Claude Fable 5.2 Freelance Ability Tested

Claude Fable 5.2 finished work that a paying client would accept in 41.3 percent of the projects in the Remote Labor Index. That figure is the reason this model matters this week. A year ago the best agent on the same index finished 2.5 percent. The jump is large, the test is built from real paid work, and most of the jobs still failed. Facts checked November 12, 2026.

Anthropic released Fable 5.2 on November 9, alongside Claude Mythos 5.2. The two names describe the same underlying model with different safeguards. Fable is the public version. Mythos is limited to organisations cleared through verification programs for cyber and life sciences work. This review covers Fable, the version a reader can actually buy, and it judges the model on the freelance result first because that result claims something concrete about paid work.

What It Is

Fable 5.2 is Anthropic most capable public model for demanding technical work, in the company own framing. It answers questions, writes text and code, works with files, and runs as an agent that can use tools over a long task. Anthropic says it can act as an advisor that other Claude models consult, or as an orchestrator that hands smaller tasks to smaller models.

The freelance claim sits inside a wider launch table. Anthropic reports gains over Fable 5.1 on every line it published, from coding to chart reading to robotics simulation. Two launch demos got most of the attention. In one, the model directed more than 1,000 subagents for eight days to build a prototype operating system from an empty folder. In another, 1,000 agents formalised a famous maths proof in a computer checked language in 30 hours. Both demos are single runs reported by Anthropic, and nobody outside the company has reproduced either one. Treat them as claims about what the company says happened, not as measured results you can plan around.

The Remote Labor Index is different in kind. It was built by the Center for AI Safety and Scale Labs, not by Anthropic. Its projects came from outside the lab, its graders are human, and its score was published on a public leaderboard. That is why this review gives the index more weight than the demos, while still reading the index closely for what it leaves out.

Features and Specifications

The core facts are unchanged from Fable 5.1 in shape and changed in price and speed. The context window, meaning the amount of text and material the model can hold in view at once, stays at one million tokens. Maximum output stays at 128,000 tokens. Anthropic says output arrives up to 30 percent faster than Fable 5.1, and an optional fast mode raises generation speed by up to 2.5 times at double the token price.

Pricing fell by exactly 20 percent on the main lines, from $10 and $50 on Fable 5.1 to $8 and $40 here. That cut sounds generous until you place it beside the rest of the Claude range. Opus 5.5 costs $4 in and $20 out, half the Fable price. GPT 6 Astra lists at $10 in and $50 out on the comparison reporting for this launch week, so Fable now sits below Astra per token and well above the model Anthropic calls its daily driver.

Effort is the second price dial, and it matters more than the rate card. A model effort setting controls how much work the model does before it answers. On FrontierCode 2.0, Anthropic system card figures reported this week put Fable 5.2 at 34.2 percent on low effort for about $2.10 per task, 45.3 percent on high effort for about $6.72 per task, and 49.1 percent on the extra high setting for about $14.19 per task. At the max setting the score fell to 45.7 percent while the cost rose to about $44.70 per task. More effort bought more score up to a point, then it bought a larger bill for a worse result.

Independent cost checks point the same way. Artificial Analysis, which runs its own model tests, ranked Fable 5.2 first on its intelligence index with a score of 65, ahead of Gemini 4 Carbon at 60. The same testing found cost per task lower than Fable 5.1 at medium effort and higher at high effort, because the new model used more tokens at the higher setting even though each token costs less. A 20 percent token cut is not a promise that your bill falls by 20 percent.

Real World Experience

This is a research review. Dirty Static did not hire Fable 5.2 for freelance projects and did not rerun the index. What follows draws on the Scale Labs leaderboard and index documentation, the paper that introduced the index, Anthropic launch material, and published coverage that names its sources. Vendor figures are labelled as vendor figures where they appear.

Start with how the index works, because the method is the strongest part of the claim. The builders paid 358 freelancers with verified Upwork accounts for samples of work they had already completed for clients. Of 550 projects collected, 240 passed review. The set spans 23 kinds of work, including game development, product design, architecture, data analysis and video animation. Each project carries the client brief, the input files and the finished work the human delivered.

