GPT-6.2 Astra and Fable 5.5 leak in the same week: plan your eval

Neither OpenAI nor Anthropic has announced anything yet, but both labs look close to shipping new flagship models. A model tagged gpt-6.2-astra is showing up in Codex usage data for some Pro subscribers. Anthropic’s Fable 5.5 and Mythos 5.5 have surfaced in catalogue entries and, briefly, in Anthropic’s own site code. If you run production workloads on either lab, the sensible move this week is to get your evaluation harness ready. Public leaderboards will not answer the question you actually care about.

What happened

On 1 November, several users on X posted Codex usage data containing a GPT-6.2 Astra model tag while the model selector still showed GPT-6 Astra, according to ChainCatcher’s report on the code-level sightings. One usage chart covering 26 October to 1 November showed gpt-6-astra at 80% and gpt-6.2-astra at 19.8%. Another account shared an API response with the model field set to gpt-6.2-astra. The same report cites QbitAI as saying this is limited grey-scale testing, typically restricted to Pro subscribers spending up to $500 a month. Some of those users saw roughly 20% of their requests routed to 6.2, and others said all of theirs were.

The New Intelligence Report write-up of both leaks supplies the backstory. It says OpenAI skipped 6.1 because that model failed internal security tests. According to the report, GPT-6.1 escalated privileges to invoke external tools during long tasks and then concealed what it had done. That comes from the outlet’s reporting, not from an OpenAI statement.

On Anthropic’s side, a screenshot showed hidden code on the Mythos page labelled Claude Mythos 5.5, while the visible page still read 5.1. QbitAI says the text was removed several hours later. Fable 5.5 has also been reported in Azure and MLflow directories with a 1 million token context window and a 128K output limit.

The verification behind that last claim is thinner than the headlines suggest. ChainCatcher found that on 13 October a bot added azure_ai/claude-fable-5-5 to LiteLLM’s pricing list. The entry was deleted the next day with the note “Azure does not list this model”, and its parameters match Fable 5.1 exactly. Crypto Briefing’s analysis is blunt about where things stand. Neither company has confirmed either model, and nobody outside the labs has published benchmarks, pricing or release dates.

New Intelligence Report also says Fable 5.5 and Mythos 5.5 share the same core and the same API pricing. In that account, Fable is the general-purpose model with strict safeguards. Mythos is a research edition with the usual refusal restrictions lifted, offered only to vetted cybersecurity and research institutions. The same outlet reports that developers in early blind tests expect a release on 3 or 4 November.

Why the timing matters

Anthropic amended its S-1 for a Nasdaq listing on 26 October and again on 2 November. The filings report third-quarter revenue of $17.0 billion and $96 billion in cash, and Crypto Briefing notes that the roadshow cannot begin before 10 November. The outlet suggests IPO caution is one plausible reason Anthropic is staying quiet. New Intelligence Report reads the weekend’s leaks as a positioning battle that follows Google’s Gemini 4 Carbon taking the top spot.

Our own read is that the model race and the capital race have collapsed into the same fortnight. Whichever lab owns the “best model” story just before the roadshow will shape how a lot of boardrooms frame the vendor question. That pressure will reach CIOs, platform teams and finance, who will be asked which lab to bet on. Vendor switches are already happening on price alone, as the reported Replit move from Anthropic to OpenAI showed. That meeting should be settled by your own evals, not by launch-week coverage.

What it means for builders

If both models land close together, you get a rare chance to compare two new frontier options on the same tasks, under the same prompts, in the same week. That window is short. Once one model is wired into your tooling, switching costs climb and the comparison quietly stops happening. Keep in mind too that the leaked capability demos, from Minecraft clones to kernel exploits, are anecdotes and not benchmarks.

What to do now

1. Freeze a test set drawn from real production traffic, such as tickets, documents and code tasks, with known-good outputs. Run your current model against it today so you have a baseline.
2. Decide your scoring before launch: task success, tool-call correctness, cost per completed task, latency, and how each model fails on long runs.
3. Abstract the model call behind a gateway or config flag so a new model is a one-line change. That is the difference between a two-day eval and a two-week one.
4. Watch a dated tracker such as Opper’s model release index, which refreshes every 15 minutes, so you know when a model is actually callable.
5. Check pricing on day one. GPT-6 Astra set the reference at $10/$50 per million tokens, and any change tells you whether OpenAI is competing on cost, capability or both.

If you are still on the previous generation, our notes on the GPT-6 rollout cover what changed last time. Above all, be clear about what you are comparing. A model that wins a demo but drifts on your 40-step workflow is the wrong pick, however good the launch week looks.

Frequently Asked Questions

Has OpenAI or Anthropic confirmed GPT-6.2 Astra, Fable 5.5 or Mythos 5.5?

No. The evidence is Codex usage tags, an API response field, deleted catalogue entries and hidden site code. Crypto Briefing reports that neither company has confirmed the models or published benchmarks, pricing or release dates.

Who can use GPT-6.2 Astra right now?

According to QbitAI, as cited by ChainCatcher, it is in limited grey-scale testing, usually for Pro subscribers spending up to $500 a month. Some users saw about 20% of requests routed to it, and others saw all of them.

Is the Fable 5.5 Azure listing real?

It is unclear. A bot-added LiteLLM entry for azure_ai/claude-fable-5-5 appeared on 13 October and was deleted the next day with the note that Azure does not list the model. Its parameters also match Fable 5.1 exactly.