GPT-6 Astra vs Claude Fable 5.1: The 72 Hours That Reset the AI Race

What actually shipped in those 72 hours

Between September 1 and September 3, 2026, the two labs at the front of the industry both shipped a flagship model, and in the middle of it most of the world briefly lost access to every assistant they use.

Anthropic went first. On September 1 it released Claude Fable 5.1 and Claude Mythos 5.1, describing them as the same model with different levels of safeguards. Fable 5.1 is generally available. Mythos 5.1 is not.

Two days later OpenAI launched GPT-6 Astra and told the world it had entered a new phase. Greg Brockman's framing was blunt: "Welcome to the AGI era."

Between those two announcements, on the morning of September 3, Azure's East US region failed and took ChatGPT, Claude, Grok, and Microsoft's own Copilot down with it.

That sequence is the story. Two labs claiming the frontier, priced identically, separated by a few benchmark categories, and both running on infrastructure that failed at the same time.

The morning every AI assistant stopped working

The failure started at roughly 6:58 a.m. ET on September 3. By 7:53 a.m. ET, Downdetector had logged more than 5,000 ChatGPT reports, and volumes peaked around 11:00 a.m. ET.

ChatGPT drew more than 37,000 reports on its own. Counted together with Codex, the figure passed 66,000. OpenAI's status page listed 15 affected ChatGPT components, including conversations, login, search, file uploads, voice mode, image generation, Deep Research, and Agent.

Claude peaked at 1,324 reports and Grok at 1,365. Both recovered inside about half an hour. Everything was back to normal by 12:42 p.m. ET.

The interesting part is who stayed up. Google Gemini logged around 500 reports and kept running, because it sits on Google's own infrastructure rather than a third-party cloud. Microsoft Copilot, running on Microsoft's own cloud, did not stay up, which points at shared infrastructure layers rather than any single customer's configuration.

For anyone shipping on top of these APIs, that is the lesson worth keeping. Three competing vendors failed together because they shared a region, so multi-model routing is only real redundancy if the models sit on different clouds. It is the same concentration risk driving the current wave of AI data center expansion.

GPT-6 Astra and the AGI claim

OpenAI positions Astra as a generational jump rather than an increment, and the pitch centres on computer use. The model navigates software the way a person does: filling forms, updating CRM records, managing calendars, operating spreadsheets and Power BI, and driving engineering tools such as KiCad and FreeCAD.

The numbers behind that claim are strong. Astra scores 72.6 percent on the OSWorld 2.0 offline subset at roughly 40 minutes per task, against GPT-5.6 Sol at 65.7 percent and about 75 minutes. That is a better result in a little over half the time. On ScreenSpot-Pro it reaches 92.7 percent.

Brockman's explanation for why this matters was that the bottleneck was never the model. "We've been bottlenecked by people writing connectors," he said. A model that can drive a normal desktop does not need one.

The safety side of the launch is louder than the capability side, though. Astra is the first model OpenAI has classified at the Critical cybersecurity threshold under its Preparedness Framework, meaning it can autonomously find unknown vulnerabilities and build exploit chains. It found two previously unknown zero-days during evaluation and scored 100 percent on ExploitBench.

Chief scientist Jakub Pachocki paired the AGI framing with a caveat worth repeating: "Progress in intelligence does not guarantee progress in alignment." That tension is exactly what AI governance work in 2026 has been trying to get ahead of, and it is now a product constraint rather than a policy question.

Fable 5.1 vs GPT-6 Astra: how the benchmarks compare

Neither model sweeps. Leadership swaps by category, which is a meaningful change from the previous generation covered in our GPT-5.6 versus Fable 5 comparison.

Read that table by column and Astra looks like the maths and computer use model, while Fable 5.1 looks like the reasoning and coding agent model. Astra also reports 88.0 percent on SRE-Bench in one attempt and 96.3 percent on long context in the 512K to 1M band, and it effectively saturates ARC-AGI-3, reported between 98.6 and 99.9 percent depending on the harness.

Anthropic's own standout number is elsewhere. Fable 5.1 hits 52.6 percent on Terminal-Bench-Science 0.1, against 24.7 percent for Fable 5, 29.0 percent for Opus 5, and 22.4 percent for GPT-5.6 Sol. Mythos 5.1 pushes Terminal-Bench 4.0 to 60.9 percent, ahead of both public models.

The customer signal matters too. Cognition said it was moving its Opus 5 traffic in Devin to Fable 5.1 on launch day, and Jane Street reported that Fable 5.1 solves more of its coding problems than either Fable 5 or Opus 5.

What these models actually cost

The headline rates are the same, which makes the surrounding structure the real decision.

Anthropic cut cache reads by 75 percent, and that single change is what moves the bill. It says typical workloads land about 25 percent cheaper than Fable 5, and heavily agentic workloads up to 45 percent cheaper. If your application replays a long, stable prefix on every turn, that is the number to model.

OpenAI argued the opposite framing. "Pricing tokens doesn't make any sense," Brockman said, pushing price per completed task as the honest metric instead. There is something to that, given Astra finishes computer use tasks in roughly half the wall clock time of its predecessor.

Both positions are self-serving, and both are partly right. Measure cost per completed task on your own workload, then check whether cache hits are doing the work you think they are.

Mythos 5.1, watermarking, and who gets the keys

The most consequential detail of the Anthropic launch is not a benchmark. It is that the strongest configuration is not for sale.

Mythos 5.1 runs the same model as Fable 5.1 with safeguards tuned for cybersecurity and life sciences, work the general safeguards would otherwise refuse. Access is limited to vetted US organisations through the Cyber Verification Program and the Life Sciences Verification Program, the latter run in partnership with the US government.

Anthropic also reports that the general safeguards got less annoying. Cybersecurity classifiers fire with 60 percent fewer false positives, and biology safeguards trigger 85 percent less often on benign requests than Fable 5 did.

Then there is provenance. Fable 5.1 embeds an invisible numerical watermark in generated text, with a detection API in private preview for regulators, law enforcement, researchers, media, fact-checkers, and educational institutions. That is EU AI Act groundwork, and it is the first time a frontier lab has shipped text watermarking as a default rather than a research demo.

On the OpenAI side, access is staged too. Astra opened through a gated programme for enterprise customers before rolling out to ChatGPT Plus, Pro, Business, and Enterprise, with a separate GPT-6 Astra Pro tier for the higher plans. Enterprise admins enable it per workspace, and it is off by default.

What to do with all this if you are shipping

The practical takeaway is smaller than the headlines suggest, and more useful.

First, stop treating the leaderboard as a purchasing decision. When two models split categories this evenly, your own evaluation set is the only thing that resolves the tie. Run both against the tasks you actually serve, at the effort levels you can afford.

Second, price the cache, not the model. Identical per-token rates mean the winner on your invoice is decided by cache read pricing and how well your prompts hold a stable prefix. That is a prompt architecture problem before it is a vendor problem.

Third, treat the September 3 outage as a design requirement. If your fallback model shares a cloud region with your primary, you do not have a fallback. The teams that stayed online were the ones already routing across providers on different infrastructure.

Fourth, plan for gated capability. Between Mythos 5.1's verification programmes and Astra's Critical cybersecurity classification, the strongest models now come with access review attached. If your roadmap depends on frontier security or life sciences capability, the paperwork is now part of the timeline, which is a shift teams following agentic coding in 2026 should budget for.

The 72 hours did not settle which lab is ahead. They settled something more useful, which is that the answer now depends entirely on what you are building.