Claude Fable 5.2 vs Opus 5.5 for agent data work

Claude Fable 5.2 costs 2x Opus 5.5 per token. Which one should plan your data agent, which should run the repeated calls, and what 100 leads cost.

Anthropic released Claude Fable 5.2 on 2026-11-09. This page is for agents that pull data: lead lists, work emails, Google results, company records. It compares Fable 5.2 with Claude Opus 5.5 on Anthropic’s published facts only, gives a routing rule for data agents, and prices a 100-lead job from looot prices read on 2026-11-09.

Short answer: start with Opus 5.5, as Anthropic advises. Give Fable 5.2 the planning and the hard judgment calls, and send the repeated lookups to a cheaper model or a script. The data calls cost the same whichever model makes them.

To connect looot to Fable 5.2 first, follow Claude Fable 5.2 MCP setup. It also has every receipt from a real 10-lead Fable 5.2 run, which this page reuses.

Is Claude Opus 5.5 better than Fable 5.2?

Not on Anthropic’s agent benchmarks. Fable 5.2 scores higher on all but one row in the table below. Anthropic still tells you to start with Opus 5.5 for most workloads. Its models overview says to “use Claude Fable 5.2 for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short.” The launch post calls Opus 5.5 its strongest daily driver.

Most of a data agent’s work is not demanding reasoning. It is “find this person, find their email, check it, enrich the company,” four calls repeated per row. The hard part is the plan and the matching: which companies fit, which person is the right one, what to do when a step misses.

How much more does Fable 5.2 cost than Opus 5.5?

Twice as much per token, with one exception: cache reads cost the same, $0.20 per million tokens on both.

Two notes on speed. Anthropic says Fable 5.2 writes output up to 30% faster than Fable 5.1, which is a claim about the previous Fable, not about Opus. Fast mode makes Fable 5.2 generate up to 2.5x faster at double the price.

The default effort differs. If you switch a running agent from Opus 5.5 to Fable 5.2 without setting effort, it moves from medium to high as well, and that alone can change how many tokens it spends. Set effort explicitly when you compare the two.

Sources, read 2026-11-09: Anthropic announcement, models overview, Fable 5.2 model page, pricing.

What do the benchmarks say about agent work?

These are Anthropic’s numbers from the launch post, with safeguards on. Each score is the model’s highest available effort unless a cell says otherwise. We did not run any of them.

Two cost points from the same post matter more than the table for a data agent. On WANDR, Anthropic says Fable 5.2 at medium effort matches Opus 5.5’s top score for about 40% of the cost per task. On FrontierCode 2.0, Fable 5.2 at high effort scores about level with Opus 5.5’s best (45.3% against 45.5%) for roughly half the cost per task. So the per-token price is 2x, yet Anthropic reports a lower cost per task at a lower effort on those two tests. Check that on your own job before you rely on it.

One independent number exists so far. Artificial Analysis lists Claude Fable 5.2 (Max, Default Fallback) at 65 on its Intelligence Index, first of 74 models it ranks, with an output speed of 103 tokens per second (Artificial Analysis, Claude Fable 5.2, read 2026-11-10). That is one setting of one model, so it says nothing about Opus 5.5 on the same scale; compare them on that site directly before you quote a gap.

Which model should plan and which should run the lookups?

Split the job by what each step needs.

Anthropic describes this setup for Fable 5.2 as an orchestrator that hands simpler tasks to smaller models. It also works the other way round: Haiku, Sonnet or Opus can consult Fable 5.2 as an advisor on the hard rows. Anthropic’s overview describes Haiku 5.5 as built “for high-volume, latency-sensitive tasks such as classification, extraction, and routing.” We have not measured Haiku on this job, so run 10 rows before you send 100.

Why not one long session? Every turn re-reads the whole conversation. In the 10-lead Fable 5.2 run, 1,653,760 tokens were cache reads and 283,977 were cache writes, against 33,175 output tokens. A longer job grows that context with every tool result. Short per-row workers keep each request small.

What does a 100-lead enrichment job cost on each model?

The job: 100 company domains in, and for each one a founder or CEO, a work email, a verify check and the company record. That is four looot runs per domain, 400 runs.

The data cost is the same on every model

looot charges per run, at the price search and inspect show before the run. The model that sends the call does not change the price. These are the four endpoints that answered in the Fable 5.2 demo, with prices read from inspect on 2026-11-09:

If the first provider answers every step, 100 domains cost $1.01 of looot credit.

