Claude Fable 5.1: How It Handles Real Knowledge Work
Claude Fable 5.1 tested on spreadsheets, decks, and financial models at different effort settings, compared against GPT-5.6 Soul for real knowledge work.
What is Claude Fable 5.1 actually good at?
Claude Fable 5.1 is Anthropic’s latest model tested here on knowledge work, the kind of open-ended tasks that don’t have a clean pass/fail signal the way code does. Instead of benchmarking it in isolation, one creator ran it against GPT-5.6 Soul on real assignments: building a discounted cash flow model in Excel, turning that into a PowerPoint deck, writing a short explainer, and even generating a 3D architectural film from scratch using Blender. The pattern that emerged mattered more than any single output: quality scaled with effort setting, and each model had a different strength depending on what stage of the work you were in.
TL;DR
- Effort settings matter as much as model choice. Fable 5.1 on low produced a finished, usable seven-sheet workbook and 13-slide deck for a real acquisition analysis, while extra effort added deeper diligence like a weighted average cost of capital, deal-close probability, and 26 linked sources.
- Low effort is a genuinely useful default, not just a cheap fallback. It’s described as good enough for a strong first pass on a financial model, which matters given tighter token limits on Anthropic’s side compared to ChatGPT.
- Fable 5.1 writes noticeably better than Fable 5, cutting back on decorative metaphors and improving causal clarity in a hundred-word writing test about Toyota’s US market entry.
- GPT-5.6 Soul’s edge is verifiability, not raw output. Its workbook included a dedicated sources sheet and a check sheet with explicit pass results, making it easier to hand off to another analyst.
- None of the models agreed on the final valuation number, landing anywhere from roughly $1.14 to $1.44 a share on the same GoPro/Starman acquisition scenario, underscoring that these are drafts to interrogate, not answers to trust outright.
- Fable 5.1 can generate a full 3D video from a single address, building a house, terrain, lighting, and camera path in Blender through code, entirely on its own initiative.
- Knowledge work resists one-shot answers. Unlike code, which either compiles or doesn’t, spreadsheets and decks need iteration, and the smart workflow may mean chaining models: draft in one, verify in another, polish in a third.
How does Fable 5.1 perform on low versus extra effort settings?
The test case was a real, current event: GoPro’s acquisition by a company called Starman. Each model was asked to research the deal, build a post-acquisition discounted cash flow model, put it in Excel, and turn it into an executive-readable deck.
On low effort, Fable 5.1 built a seven-sheet workbook and a 13-slide deck from a relatively short prompt. The formulas worked, the scenarios meaningfully shifted the valuation, and the model accounted for real factors like cash burn and the possibility that the combined company would need more funding. It landed on a base case of $1.15 a share, with bear and bull cases around $1.14 and $1.44 (reported in the run as $114 and $144, likely typographical shorthand for per-share cents-level rounding in the original commentary). The gaps: no dedicated sources sheet, no checks sheet, meaning the file was complete but harder to audit.
On extra effort, the workbook grew to nine sheets and 15 slides. This version treated GoPro and Starman as separate businesses, added an explicit probability of deal closure, used a weighted average cost of capital, added an exit multiple check, and linked 26 sources. It separated deal value today from value at close and arrived at a different per-share figure, $1.30 (again reported in the source as $1.3). The extra setting didn’t just pad the document length. It surfaced due-diligence questions that could plausibly change an investment decision.
The practical takeaway: low is fast enough for an honest first draft, and extra is worth the added tokens when you already know where you’re headed and need to stress-test assumptions.
How does Fable 5.1 compare to GPT-5.6 Soul on the same task?
GPT-5.6 Soul, run at extra-high effort, produced a tighter package: a 10-sheet workbook and a 10-slide deck. Its base case came in more optimistic at $1.21 a share. The standout feature wasn’t the number, it was structure. Soul included a dedicated sources sheet and a check sheet with an explicit pass result, which makes the workbook far easier to hand to another analyst for review. The tradeoff was visual polish: the PowerPoint output was described as less attractive than Fable’s, with Anthropic’s models generally producing better-looking front-end design.
