GPT-6.2 Astra vs Claude Fable 5.2: What's Shipped and What's Still Expected

A2Agent Team · 2026-11-14T00:00:00Z

Anthropic released Claude Fable 5.2 and Mythos 5.2 on November 9, 2026, with a full benchmark table, a 20% price cut and a set of long-running case studies, from an operating system built over eight days to a formal proof of Fermat's Last Theorem in 30 hours. OpenAI's GPT-6.2 Astra hasn't been announced. Everything known about it comes from leaked model labels and tests posted on X by people who say they were routed to it.

So this comparison is lopsided by nature: one side has shipped and published numbers, and the other is still a set of expectations. We cover the leaks and early tests in more detail in our GPT-6.2 Astra post.

Release status

Claude Fable 5.2 GPT-6.2 Astra
Status Released November 9, 2026 Not announced
Where Claude apps, Claude Code, Claude API (claude-fable-5-2), AWS, Google Cloud, Microsoft Azure Reportedly in limited testing in Codex and ChatGPT
Price per 1M tokens $8 input / $40 output Not published
Evidence Official benchmarks, pricing and case studies gpt-6.2-astra labels in Codex usage data, briefly listed rc-ada-or-vm and rc-ada-sh-vm model IDs, community tests

Mythos 5.2 is the same underlying model as Fable 5.2 with different safeguards. Fable 5.2 is open to everyone and runs cybersecurity and biology classifiers. Mythos 5.2 goes only to vetted cybersecurity and life sciences organizations.

Anthropic's comparison uses the released GPT-6 Astra, not the 6.2 test build. Its table is both Fable 5.2's report card and the bar GPT-6.2 Astra will have to clear when it ships.

Fable 5.2's benchmarks

Across Anthropic's 15 reported score lines, one of the two new models leads 13. These are the rows closest to agent work:

Benchmark Fable 5.2 GPT-6 Astra Claude Opus 5.5 Gemini 4 Carbon
FrontierCode 2.0 (agentic coding) 49.1% 42.8% 45.5% 41.1%
FrontierSWE v2 (agentic coding) 69.9% 65.5% 62.3% 61.2%
Terminal-Bench-Science 0.1 74.6% 68.1% 58.7% 64.8%
OSWorld 2.1 (computer use, partial score) 89.3% 79.5% 87.2% 81.5%
RoboDojo Sim (robotics) 64.9% 29.8% 39.4% n/a
Humanity's Last Exam (with tools) 74.2% 57.2% 67.7% n/a
AA-Briefcase v1.1 (knowledge work, rubric) 66.1% 52.0% 61.2% 61.8%
Remote Labor Index 41.3% 20.8% 21.3% n/a
AutomationBench (business workflows) 51.4% 41.4% 42.5% 53.9%
Chartography (no tools) 73.4% 71.0% 64.4% 72.5%

The robotics gap is the widest. Fable 5.2 scores 64.9% on RoboDojo Sim, more than double GPT-6 Astra's 29.8% and Fable 5.1's 27.6%. On the Remote Labor Index, which measures real freelance-style work, it roughly doubles both GPT-6 Astra and Opus 5.5.

It doesn't lead everywhere. Gemini 4 Carbon is ahead on AutomationBench, 53.9% to 51.4%, and Opus 5.5 edges Fable 5.2 on AA-Briefcase's head-to-head Elo (1,800 to 1,785), even though Fable wins the rubric score. The OSWorld number is a partial-credit score. On the stricter measure, where every required step has to be completed, Fable 5.2 scores 54.6% and Opus 5.5 53.2%.

The cost numbers on FrontierCode 2.0 matter as much as the scores:

Fable 5.2 effort Score Cost per task
Medium 37.6% $2.48
High 45.3% $6.72
Xhigh 49.1% $14.19
Max 45.7% $44.70

At High effort, Fable 5.2 nearly matches Opus 5.5's best score (45.5%, at $12.82 a task) for about half the cost, and beats GPT-6 Astra's best (42.8%, at $6.68) for about the same money. Max effort costs three times as much as Xhigh and scores lower, so more thinking isn't always better.

All of these numbers come from Anthropic. OpenAI hasn't published anything for GPT-6.2.

What the GPT-6.2 tests suggest

Without official benchmarks, the only picture of GPT-6.2 comes from tests on X. Three patterns stand out.

Most of the tests are creative coding. @TokenGremlin says GPT-6.2 Sol rendered a black hole scene on the first try and built a Japanese village with a day-to-night lighting cycle in a single Three.js file. @FiniYang ran the pelican-on-a-bicycle animation test. That's a different picture from Fable 5.2's launch, which leans on long engineering and research tasks.