Real runs cost more, and the demo shows why. Its 69 runs cost $0.442164 for 10 finished leads, $0.044 per lead, which is about $4.42 for 100. Two things raised it. AI Ark ran first on all 18 people searches, found no one and billed $0.016173 each, $0.29 in total, before LeadMagic answered. And 8 of the 18 domains gave no founder or CEO, so their searches bought nothing. With fallback, a miss moves to the next provider for the same job inside the cap you set, and route.attempts on each run shows who billed what.

The model cost is an estimate

We did not run this 100-lead job on either model. The numbers below are list-price arithmetic on token counts we assumed, so read them as an order of magnitude, not a measurement:

• Planner: 30,000 input and 8,000 output tokens in total, for the plan and the final review.
• Worker, per domain: 5 turns of 8,000 input tokens each (instructions, tool definitions, results so far), so 40,000 input tokens, plus 1,500 output tokens. For 100 domains, 4,000,000 input and 150,000 output.
• No prompt caching and no batch discount, which would both lower these figures.

Add the data cost, $1.01 to about $4.42, to any row. With Haiku 5.5 doing the per-row work, the looot data costs more than the model does.

For one measured point: the 10-lead Fable 5.2 demo ran as a single session and its tokens came to $6.20 at list price. The same token counts at Opus 5.5 rates would be $3.27. Opus 5.5 would not spend exactly the same tokens, so that is a repricing, not a test of Opus.

Does the model change what the data calls cost?

No. A looot run is priced by the endpoint and its input, and the charge lands in the run’s receipt. Fable 5.2, Opus 5.5, Haiku 5.5 and a shell script calling the REST API all pay the same $0.005 for a kittai-find-email call. What the model changes is the token bill, and how many calls it decides to make.

What looot offers, read from the live catalog and balance on 2026-11-09:

• One key, one MCP connection. The catalog listed 4,144 endpoints from 310 providers, 2,479 of them runnable without your own account. Server URL: https://api.looot.ai/mcp.
• Prepaid credit, per call. Top-ups start at $5. See Before your first run.
• A price before you spend. search and inspect are free and show the most a run can cost.
• Fallback on a miss. A run can try the next provider for the same job, up to the maxCostUsd you set.
• You pay for attempts that ran. An input the provider never received costs nothing. A provider that ran and found nothing can still bill, as AI Ark did above.

When is Fable 5.2 worth twice the price?

Move a step to Fable 5.2 when Opus 5.5 fails it on your own checks, not by default. For a data agent that usually means:

• The plan is the hard part. A vague target (“open-source business software vendors, founder-led”) where most search results are directories, listicles and forums to throw out.
• Matching across sources. Deciding that the person from one provider and the company from another are the same, or that a reseller is not a vendor.
• Long, many-step jobs. Anthropic positions Fable 5.2 for long-horizon agentic work, and its largest gaps over Opus 5.5 are on WANDR and Remote Labor Index.

Keep Opus 5.5, or a smaller model, on fixed steps with clean inputs. Set maxCostUsd on every run so a model that loops cannot spend more data credit than you planned.

FAQ

Is Claude Fable 5.2 better than Opus 5.5 at web search for agents?

On Anthropic’s WANDR agentic search benchmark, Fable 5.2 scores 82.1% against 72.3% for Opus 5.5. Anthropic also says Fable 5.2 at medium effort matches Opus 5.5’s top WANDR score for about 40% of the cost per task. These are Anthropic’s numbers; we have not measured search quality on looot.

Can Fable 5.2 and Opus 5.5 share one looot key?

Yes. The key and the credit belong to your looot organization, not to a model. A planner on Fable 5.2 and workers on Haiku 5.5 can all call the same MCP server and spend the same balance.

Does Fable 5.2 spend more tokens than Opus 5.5 at default settings?

Its default effort is high and Opus 5.5’s is medium, so a switch without setting effort also raises effort. Set it explicitly on both before you compare costs.

Should I use Fable 5.2 for every lead lookup?

No. Each lookup is a fixed call with a fixed price on looot. Paying 2x per token for the model that sends it buys nothing on that step. Use Fable 5.2 where judgment is the bottleneck.

Is Opus 5.5 still Anthropic’s recommended default?

Yes. Anthropic’s models overview says to start with Claude Opus 5.5 for most workloads, and to use Fable 5.2 for demanding reasoning and long-horizon agentic work.