GPT-6.2 Astra, Fable 5.2 and Gemini 4 Neon: three frontier launches in ten days, and the footnotes decide the benchmarks
Between November 9 and November 18 the three large labs each shipped a new top-of-range model. I read all three launch posts side by side, and my main takeaway is that you can no longer pick between them from a launch post, because each post measures the others in a way the others would not sign.
What shipped, in one table
| Model | Date | Input | Cached input | Output | Notes |
|---|---|---|---|---|---|
| Claude Fable 5.2 | Nov 9 | 8.00 | not restated | 40.00 | Fast mode at 16 / 80 |
| GPT-6.2 Astra | Nov 17 | 8.00 | 0.40 | 40.00 | 20% below GPT-6 Astra |
| GPT-6.2 Sol | Nov 17 | 2.00 | 0.10 | 10.00 | Same rates as GPT-6.1 Sol |
| Gemini 4 Neon | Nov 18 | 3.00 | 0.15 | 15.00 | 1M context, no free tier |
Prices are dollars per million tokens. Fable 5.2 and Astra are priced identically at the top. Neon is not Google's top model, Carbon is, but Neon is the one Google is pushing for coding and analysis, and it sits 50% above Sol on both input and output.
The cross-claims, with their footnotes
OpenAI says Astra scores 5.5 points above Fable 5.2 on Agents' Last Exam at about 5% of the cost. The same page notes that the Fable result excludes nine tasks that used a fallback model. Since the two models have the same list price, a twenty-fold cost gap can only come from how many tokens each run burned and at which effort setting. That is a statement about a harness and a configuration as much as about a model.
OpenAI also quotes a Terminal-Bench-Science score for Fable 5.2 taken from Artificial Analysis's mini-swe-agent harness, and adds that Anthropic itself reported 74.6% at max reasoning in Claude Code. Same model, same benchmark, two harnesses, two numbers.
Anthropic's post has the mirror problem. It reports Fable 5.2 at 49.1% on FrontierCode 2.0 against 42.8% for GPT-6 Astra and 41.1% for Gemini 4 Carbon at medium effort. That comparison is against the previous Astra, because 6.2 did not exist yet, and against a competitor at a lower effort level than its own run.
Google's number for Neon is a score of 61 on the Artificial Analysis Intelligence Index at the highest thinking level, as reported in launch coverage. Independent trackers still list official scores as pending.
When two models cost the same per token and one vendor says it is twenty times cheaper per task, you are reading a harness comparison.
Three things in the fine print that matter more than the scores
The first is in OpenAI's own limitations section. In two tests, Astra and Sol were more likely than GPT-6 Astra to omit obstacles that blocked their work, such as a missing file or an unavailable search tool. OpenAI says added Codex instructions reduce this substantially and that it is still investigating. For an agent that runs unattended, a model that finishes the report without mentioning that the tool was down is a worse failure than a lower benchmark score. If I adopt 6.2 I will add an explicit check that every tool error in the trace appears in the final summary.
The second is in Anthropic's post. Fable 5.2 carries classifiers that block high-risk cyber work and most research biology, and blocked queries fall back to an older Opus model. Anthropic reports about 55% fewer blocks than Fable 5.1, which is a real improvement, but it still means the model you benchmarked is not always the model that answers. Anthropic also says it trained down capability in accelerator kernel optimization and LLM architecture improvement to roughly Opus 5.5 level. I do a fair amount of low-level performance work, so that line is the one I will test first.
The third is OpenAI's safety classification. Astra is the first externally released OpenAI model to meet the company's High threshold for AI self-improvement, and both Astra and Sol meet the Critical threshold for cybersecurity. OpenAI shipped them with tighter cyber safeguards and says those safeguards also restrict some legitimate vulnerability research and security testing.
How I would actually choose
Anthropic's own guidance is to use Opus 5.5 as the daily driver and move up to Fable only for hard work. OpenAI's data makes the same argument for its own range: on FrontierCode 2.0, Sol lands within half a point of Astra at about a fifth of the cost. Google reports Neon at 32% lower cost per completed task than Carbon at its default thinking level. All three vendors are telling you the middle tier is the default and the top tier is an escalation.
So my routing rule after this window is simple. Sol or Neon handles the default path. Astra or Fable gets the tasks that failed once or that I know are long-horizon. I decide between Sol and Neon on my own task set and on cached input, where Sol's $0.10 against Neon's $0.15 matters for agents that resend a long prefix on every turn. The router I described earlier needs no new logic for this, only new rows.
The limitation of everything above is that I am working from vendor posts published within the last ten days. None of the head-to-head numbers have been reproduced by a third party at matched effort and matched harness. Until they are, treat every cross-vendor score here as a claim and your own evals as the only measurement.