Guide · Agentic AI · as of 10 November 2026

Claude Fable 5.2 is out. We tested it against Opus 5.5 on our own code.

On 9 November Anthropic released Claude Fable 5.2, its most capable model for technical work, at a lower price than Fable 5.1. In the same announcement it says something a model launch rarely says: start with the cheaper model. Anthropic recommends Claude Opus 5.5 as the default for most coding and knowledge work, and Fable for the hardest problems. We build software with Claude every day, so the next morning we tested that advice on a real job from our own code. Three models wrote the same function 18 times, and we graded every run against the engine behind our public UAE gratuity calculator. Here is what changed in 5.2, what it costs, what our test showed, and how we now choose a model for each task.

What Anthropic released

A stronger model, at a lower price.

Fable 5.2 is on the API as claude-fable-5-2, in Claude Code and the Claude apps, and on AWS, Google Cloud and Microsoft Azure. It costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1. Anthropic says it writes output up to 30% faster, uses fewer tokens than 5.1 at the default Medium effort, and was blocked by its safety filters about 55% less often than 5.1. The benchmark gains are large. On FrontierSWE v2, a test of real software engineering tasks, Anthropic reports 69.9% for Fable 5.2 against 62.3% for Opus 5.5. Claude Mythos 5.2 is the same underlying model without the cyber and biology safeguards, available only to vetted organisations through Anthropic’s verification programs. Two details matter if you use Claude through a subscription. Medium is the default effort in the Claude apps and Claude Code. And on Pro plans and standard Team seats, Fable can use up to half of your weekly usage limit, with one to three Fable resets depending on how long you have subscribed.

What each model costs

per million tokens, input / output

Fable costs twice what Opus does.

  • Claude Fable 5.2: $8 / $40. Fast mode, up to 2.5 times quicker output: $16 / $80.
  • Claude Fable 5.1: $10 / $50. Same tier, now replaced at a 20% lower price.
  • Claude Opus 5.5: $4 / $20. Anthropic’s recommended default, “our strongest daily driver”.
  • Claude Sonnet 5.5: $2 / $10. A quarter of the price of Fable.
  • Claude Haiku 5.5: $0.10 / $0.50 for prompts under 100,000 tokens. For lookups, sorting and short replies.

The claim worth testing

Half the cost per task?

Price per token is not the price of a task. A stronger model can finish in fewer steps and spend less in total. Anthropic makes exactly that argument: at High effort, Fable 5.2 scores about the same as Opus 5.5’s best score on its FrontierCode benchmark (45.3% against 45.5%), “for roughly half the cost per task”. That is a claim about hard problems, where Opus needs its highest effort to get there. Most of the work a business gives an AI is not that hard. So we asked a narrower question: on an ordinary, well-defined job from our own code, which model gets it right, and what does each correct answer cost?

Our test

10 November 2026

One brief, graded by our own engine.

The job: write a function that calculates UAE end-of-service gratuity under the current Labour Law, Federal Decree-Law No. 33 of 2021. It is a real task with real traps: no gratuity under one full year, 21 days of basic wage a year for the first five years and 30 after, unpaid days left out of the service, and a cap of two years’ wage. We gave each model the same short brief, with no law text and no web access, so it had to know the rules itself. Each model ran in Claude Code three times at each setting, wrote the function, tested it and wrote down its legal assumptions. We then ran every function on the same 40 cases, from 364 days of service to 30 years, and compared it with the engine behind our free gratuity calculator, to the fil. Before trusting the grader we tested it: our own engine scored 40 of 40, and a deliberately wrong version scored 3.

The free UAE gratuity calculator

What we found

3 runs per setting, cost at API prices

All of them were right. The bills were not.

  • Fable 5.2, Medium effort: 3 of 3 exactly right. About $0.50 a run, a little over a minute each.
  • Fable 5.2, High effort: 3 of 3 exactly right. About $0.88 a run, about two minutes each, twice the writing of Medium for the same answer.
  • Opus 5.5, High effort: 3 of 3 exactly right. About $0.31 a run, under a minute and a half.
  • Opus 5.5, Medium effort: 2 of 3 matched our engine. The third used a daily wage of basic × 12 ÷ 365 instead of basic ÷ 30, which pays 1.4% less. It wrote that choice down. About $0.24 a run.
  • Opus 5.5, Xhigh effort, its setting just below Max: 2 of 3 matched; the third made the same 12 ÷ 365 choice. About $0.68 a run and over three minutes each. More thinking, a bigger bill, no better answer.
  • Sonnet 5.5, High effort: 3 of 3 exactly right. About $0.13 a run, under a minute. Less than half of what Opus 5.5 cost at High, for the same answer.

What it means

Pay for the model the task needs.