The jobs are not toy tasks. At the median, a project took a human 11.5 hours and was worth $200. Averages run higher, at 28.9 hours and $632.60, because the set includes jobs worth more than $10,000 and jobs that took more than 100 hours. Together the projects total $143,991 in reported value. Leaderboard scores run on 230 projects kept private, which limits the chance that a model has seen the exact jobs in training.

Grading is human and blunt. A trained evaluator opens the agent files next to the freelancer files and scores the agent work 1, 2 or 3. A score of 1 fails. A score of 2 meets the standard. A score of 3 exceeds it. The headline rate, called the automation rate, is the share of projects scored 2 or 3. Three evaluators review each deliverable, the majority vote counts, and Scale Labs reports 94.4 percent agreement among evaluators. The index team has tried to automate this grading and reports that AI graders scored past models far too high, in part because they could not open professional tools properly and missed work that looked polished but was broken inside the file.

Against that method, the leaderboard is stark. Fable 5.2 sits at 41.30 percent. Opus 5.5 is second at 21.25 percent. GPT 6 Astra is third at 20.83 percent. Fable 5.1 records 17.92 percent and Gemini Argon 4 records 17.08 percent. Anthropic launch table prints the same Fable result in rounded form, 41.3 percent at high effort, which matches the leaderboard entry added on November 9.

Scale Labs said the largest gains for Fable 5.2 came in audio and music work, architecture and interior design, and 3D modelling. Nearly 50 of the projects it completed had never been completed by an earlier agent. The index site describes the pace as a doubling about every 3.3 months over the past year. From 2.5 percent in October 2025 to 41.3 percent in November 2026 is a little more than four doublings in 375 days, so the arithmetic behind that pace claim roughly holds.

Context from our own coverage sharpens the point. Writing work is already feeling this pressure, as our report on writing freelance postings falling in a year showed. A benchmark built from whole projects speaks to that market in a way a quiz score cannot. A client does not buy a quiz answer. A client buys a file they can use.

What Works

The first strength is the size and shape of the lead. This is not a one point win on a crowded board. Fable 5.2 leads the next model by about 20 points, which is more than the entire gain the board recorded from its launch best to the previous leader. On the Anthropic table the pattern repeats in other long task measures, with FrontierSWE v2 at 69.9 percent against 65.5 percent for GPT 6 Astra and 62.3 percent for Opus 5.5, and Terminal Bench Science at 74.6 percent against 58.7 percent for Opus 5.5. Those wider figures are Anthropic reported figures, not independent results, and this review treats them as supporting evidence rather than proof.

The second strength is that the test punishes the usual agent tricks. A pretty summary does not pass. The grader opens the actual files in tools such as Blender for 3D work or design software for plans, then compares the result with work a professional sold. Projects that need many files to agree with each other expose weak agents fast. Passing four in ten of those jobs is a real capability signal, even with every caveat in the next sections attached.

The third strength is candour in the buying advice, which is rare at a launch. Anthropic tells most users to start with Opus 5.5 for coding and knowledge work and to move up to Fable only for the most ambitious work. That advice matches the price and the scores. It also gives this review a clean test for its verdict. If the maker says the cheaper model is the default, a buyer needs a specific reason to pay double. The freelance result is the first public reason that is strong on its own terms.

For a sense of where Fable sits in daily Claude use, compare the office tools in our Claude Sheets and Whiteboards review and the agent setup in our Claude in Chrome and Cowork review. Those products are about staying inside a file with a team. Fable is about handing a whole job to the model and judging what comes back.

What Does Not

The failure rate is the headline nobody puts on a launch slide. Fable 5.2 failed 58.7 percent of the projects in this set. If you hired a freelancer who missed most briefs, you would stop hiring them. An agent is cheaper to retry than a person, and a failed attempt wastes time and tokens rather than a fee for finished work. Still, any workflow built on this model needs a human who checks every deliverable and a plan for the majority case, which is failure.

