AI AutomationProductivity

Fable 5.2 costs 2x Opus 5.5. Anthropic's own page says start with Opus.

·5 min read

Anthropic released Claude Fable 5.2 and Mythos 5.2 on November 9. It was the top story on Hacker News within hours, and the benchmark table is impressive. But one sentence in the same launch post matters more if you are a solo builder paying for tokens: "we recommend that users start with Claude Opus 5.5, our strongest daily driver."

That is Anthropic recommending the cheaper model on the day it launches the new one. So before you switch everything, here is what the launch post actually says, and a test that takes an afternoon.

What launched, and who can use it

Fable 5.2 is available on the Claude apps, the API (claude-fable-5-2), Amazon Web Services, Google Cloud and Microsoft Azure. Mythos 5.2 is a different story. It is limited to vetted organizations through Anthropic's Cyber Verification Program and Life Sciences Verification Program, so most of us will not touch it.

Fable's safeguards also block most high-risk cybersecurity and research biology tasks. According to the post, blocked queries fall back to Opus 4.8 for cyber and Opus 5 for biology. If your work touches security tooling, expect some requests to be answered by an older model.

This week, eligible paid plans get extra Fable usage. The post says plans that include Fable get one to three Fable resets depending on how long you have subscribed, and Pro plans and standard Team seats can use it for up to 50% of weekly usage limits. That is a launch promotion, not the permanent limit.

Number 1: 49.1% vs 45.5%

On FrontierCode 2.0, Anthropic lists Fable 5.2 at 49.1% and Opus 5.5 at 45.5%. The table note says every score is each model's highest available effort. Fable 5.1 scored 32.4% on the same test, so the jump from 5.1 to 5.2 is large. The gap to Opus 5.5 is 3.6 points on one coding benchmark.

The rest of the table points the same way, with Fable 5.2 ahead of Opus 5.5 on most rows (FrontierSWE v2 69.9% vs 62.3%, Humanity's Last Exam with tools 74.2% vs 67.7%). It is not ahead everywhere. On AutomationBench the table lists Gemini at 53.9% against Fable's 51.4%. These are Anthropic's own numbers, run by Anthropic, so treat them as a starting point. Artificial Analysis, an outside tracker, lists Fable 5.2 at 65 on its Intelligence Index, first out of the 74 models it tracks.

Number 2: the bill

Anthropic's pricing page lists:

| Model | Input per 1M tokens | Output per 1M tokens | |---|---|---| | Fable 5.2 | $8 | $40 | | Opus 5.5 | $4 | $20 | | Fable 5.1 | $10 | $50 |

Fable 5.2 is 20% cheaper than 5.1, which is the good news. It is still exactly double Opus 5.5 on both input and output. Fast mode is $16 and $80 for Fable 5.2, and $8 and $40 for Opus 5.5.

Number 3: at High effort the gap closes

Here is the catch inside the launch post. In Anthropic's words, "Fable 5.2 at High effort scored about level with Opus 5.5 at Max effort, for roughly half the cost per task." The chart note gives the figures: 45.3% for Fable at High and 45.5% for Opus 5.5's best score.

Read that carefully. It says Fable 5.2 can match Opus 5.5's best result for less money than Opus 5.5 spends to get there. It does not say Fable beats Opus 5.5 at every effort level. Opus 5.5 at a lower effort setting, which is how most of us run it, is not part of that comparison. The post also says that at Medium effort, the default in Claude apps and Claude Code, Fable 5.2 finishes tasks with fewer tokens than Fable 5.1 did.

So per token Fable costs double. Per finished task it might cost less or more than Opus 5.5, depending on how hard your tasks are and how much effort you give each model. Nobody can tell you that from a benchmark table. Your own tasks can.

The other claims in the post

The post also says a team of 1,000 Fable 5.2 agents formalized a proof of Fermat's Last Theorem in 30 hours, and that Opus 5.5 with the same setup reached 26% in 18 hours. Anthropic adds its own caveats: one run per model, the agents had no mathematical references, and an earlier effort by an unreleased model took 11 days. It also says the two models helped with two open black hole physics problems, with results that rely on holographic duality and are linked as preprints. I would read those as research demos, not as a reason to change your coding setup.

The switch test (an afternoon, no code)

Copy this into a note and fill it in.

SWITCH TEST: Opus 5.5 vs Fable 5.2
1. FREEZE  Pick 5 real tasks from the last month of my own work.
           Save the exact prompts and the files they used.
2. MATCH   Run each task twice:
           A) Opus 5.5 at the effort I normally use
           B) Fable 5.2 at High
3. RECORD  For each run: finished? (yes / needed a fix / failed)
           tokens in, tokens out, minutes, my corrections.
4. PRICE   cost = (tokens in x input price) + (tokens out x output price)
           Opus 5.5: $4 in, $20 out per 1M.  Fable 5.2: $8 in, $40 out per 1M.
5. DECIDE  Switch only the task types where B finished and A did not,
           or where B cost less per finished task.

Two notes on it. First, count a run that needs a fix as unfinished, because fixes cost you time even if they cost no tokens. Second, five tasks is a small sample, so use it to find the task types where Fable clearly wins, not to crown a winner.

What to do this week

  • Do not move your whole workflow because of a launch table. Anthropic itself says to start with Opus 5.5.
  • Use the extra Fable usage this week to run the switch test, since the promotion makes the test cheap on a paid plan.
  • If you pay by the token through the API, run the test with a budget cap, because Fable at High is not free.
  • Keep Fable for the hard tasks where a retry costs you the most, and keep Opus 5.5 for everything routine.

Which of your tasks would you test first?

Sources: Anthropic launch post, Nov 9, Anthropic pricing, Artificial Analysis.