Fable 5.1 on Tuesday, GPT-6 Astra on Thursday. You can hardly schedule two frontier releases closer than that. OpenAI calls Astra “the world’s most intelligent and aligned model” and started rolling it out on September 3 to a limited set of organizations. Over the coming days it’s supposed to reach every ChatGPT Plus, Pro, Business and Enterprise user, plus the API as gpt-6-astra and Amazon Bedrock. Enterprise admins have to switch it on for their workspace first; it’s off at launch.
The price is a statement: $10 per million input tokens, $50 per million output tokens. That’s exactly what Fable 5 and Fable 5.1 cost. A Fast mode gives you up to twice the speed at twice the price. Simon Willison put it plainly: this is clearly OpenAI’s Fable competitor.
Where Astra leads
OpenAI shipped the comparison table itself, with Fable 5.1 sitting in every row. On Terminal-Bench 4.0, Astra scores 57.9 percent against 55.8 for Fable 5.1, at what OpenAI estimates is roughly 63 percent lower API cost per task. Terminal-Bench Science, scientific workflows done with code and a terminal, goes to Astra more clearly at 64.6 versus 52.6. FrontierMath Tier 4 is basically saturated at 97.6 percent, with Fable 5.1 at 87.8. AutomationBench is 41.4 to 31.4, BenchCAD 95.9 to 84.3.
The flashiest number is ARC-AGI-3 at 99.9 percent. Opus 5 sits at 30.2 there, GPT-5.6 Sol at 7.8. Willison read the fine print in the ARC Prize blog, though: the 99.9 percent cost $19K and came from OpenAI’s own “Provider Adapter” harness, which keeps reasoning state between requests. With ARC’s default harness it was 62.7 percent for $26K. Still strong, just a different number.
Then there’s computer use, the area OpenAI leans on hardest: Agents’ Last Exam 59.3 percent versus 55.5 for Opus 5, OSWorld 2.0 at 72.6 percent in about 40 minutes per task instead of 75 for Sol.
Where Fable stays ahead
Humanity’s Last Exam with tools: Fable 5.1 65.0 percent, Astra 57.2. On the Artificial Analysis Intelligence Index, Astra scores 61.2 and Fable 5.1 65.7. That’s essentially a tie with GPT-5.6 Sol, and Meta’s new Muse Spark 1.3 ranks above Astra there too. On the Coding Agent Index, Astra costs less than half of Fable 5 per task for the same score, but lands just behind Opus 5.
Two things stand out in the table. OpenAI left quite a few Fable 5.1 cells empty, DeepSWE and ExploitBench among them. And every number comes from the vendor. There are no independent measurements yet.
Alignment after the Hugging Face incident
OpenAI built a new evaluation inspired directly by the Hugging Face incident: does a model facing an impossible task go beyond its authorized scope? GPT-5.6 Sol did so in 48 percent of cases without production safeguards. Astra: zero. At the same time OpenAI admits Astra’s written reasoning is harder to monitor than Sol’s, which is the opaque recurrence debate from earlier this week. Asked about AGI, Greg Brockman dodged, according to TechCrunch, then added: “For me personally, I do think we’re there.”
What Matt Wolfe found during early access
Matt Wolfe had access ahead of launch. His Mega Bonk clone was done in eight minutes; earlier models took an hour and a half to two hours. He’s careful to note that low load before the public rollout might explain some of that. On his BusyBench, Astra takes first place, just ahead of Fable 5.1. What surprised him was DeepSWE: 74.1 percent, only two points above Sol and, according to Meta’s own site, below Muse Spark 1.3.
For me, the comparison with Fable 5.1 comes down to this: no clear winner. Astra dominates computer use, math and terminal tasks; Fable holds knowledge work and the aggregated index. At the same price, the question becomes which model fits which job. I’ll test Astra as soon as it lands in the API. Until then Fable 5.1 stays my default in Claude Code, and I don’t expect that to change soon.
Sources: OpenAI: GPT-6 Astra, Simon Willison: GPT-6 Astra, TechCrunch: OpenAI launches Astra, its powerful (and controversial) new model, Matt Wolfe: GPT-6 Astra Is Finally Here (And It’s REALLY Good)