In October 2025, the best AI agent could finish 2.5% of real freelance projects to a client's standard. By June 2026, the official Remote Labor Index record was 16.1%, set by Claude Fable 5. On November 9, Anthropic reported that Fable 5.2 finishes 41.3% in its own tests, a result not yet confirmed on the independent leaderboard.

Fable 5.2 is Anthropic's most capable generally available model, now 20% cheaper per token and up to 30% faster than the version it replaces.

It matters because the work it does best (ambiguous engineering, numbers under scrutiny, long research runs) is exactly the work that stalls a small team.

You'll learn where it beats Anthropic's everyday model, what it really costs, and which jobs to hand it first.

Below is the full breakdown, written for founders rather than benchmark watchers.

Inside you'll find:

Most launch coverage leads with black holes and Fermat's Last Theorem. For a founder, the headline is the price list.

Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1. Outputs arrive up to 30% faster. An optional fast mode generates up to 2.5x faster, at double the price ($8 / $40 becomes $16 / $80).

The rest of the release notes, translated:

Companies were already paying for this tier before the price cut. Fable accounted for 11% of spending on Anthropic's models in August, according to Ramp's AI Index as reported by The Next Web.

→ Cheaper per token is not the same as cheap.

Fable still costs more per token than Claude Opus 5.5, which Anthropic itself recommends as the everyday default. The real saving shows up per finished task. On Anthropic's FrontierCode 2.0 coding test, Fable 5.2 at High effort scored 45.3%, level with Opus 5.5's best of 45.5%, for roughly half the cost per task.

The most telling detail: the price cut matters less than the efficiency. Fable 5.2 reaches the same answer with fewer steps, so a hard job can end up cheaper on the "expensive" model.

Most early-stage teams don't lack ideas. They lack senior engineering hours.

That is where Fable 5.2 is strongest. Anthropic says Claude now writes over 90% of the changes merged into its core codebase, with human guidance. A team of agents running an earlier version of Fable 5.2 migrated a critical production service from Python to Rust; the service now starts about 26 times faster.

Early testers report the same pattern on their own code:

On the benchmarks, Fable 5.2 scored 49.1% on FrontierCode 2.0 at its highest effort setting, the top score of any released model Anthropic tested, and 69.9% on FrontierSWE v2 versus 62.3% for Opus 5.5.

→ The win isn't writing code faster. It's finishing the projects that never make the sprint.

For a 4-to-15-person team, that usually means:

Investors rarely reject a round over slide design. They reject it when the numbers don't reconcile.

Fable 5.2's biggest gains are in judgment on messy, data-heavy work, which is exactly what a data room is. Fundamental Research Labs asked it to pull a company's financials from a 10-K into a forecasting model. The template was ambiguous, yet the model inferred the author's intent, caught that regional numbers didn't sum to the reported total and got every cell right. It scored 91% on their evaluation versus 86% for the next best model.

Other testers saw the same discipline:

It is also more willing to admit uncertainty. Cognition reported that it says it doesn't know instead of inventing an answer. Anthropic's own audit of the underlying model found fewer hallucinated inputs and fewer false claims of finished work than Opus 5.5.

→ Treat it as the skeptical associate who reads your data room before the investor's associate does.

Concrete jobs before your next raise:

In most startups, the real constraint isn't headcount. It's the founder's calendar.

Anthropic calls Fable 5.2 its strongest lead agent. It can run ambiguous work for days, coordinate hundreds or even more than a thousand subagents, and play two useful roles: Orchestrator, delegating simple tasks to cheaper models, or Advisor, where cheaper models consult it on hard problems.

An Uber engineer had it run a 10 to 12 hour security audit. It sent out about 170 reviewer agents and returned with a write-up, fix pull requests and a policy proposal. Datadog compared it to an incident commander that dispatches faster models for quick signals. At Jump Trading, it played every role in a research system at once: planner, coordinator, coder and job submitter.

Closer to a founder's reality is ChatPRD. Claire Vo runs the bootstrapped startup, used by more than 100,000 product managers, as its only human, with nine AI agents covering jobs like sales inquiries, her calendar and podcast analytics. Testing Fable 5.2 for Lenny's Newsletter, she used it to re-architect ChatPRD's product-intelligence pipeline.

The clearest stress test was mathematical. A team of 1,000 Fable 5.2 agents fully formalized the proof of Fermat's Last Theorem in 30 hours, with minimal human help. The same setup running Opus 5.5 stalled at about a quarter of the work.

→ The lesson for founders isn't the math. It's the delegation pattern.

Put Fable 5.2 on planning, judgment and review; let cheaper models handle volume. Good first candidates:

Hardware founders face a brutal loop: you need a prototype to raise, and you need money to build the prototype.

Anthropic tested Fable 5.2 on exactly that. From a one-paragraph prompt, it designed three working devices with circuit-board maker Diode: a home voice-assistant hub, a USB speakerphone and a handheld virtual pet. It wrote the specs, picked parts from US distributors, routed four-layer circuit boards, modeled the enclosures and wrote the firmware.

Each board was ready for manufacturing in under a day, and all three devices worked without electrical modifications. Anthropic is clear about the limit: these are prototypes, not yet validated, certified or designed for mass production.

The same pattern holds in scientific software. Discovered Materials used Fable 5.2 to make a standard materials-science simulation about 36 times faster on the same GPU, cutting 13 benchmark cases from 3.3 hours to under four minutes.

→ For deep-tech teams, the first demo can shrink from a quarter-long project to days of work plus a manufacturing partner.

Humans still own validation, certification and the decision to build. What changes is how much proof you can show an investor before you spend your seed round.

No model comes without trade-offs, and a vendor's launch post won't list them for you. Before you build a workflow on Fable 5.2, weigh these six.

→ Fable 5.2 is a lever, not a hire. It multiplies a founder who checks its work; it doesn't replace one.

The most expensive mistake is using the most powerful model for everything.

Anthropic's own guidance is to start most coding and knowledge work on Opus 5.5, then move up to Fable only when the job demands it. Use this rule of thumb.

Escalate to Fable 5.2 when at least two are true:

Stay on Opus 5.5, or drop to Sonnet or Haiku 5.5, for: emails, investor updates, meeting notes, routine code and slide copy.

Check the cost before you scale a workflow:

Example: a step that reads 200,000 tokens and writes 20,000 costs $1.60 + $0.80 = $2.40.

Where to start today:

A prompt to run before your next investor meeting:

Your one-week test: pick the job above that maps to your biggest bottleneck. Run the same prompt on Opus 5.5 and on Fable 5.2, then compare quality, time and cost per finished task. Write the winner down as your team's rule.

The goal isn't to use the most powerful model. It's to know exactly which work deserves it, so you ship faster without burning runway on tokens.

For six more copy-paste prompts mapped to founder pain points, see our companion resource, Claude Fable 5.2 for Founders.

All performance figures are Anthropic's or its testers' own pre-release results. Independent benchmarks were not yet available at publication.