Claude Fable 5.2 vs Opus 5.5: Pricing, Benchmarks, and What a Coding Task Really Costs

Anthropic released Claude Fable 5.2 on November 9, 2026. The API model ID is claude-fable-5-2. It costs $8 per million input tokens and $40 per million output, which is 20% less than Fable 5.1 and still exactly twice the list price of Opus 5.5.

Anthropic’s headline claim is about cost per task: on FrontierCode 2.0, Fable 5.2 at high effort matches the best Opus 5.5 score for about half the cost. We wanted to see what that looks like inside a coding agent, so we ran both models on the same task at two effort levels, three runs each.

Fable 5.2 vs Opus 5.5: which should you use?

Start with Opus 5.5. Anthropic says the same thing in its own docs: Opus is the place to begin for most workloads, and Fable 5.2 is the step up when your evals on Opus at higher effort still fall short.

Move to Fable 5.2 for long, open-ended technical work where the extra capability pays for itself, and set the effort level on purpose. Medium is a reasonable place to try first in an agent. High is where the cost climbs.

Here is what we measured, averaged over three runs of each setup:

Three things stand out:

• At the same level, Fable 5.2 cost 1.9x as much at medium and 4.5x as much at high. At medium that is almost exactly the list price gap, even though Fable 5.2 needed fewer steps.
• Changing only the model ID raised the bill the most. Opus 5.5 defaults to medium on the API and Fable 5.2 defaults to high. In our runs that one swap made the same task 7.9x more expensive and 6.3x slower.
• Fable 5.2 at medium cost about the same as Opus 5.5 at high ($1.87 vs $1.70) and finished in less than half the time.

What we measured

We ran all four settings through Atomic Agent, our open-source agent, on a pre-release build of its next version, in cloud mode through OpenRouter with the provider pinned to Anthropic. Each run got a fresh, isolated copy of the same Python project and the same one-line prompt. The time limit was 20 minutes for the Opus 5.5 medium runs and 60 minutes for the rest, and none of them hit it.

The task is a warehouse stock library of about 1,400 lines with 13 planted bugs and one missing feature, stock holds. The agent has to read a written spec, fix the bugs, build the feature, and get the test suite passing. We then score the result against 59 hidden tests the agent never sees. The untouched code passes 20 of them, and two deliberately naive fixes we wrote pass 30 and 44, so the suite does separate weak work from complete work.

Cost is the total OpenRouter reported for each request. The runs were fairly consistent: Opus 5.5 at medium ranged from $0.96 to $1.00, Fable 5.2 at medium from $1.78 to $2.03, and Fable 5.2 at high from $7.20 to $8.12. Opus 5.5 at high was the noisiest, from $1.54 to $1.98 and 13 to 25 steps.

For the two default settings we sent no effort value, so going by Anthropic’s docs each model ran at its API default. For the other two we set reasoning.effort in the OpenRouter request.

Limits of this test. It is one task, three runs per setting. Every setting scored 59 out of 59, so this test says nothing about which model is smarter. It shows how much thinking and how many steps each setting spends to reach the same result. We did not test xhigh or max on either model, or Fable 5.2’s fast mode. Timing is less controlled than cost: the Opus 5.5 medium runs went one after another earlier the same day, while the other nine ran partly or fully in parallel. On a task hard enough that Opus 5.5 fails, the extra spend on Fable 5.2 could easily be worth it, and this test cannot show that.

Pricing

List prices from Anthropic’s pricing page on November 9, 2026. OpenRouter lists the same base rates.

A cache read costs the same $0.20 on both models, but it was a small share of our bill. Cache writes, fresh input, output, and thinking all double, which is why Fable 5.2 at medium landed at 1.9x Opus 5.5 even with fewer steps.

Benchmarks from Anthropic

Anthropic’s announcement figures, at each model’s highest available effort unless noted. The Fable FrontierCode scores are at xhigh. These are the lab’s own numbers.

Fable 5.2 leads Opus 5.5 on every row here. It does not win everywhere: Gemini 4 Carbon takes AutomationBench at 53.9%, and on the Elo version of AA-Briefcase, Opus 5.5 stays slightly ahead (1800 vs 1785).

The cost claims are the part worth reading closely. On FrontierCode 2.0, Anthropic reports Fable 5.2 at high effort scoring about the same as Opus 5.5 at its best (45.3% vs 45.5%) for roughly half the cost per task. That is not a comparison at equal settings. It puts Fable 5.2 at high against Opus 5.5 at its most expensive level, on a benchmark where both score under 50%. Our test went the other way at equal settings on a task both models finish. The two results do not contradict each other, and they answer different questions.