The conclusion drawn wasn’t that one model beat the other. It was that they’re useful at different points in a workflow. Fable 5.1 on low gives a fast first view of a problem. Fable 5.1 on extra surfaces harder questions and looks better doing it. Soul makes the underlying logic easier to verify. A realistic process might mean drafting in Fable, checking structure in Soul, then finishing the polish back in Fable.
Does Fable 5.1 write better than earlier versions?
A hundred-word writing test asked each model to explain how Toyota entered and won the US car market. This constraint forces a model to choose what to keep and what to cut, which reveals a lot about its judgment.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Fable 5 packed in the failed 1957 launch, the Corolla, the oil crisis, the Toyota production system, US manufacturing, Lexus, and Toyota passing GM, all in a hundred words. That’s a lot of information, but it leaned on stock phrases like “thirsty Detroit models,” adding flourish before fully explaining the underlying logic.
Fable 5.1 kept the key facts (the failed Crown launch, the Corona and Corolla, the oil crisis, Toyota passing Volkswagen as the top import brand, Japan’s export limits pushing Toyota to build US plants) but connected them with clearer cause and effect and cut the decorative language.
Soul took a different approach entirely, using fewer dates and milestones in favor of a smoother narrative arc from Toyota’s failed first car to its reputation for reliability. Fewer facts, but a clearer overall story.
None of these is objectively “best.” The choice depends on audience: a general audience might read better with Soul’s narrative flow, while an executive audience might prefer Fable 5.1’s fact density and causal clarity.
Can Fable 5.1 handle visual and video generation?
One of the more unusual demonstrations involved giving Fable 5.1 a real property address in Seattle and asking it to build a cinematic architectural walkthrough using Blender, a 3D modeling tool that relies on code. There was no existing 3D model, just an address and a loose brief.
The result was a 37-second film: a camera path starting outside the house, moving through the entrance, into the living room, up toward the ceiling, and back outside above the property. Fable built the house, terrain, interiors, lighting, landscaping, and camera path entirely in code, rendered stills, inspected them, revised scenes, checked a motion preview, and rendered the final sequence, all without human intervention in Blender itself.
The output wasn’t flawless (trees were stylized, some glass rendering was simple), but as a way to quickly communicate a visual concept to a client or collaborator, generating a rough cinematic draft in under an hour is a meaningfully different capability than what earlier models produced. Fable 5’s version of the same task was rougher and shorter. Soul’s version was clean but very brief.
Why does knowledge work resist a single “best model” ranking?
Code has a built-in verification signal: it compiles or it doesn’t, tests pass or fail. That’s part of why code-focused agent tasks have taken off. Knowledge work, spreadsheets, decks, written analysis, doesn’t offer that courtesy. There’s no compiler that tells you a valuation model or a deck is “done.” Quality is judged by whether the reasoning holds up, whether the assumptions are visible, and whether another person can pick up the work and trust it.
That’s why comparing Fable 5.1 and GPT-5.6 Soul isn’t really about crowning a winner. It’s about matching a model and effort setting to the stage of work you’re in: fast drafting, deep diligence, verification, or polish.
Frequently Asked Questions
What is Claude Fable 5.1?
It’s Anthropic’s model tested here on real-world knowledge work tasks, including building financial models in Excel, generating PowerPoint decks, writing explanatory text, and producing 3D video through Blender, across different effort (compute) settings.
What’s the difference between low and extra effort settings in Fable 5.1?
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
Low effort produces a complete, usable draft quickly, such as a seven-sheet workbook with working formulas, but skips extras like a sources sheet or checks sheet. Extra effort adds deeper diligence: more sheets, cited sources, alternative valuation methods like weighted average cost of capital, and explicit probability assumptions.
Is Fable 5.1 better than GPT-5.6 Soul?
Neither model won outright in the tests described. Fable 5.1 produced more polished decks and stronger causal writing, while Soul’s output was easier to verify thanks to a dedicated sources sheet and checks sheet with pass/fail results.
Can Fable 5.1 really generate video?
Yes, in the sense that it can write code that drives Blender, a 3D modeling program, to build a scene, position a camera, and render a video sequence, without the user needing any Blender experience.
Why did the models disagree on the GoPro/Starman valuation?
Complete financial history wasn’t publicly available for the acquiring company, so each model had to build scenarios around missing information. The differing per-share estimates (roughly $1.14 to $1.44 across runs) reflect different assumptions, not necessarily correctness.