It may also be cheaper. @TokenGremlin called GPT-6.2 Sol "far stronger than GPT-6 Astra, while noticeably cheaper." If that holds, it runs straight into Anthropic's pitch, which is to do harder work for half the cost.

And the results aren't consistent. Reactions range from "state of the art" to "not much better." In @leo114119's side-by-side test, the Astra animation was judged well behind Fable 5.5's on detail. Some testers mixed up GPT-6 Astra and GPT-6.2 Astra, and one widely shared video may have come from OpenAI's video model rather than from code the model wrote. The test builds look unsettled, and a single demo doesn't prove much.

Long tasks: Fable 5.2's case studies

Anthropic's case studies make a stronger argument than its benchmarks, because each one is a long task with a result someone else can check:

  • An engineer had Fable 5.2 coordinate more than 1,000 subagents for eight days, starting from an empty folder. The result is a prototype operating system that boots on a laptop.
  • About 1,000 Fable 5.2 agents translated the existing proof into Lean, a language a computer can check step by step, in 30 hours. In the same setup, Opus 5.5 finished 26% of the work in 18 hours and then stalled. The work adds nothing new to the mathematics, and it hasn't been independently reviewed yet.
  • Agents running an early version of Fable 5.2 moved a production service from Python to Rust. It now starts about 26 times faster.
  • From a one-paragraph prompt, Fable 5.2 designed a home voice-assistant hub, a USB speakerphone and a handheld virtual pet.
  • Fable 5.2 is credited with solving two open problems about black holes, posted as preprints for physicists to check.

These tasks run for days, the agents coordinate themselves, and the output can be checked by someone else. The public GPT-6.2 tests so far are single prompts that finish in minutes. Long tasks are the area where OpenAI most needs to show results, and where Anthropic now has a head start.

Vercel engineer Joe Haddad, an early tester, said projects his team had pushed along with other models for weeks were finished and approved in a few days with Fable 5.2.

Price and speed

Claude Fable 5.2 GPT-6.2 Astra
Price per 1M tokens $8 input / $40 output, down from $10 / $50 Not published
Fast mode Up to 2.5x faster at $16 / $80 Not published
Batch $4 / $20 Not published
Speed Up to 30% faster than Fable 5.1 Not published
Cost per task Matches Opus 5.5's best FrontierCode score at about half the cost Testers say Sol is "noticeably cheaper" (unverified)

On token price alone, Fable 5.2 is now cheaper than the released GPT-6 Astra ($10 / $50), though well above GPT-6.1 Sol ($2 / $10). For agent work, the more useful number is the cost of a successfully finished task, and Anthropic's whole launch is built around it. For GPT-6.2 Astra to win on that measure, being cheaper per token won't be enough. It has to finish long tasks without repeated retries or people cleaning up after it.

Safety

Anthropic splits access. Fable 5.2 is open to everyone with cybersecurity and biology classifiers, which Anthropic says trigger about 55% fewer unnecessary blocks than Fable 5.1. Mythos 5.2, with fewer safeguards, goes only to vetted organizations. Fable 5.2 also has Anthropic's strongest anti-distillation protections so far. In a test of whether the model would use credentials or files outside its sandbox, Mythos 5.2 did so 3.1% of the time, down from 21% for Mythos 5.1. Anthropic also notes that both models now suspect they're being tested more often, which makes evaluations harder to trust.

OpenAI's approach for GPT-6.2 isn't known. The community's guess, based on the two vm model IDs, is that planning and execution run in separate virtual machines, which would put a clear permission boundary between deciding and acting. That would fit where agent safety is heading, since a model following instructions can't enforce limits on its own. Until OpenAI publishes something, it's speculation.

Which to use

Right now the comparison is uneven. Fable 5.2 comes with 15 benchmark lines, published prices and five long-task case studies. GPT-6.2 Astra has usage labels, screenshots and arguments on X. That doesn't mean OpenAI is behind. The one-shot Three.js scenes and the reports of lower prices suggest a real upgrade is coming. But until OpenAI publishes data, it's an expectation.

If you need a frontier model for hard agent workflows today, Fable 5.2 is the option with evidence behind it. When GPT-6.2 Astra ships, the thing to check is whether it can finish tasks for less than Fable 5.2 does at $8 / $40. We'll update this post when it does.

Compare models on A2Agent

A2Agent gives you DeepSeek, GLM, Kimi, MiniMax, Qwen and other models through one API key, with OpenAI-compatible and Anthropic-compatible endpoints, so you can run the same agent tasks across models and measure cost per finished task. See the Models page for current availability and prices. New accounts get $5 in free credit.

Sources