The index team own analysis of failed work, from the paper that launched the test, shows why the failures matter. Of failed deliverables, 45.6 percent were poor or amateur in quality. A further 35.7 percent were incomplete, such as a video far shorter than the brief asked for. Technical and file faults accounted for 17.6 percent, and 14.8 percent contradicted themselves across files, such as a house design that changed between views. A deliverable could fail in more than one way, so those shares overlap. They describe the launch era agents rather than Fable 5.2 alone, but Scale Labs still names industrial product design and animation as areas that remain far from solved.

Two missing numbers limit how far a buyer can trust the score. Scale Labs states that there is no margin of error figure for the automation rate, so readers cannot tell how wide the uncertainty around 41.3 percent runs. The leaderboard also gives one score per model, with no published rate for how often the same project passes on a second run. A model that passes a job once in three tries costs a different amount from a model that passes it every time, and the public record does not separate those cases.

The setup behind the scores also changed during the year the board climbed. Early agents ran in standard setups with a tight per project budget. Later runs used stronger agent software for each model family, a desktop with more than 30 professional applications, up to 24 hours per project, and a second agent that critiqued the work before delivery. That is fair as a test of peak capability, and the index team describes it openly. It also means the rise from 2.5 percent to 41.3 percent mixes better models with a better harness. The published scores do not separate the two, so do not credit the model alone for the whole climb.

Cost is the last limit, and it is the one a buyer feels first. Fable 5.2 costs twice Opus 5.5 per token. Its safeguards block most cyber work and much biology research work on the public version, and Anthropic reports that blocked cyber queries route to an older Opus model. Paid plan access carries weekly limits and launch reset rules that vary by plan and account age. None of that makes the model bad. It makes the model a specialist purchase, not a default upgrade.

Competition

The choice is rarely Fable or nothing. For most buyers the real contest is Fable against Opus inside the same account, then against the best agent from a rival lab for the same job.

One rival result deserves a fair note. On AutomationBench, a business workflow test in the Anthropic table, Gemini 4 Carbon scored 53.9 percent against 51.4 percent for Fable 5.2. On a knowledge work ranking in the same table, Opus 5.5 held an Elo lead of 1800 to 1785. Fable does not win every line, and a team whose work looks like office workflow automation should weigh that line before it pays the Fable rate for every task.

Final Verdict

Use Fable 5.2 when the job looks like the jobs in this index. The brief is written, the input files exist, the deliverable is a set of files a person can open and judge, and the work is hard enough that Opus 5.5 stalls or produces work you have to redo. In that narrow and valuable case, the freelance result is the best public evidence any current model has produced, and the price can pay for itself in one accepted project that would otherwise take a human a day or more.

Start with Opus 5.5 for everything else, exactly as Anthropic advises. Run the same job on both models once, compare the files yourself, and count the tokens each run used at the effort setting you actually plan to keep. That single check tells you more about your bill and your pass rate than any leaderboard can, because your briefs, your files and your standard of acceptance are the test that matters to your work.

The honest summary is two numbers held at once. 41.3 percent is a genuine leap on a test built from real paid work and graded by people. 58.7 percent is the share that still failed. A buyer who plans for both numbers will get real value from this model. A buyer who plans only for the first number is buying a launch slide, not a worker.

This is a research review. It draws on the Scale Labs Remote Labor Index leaderboard and index documentation, the paper that introduced the index, Anthropic launch material for Fable 5.2 and Mythos 5.2, and published cost testing by Artificial Analysis as reported this week, all checked November 12, 2026. No hands on testing was performed.

Sources

• Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2
• Scale Labs: Remote Labor Index leaderboard
• Center for AI Safety and Scale AI: Remote Labor Index project site
• Mazeika et al.: Remote Labor Index, Measuring AI Automation of Remote Work, arXiv 2510.26787
• MACCUS: On a test of real freelance projects, the best AI agent now completes 41 percent of them
• Devlery: Claude Fable 5.2 doubles real freelance task completion to 41 percent and costs 20 percent less
• Images: Dirty Static house style illustrations