On this job, Fable 5.2 at Medium cost twice what Opus 5.5 did at Medium and almost four times what Sonnet 5.5 did at High, and was no more correct. Anthropic’s own comparison is Fable at High against Opus at its top effort. On our task that is about $0.88 against $0.68: Fable cost around 30% more, not half, though it did finish sooner, in about two minutes against three. That does not make Anthropic wrong. Its half-the-cost figure is for problems hard enough that Opus needs its top effort to solve them. Ours was solved by every model at an ordinary setting, so the extra power had nothing to do, and raising the effort only bought more thinking. The one disagreement is also worth noticing. Whether a day’s wage is a thirtieth of a month or a 365th of a year is a real question of interpretation, and calculators in the UAE differ on it. The law fixes a month at 30 days for its own purposes, which is why our engine divides by 30. A model can make either choice. What made it safe was that it wrote the choice down, so a person could check it. Read the assumptions an AI writes down, whatever model wrote them. Our own usage already leaned this way. Over the last week, weighted by cost, 73% of our Claude usage was Opus 5.5, 17% Sonnet 5.5 and 10% Fable 5.1. Two weeks ago it was worse: subagents were starting on the most expensive model because nobody had set one.

How we run several AI coding agents at once

How we choose a model for each task now

Start low. Climb when it fails.

  • Lookups, sorting, search and short replies: Haiku 5.5, or Sonnet 5.5 when the answer needs judgement.
  • Everyday coding, writing and review: Opus 5.5 at Medium or High. This is our default, and Anthropic’s.
  • Long, hard, many-step work, like a migration across a large codebase or a bug nobody can find: try Fable 5.2 at High, and compare its cost per finished task with Opus at its top effort.
  • In Claude Code, set it per job: claude --model claude-fable-5-2 --effort high for one hard session, and a model on every subagent you define, so none starts on the largest one by default.
  • Ask every model to write down its assumptions, and read them. That is where our one disagreement showed up.
  • Measure cost per finished task, not price per token. Claude Code reports the cost of every non-interactive run, so a short test like ours costs a few dollars.

Why this matters to us

The model is a choice we make per job.

We are an AI lab in Dubai. The software we build for UAE businesses, from audit tools to our free calculators, is written by AI agents and checked against engines like the one in this test. A new model is a reason to test, not a reason to switch everything. When a client asks which model their app should run on, the answer comes from a test on their own task, with the cost of each correct answer written next to it.

Custom software development in Dubai

Questions

Questions about Claude Fable 5.2.

  • When was Claude Fable 5.2 released?

    Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2 on 9 November 2026. Fable 5.2 is available the same day on the Claude API as claude-fable-5-2, in Claude Code and the Claude apps, and on AWS, Google Cloud and Microsoft Azure.

  • How much does Claude Fable 5.2 cost?

    $8 per million input tokens and $40 per million output tokens, with cache hits at $0.20 per million. That is 20% less than Fable 5.1 ($10 and $50) and twice the price of Claude Opus 5.5 ($4 and $20). Fast mode costs $16 and $80.

  • Claude Opus 5.5 vs Fable 5.2: which should I use?

    Start with Opus 5.5, as Anthropic itself recommends. Move a task to Fable 5.2 when Opus fails it or needs its highest effort. In our test on an ordinary coding job, all three models got it right. Fable at Medium cost twice what Opus did at Medium, and Fable at High cost about 30% more than Opus at its top effort.

  • How do I use Fable 5.2 in Claude Code?

    Choose it with /model inside a session, or start one with claude --model claude-fable-5-2. Medium is the default effort; raise it with --effort high or /effort for a hard task. On Pro plans and standard Team seats, Fable can use up to half of your weekly usage limit.

  • What is the difference between Claude Fable 5.2 and Claude Mythos 5.2?

    Anthropic says they are the same underlying model. Fable 5.2 has safeguards for cyber security and biology and is open to everyone. Mythos 5.2 runs without those safeguards and is available only to vetted organisations through Anthropic’s verification programs.

  • Is Fable 5.2 better at coding than Opus 5.5?

    On Anthropic’s benchmarks, yes: 69.9% against 62.3% on FrontierSWE v2. On our own task both were correct, so the difference showed up only in the cost. The harder and longer the task, the more the stronger model is likely to earn its price.

Sources

read 10 November 2026

Check it at the source.

  1. Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2 (9 November 2026, benchmarks, effort and access)
  2. Claude API pricing (per-token prices for every model, read 10 November 2026)
  3. Ministry of Human Resources and Emiratisation (Federal Decree-Law No. 33 of 2021, Article 51 (gratuity) and Article 67 (a month is 30 days))
  4. Claude Code docs: model configuration (choosing a model and effort per session)

A question?

Choosing a model for your app?

Tell us what your AI has to do. We will test it on your own task and show you which model gets it right, and what each correct answer costs.

Talk to usWhatsApp +971 50 728 5494hello@brevdigital.com