Nick Saraev's video, "GPT-6.2 Astra Just Dropped, And It's Not Great," drew about 106,000 views within a day.

Later hands-on tests told a different story. Nate Herk, Chase AI and Pat Simmons all published side-by-side comparisons within a day, and the common finding was that Astra is not weak. It behaves differently. It thrives on concrete, constrained tasks and disappoints people who expect it to take a vague goal and run with it.

Same price per token, different price per job

Token price is only half of the bill. The other half is how many tokens a model burns to finish the work.

  • Independent index: Artificial Analysis currently ranks Fable 5.2 at about 65 and Astra at 62 on its Intelligence Index, so Fable leads on broad quality. Fable also leads FrontierCode 2.0, 49.1% to 44.2%.

  • Cost per task: Kingy.ai reports 78.3% on Agents' Last Exam at $2.89 per task for Astra, against 72.8% at $60.05 for Fable. Those figures are vendor-adjacent and not independently confirmed, so treat them as a direction, not a guarantee.

  • Reviewer tests: In the multi-step builds reviewed for today's episode, testers reported Astra finishing 70% to 95% cheaper than Fable, and in far less time (roughly 12 to 50 minutes against one to five hours). Results vary by task, and the independent index implies a smaller gap of roughly 4x.

One more caveat from the sweep: Artificial Analysis reports a 59% hallucination rate for Astra on its AA-Omniscience test at max effort, so a cheap run is not the same as a correct run.

Why Astra stops early on vague prompts

Reviewers describe Astra as taking the shortest path to a finished state. Given a loose prompt such as "build a game," one tester reported Fable spending hours on a playable 3D world while Astra returned a flat 2D shooter in minutes. On loose prompts like "build a AAA shooter," Astra often marked the work complete after 15 to 30 minutes instead of polishing.

The fix is a blueprint. When the prompt includes strict parameters, a design system or a visual reference, testers reported much stronger output. One reviewer's 3D mechanical watch website, built from a tightly specified prompt, is the example cited most often.

Where Astra wins and where Fable still leads

  • Astra: browser and computer use, code review and static analysis, high-volume agent runs where cost per task matters. In one head-to-head bug hunt, a reviewer reported Astra finding 84 confirmed bugs for $11, against 71 for Fable at $257 (a single test, not independently verified).

  • Fable 5.2: long autonomous builds, richer animation and visual taste, and open-ended creative work. Tech With Tim reported giving it one prompt overnight and waking to a 3D game, a programming language and a multiplayer whiteboard. The trade-off is usage: his three builds used about 25% of his weekly Fable limit in roughly four hours.

The self-improvement milestone

OpenAI's release text says Astra is its first externally released model to meet the High threshold for AI self-improvement, meaning OpenAI rates it as able to accelerate AI research. OpenAI describes the effect as equivalent to giving each of its researchers an additional capable research engineer. The same text notes why it tracks this capability: models that help develop future AI systems could introduce new risks. Reports say Anthropic added safeguards limiting similar skills in Fable 5.2.

What to do today

  1. Route by job type. Send code review, QA passes and browser tasks to Astra. Keep creative direction, long builds and nuanced editing on Fable.

  2. Write a blueprint, not a wish. For design or build prompts, paste a strict visual reference and clear completion criteria. Avoid "be creative" and "go nuts."

  3. Keep a human sign-off. Astra optimizes for brevity and can skip edge cases. Review anything customer-facing or financial before it ships.

FAQ

Is GPT-6.2 Astra cheaper than Claude Fable 5.2?
Per token, no: both list at $8 input and $40 output per million tokens. Per finished job, testers report Astra costing much less because it uses fewer tokens, with reported savings between roughly 4x and 20x depending on the test.

Which model is better, Astra or Fable 5.2?
It depends on the job. Fable leads the independent Intelligence Index and long creative builds. Astra leads on speed and cost for code review, browser automation and other high-volume agent work.

Why do some people say Astra is bad?
Most complaints involve vague, open-ended prompts, where Astra tends to take the shortest path and stop early. Reviewers who gave it strict constraints reported much better results.

How should a small business use both models?
Use a two-tier setup: Astra for volume tasks where cost per run matters, Fable for creative and exploratory work, and a human check before anything goes out.

Companion episode:

Want more plain-language AI breakdowns and automations like this? Get them free at joebuildsai.com.

Reply

Avatar

or to participate