A closer look at Claude Fable 5.2

AI-powered agencies have a margin opportunity.

This is the result that makes me think about agency economics. Fable completed 41.3% of the freelance projects in the evaluation, up from 17.9%. That’s roughly four in ten projects meeting the human professional’s standard, compared with fewer than two before. If you sell a finished deliverable for a fixed fee, reducing the production hours leaves more of that fee in your business. You can take on another project, spend more time with the client, or keep the margin.

Hardware prototyping could be the next big wave.

Fable’s tool-assisted CAD reconstruction score jumps from 44.4 to 83.5, almost doubling. That means better results turning reference designs into editable geometry. If small teams can explore more versions before paying for extensive engineering work, I’d expect new AI businesses around hardware prototyping. The opportunity is cheaper design iteration, with engineers still checking what can actually be built.

Let your agents take the night shift.

Give an agent a well-defined engineering job, a working environment and tests, and it can keep going while you’re offline. Fable scores 69.9% on long-running engineering tasks, up from 56.3%.

There are concrete examples in Anthropic’s launch report: Uber describes a 10 to 12-hour security audit that returned a write-up and fix pull requests. Inside Anthropic, an engineer had Fable coordinate more than 1,000 subagents for eight days to build a prototype operating system that booted on a laptop. Those are the kinds of runs that make the night-shift idea tangible.

Agents can finish over half the business workflows in this test.

Fable completes 51.4% of business workflows end to end, up from 31.4%. The test covers sales, marketing, finance, support, operations and HR across 47 apps. Think about a sales handoff: identify the right company, update the CRM, apply the routing policy and send the next person the context. You can hand an agent the whole sequence instead of prompting it through each step. Carbon leads this test at 53.9%, so you have more than one model worth considering for this work.

Legal work still needs people who know the law.

Fable’s all-pass rate is 7.2%, compared with Grok’s 9.4%. Every requirement has to be met without a material hallucination, and the models still struggle.

I expect more well-paid legal evaluation work as labs try to close this gap. Mercor already lists legal AI-training projects at $100 to $150 an hour, and Surge AI recruits legal specialists into its expert network. Writing demanding test cases and catching mistakes in model answers requires the expertise these models are still missing.

More of your agent’s code is ready to merge.

Fable’s score on producing mergeable code jumps from 32.4% to 49.1%. If agents handle more of the implementation, there’s more value in understanding a client’s problem, connecting their systems and getting the solution into production. That’s the work of a forward-deployed engineer (FDE).

The hiring data is already moving: Pave ranks FDE as its second-hottest tech job in Q3 2026, while Top FDE Jobs counted 1,465 postings across 499 companies in October. My bet is that better coding agents make this combination of customer judgment and hands-on engineering even more valuable.

Use High effort for routine coding before paying for Max.

Fable scores 56.9% at High and 60.0% at Max. Moving to Max buys another 3.1 points while using roughly 3.6 times as many output tokens. For routine fixes, High is a sensible starting point; save Max for difficult changes that benefit from more reasoning. Across thousands of coding tasks, the effort setting can make a substantial difference to your bill.

Fable 5.2 is your research intern.

Remember when OpenAI said it had reached its automated research-intern milestone, with agents already running throughout its researchers’ daily work? Fable gives you a way to put that idea to work yourself. It scores 74.6% on scientific workflows, up from 52.6%. You can ask an agent to set up a computational experiment, write and run the analysis code, compare outputs, and collect the results for you to review.

Better at reading charts.

Fable goes from 44.8% to 73.4% on reading professional charts, without tools. You can ask it to pull quarterly values from a chart in a client’s deck, compare trends across reports, or identify which segment drove a change. Carbon is close at 72.5%, so the big improvement here is over the previous Fable.

Turn a folder of files into client spreadsheets and presentations.

Think about a client project that starts with research, source documents and several spreadsheets. You need to reconcile the information, build an analysis and turn it into a presentation someone can use. AA-Briefcase tests this kind of complex knowledge work, with linked tasks and thousands of source files. Fable reaches a 66.1% rubric score, up from 61.5%. It leads on meeting the requirements, while Opus has the higher combined quality rating, 1800 Elo versus 1785.

Build and audit financial models for client work.

For example, you could give an agent a company’s financial statements and ask it to build a revenue forecast, check the formulas and prepare a memo explaining the assumptions. Real-World Finance tests building and auditing financial models and producing client-ready work. Its grader preferred Fable’s output to Opus’s in 74% of comparisons. Fable missed the three-hour deadline on six of 123 tasks, though, so this looks more useful for an overnight analysis than a spreadsheet you need just before a meeting.