Effort levels, and why the default matters

Fable 5.2 has five effort levels: low, medium, high, xhigh, and max. Thinking is always on and cannot be turned off, so effort is your main control.

Defaults differ by where you use it. On the Claude API, Fable 5.2 defaults to high. In Claude apps and Claude Code, Anthropic made medium the default. Opus 5.5 defaults to medium on the API.

So a third-party agent or your own code that sends no effort value gets high on Fable 5.2. That is what happened in our default runs: roughly 80 to 90 seconds per step on average and a $7.70 task. Going from high to medium cut the cost by 76% and the time by 84%, with the same 59 tests passing.

Anthropic’s guidance is to start at high, step down to medium or low where your evals show quality holds, and measure before moving to xhigh. On our task, quality held at medium. For routine agent work, test medium early. One thing to watch at lower effort: Anthropic notes that on long tasks the model sometimes ends a turn early with a list of what remains, and this happens more at low and medium. We did not see it in our runs.

What changes if you switch

Anthropic lists four changes from Fable 5.1 that can break code:

• History edits now fail on every account. If your code changes an earlier message, the system prompt, or the tool list in a conversation with thinking blocks, Fable 5.2 returns a 400 error. Fable 5.1 enforced this only on accounts created on or after August 31, 2026. A beta drop_block setting lets you continue while you fix the edit. Claude Code and Anthropic’s own agent tools handle this for you, but custom agents that trim or rewrite history need testing.
• Thinking blocks do not carry to other models. Only Fable 5.2 and Mythos 5.2 can read them, so a router that moves conversations to Opus 5.5 loses that reasoning.
• Computer use goes through computer_toolset_20260801. The older computer_20251124 tool returns an error everywhere except Amazon Bedrock.
• The advisor tool needs a Fable 5.2 or Mythos 5.2 advisor when Fable 5.2 is the executor.

Coming from older models, two rules also apply, though neither is new since Fable 5.1 or Opus 5.5. Forced tool use (tool_choice set to any or a named tool) returns a 400, so use auto with strict: true. Thinking is always on, so leave out manual thinking budgets, along with temperature, top_p, and top_k.

Fable 5.2 also requires 30-day retention and is not available under zero data retention unless Anthropic expressly authorizes it.

Safety classifiers also still apply. Fable 5.2 blocks high-risk cyber work and most research biology tasks. A declined API request comes back with stop_reason: "refusal", and you can opt in to a fallback on Opus 4.8 or Opus 5. Anthropic says blocks fell about 55% compared with Fable 5.1 in early testing on claude.ai, Claude Code, and Cowork, so security and biology teams should retest. Mythos 5.2, the same model for verified organizations, remains limited access.

How to try Fable 5.2

To repeat our comparison on your own tasks, download Atomic Agent for Mac, Windows or Linux, or install the terminal version. The public release is a little older than the pre-release build we tested, so your numbers may differ.

curl -fsSL https://atomicagent.io/install | sh
atomic-agent

On Windows, run irm https://atomicagent.io/install.ps1 | iex in PowerShell. The source is on GitHub.

Connect OpenRouter with your own key and choose anthropic/claude-fable-5.2 from the live OpenRouter model list; see our OpenRouter guide. Atomic Agent sends no effort value by default, so you get high. To run at medium, add one line to your openrouter entry under llm.providers in ~/.atomic-agent/config.json, and leave the rest of the entry as it is:

"extraBody": { "reasoning": { "effort": "medium" } }

That setting applies to every model on that provider entry, so change it when you switch models. More on providers in our models guide. If you are comparing coding agents rather than models, see our list of open-source Claude Code alternatives.

If you only want Fable 5.2 itself, it is on the Claude API, Claude Code, Amazon Bedrock, Google Cloud, and Microsoft Foundry.

The takeaway

Fable 5.2 is the stronger model on Anthropic’s benchmarks and a real price cut from Fable 5.1. It is still twice the list price of Opus 5.5, and on our one task Opus 5.5 did the job for less at every level we tried. Keep Opus 5.5 as the default. Bring in Fable 5.2 when Opus 5.5 falls short, and set the effort level yourself, because the default on the API took our task from under a dollar to nearly eight.

For earlier context, see our notes on Fable 5.1 and how Sonnet 5.5 compares with Opus 5.5.