Claude Fable 5.2: Every Score, and When I'd Pick It Over Opus 5.5
Every score from the launch, what a job really costs, when I'd switch from Opus 5.5, and the one prompt behind the three builds in my video.
Back to all free guidesSo this is my full breakdown of Claude Fable 5.2, the model Anthropic released on 9 November. You get every score from the launch, what a job on it costs, where I'd use it instead of Opus 5.5, and the exact prompt I used for the three builds in my video.
The short version
- What it is. Anthropic calls Fable 5.2 its "most capable generally available model for technical fields". Mythos 5.2 is the same model with lighter safeguards, and it only goes to vetted organisations, so Fable is the one you can actually use.
- Where you get it. The Claude apps on Pro, Max, Ultra, Team and Enterprise, but not the free plan. On the API it's
claude-fable-5-2. - The price went down. $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1. That's 20% off both.
- It's quicker. Up to 30% faster than Fable 5.1, and Fast mode speeds it up by up to 2.5 times for double the price ($16 and $80).
- Anthropic's own advice. Start most coding and knowledge work on Opus 5.5, and switch up to Fable when the job needs it.
Every score from the launch
These are the numbers from Anthropic's launch table, at the highest effort each model offers. Fable 5.2 is the first number in every row.
| Test | What it measures | Fable 5.2 | Fable 5.1 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|---|---|
| FrontierCode 2.0 | Merge-ready code changes | 49.1% | 32.4% | 45.5% | 42.8% | 41.1% |
| FrontierSWE v2 | Agentic coding | 69.9% | 56.3% | 62.3% | 65.5% | 61.2% |
| AutomationBench | Business workflows | 51.4% | 31.4% | 42.5% | 41.4% | 53.9% |
| WANDR | Agentic search | 82.1% | 68.7% | 72.3% | - | - |
| OSWorld 2.1 | Using a computer | 89.3% | 83.5% | 87.2% | 79.5% | 81.5% |
| RoboDojo Sim | Controlling a robot | 64.9% | 27.6% | 39.4% | 29.8% | - |
| AA-Briefcase v1.1 | Knowledge work (rubric) | 66.1% | 61.5% | 61.2% | 52.0% | 61.8% |
| AA-Briefcase v1.1 | Knowledge work (Elo) | 1785 | 1669 | 1800 | 1569 | 1611 |
| Remote Labor Index | Paid remote work | 41.3% | 17.9% | 21.3% | 20.8% | - |
| Humanity's Last Exam | Hard questions, with tools | 74.2% | 65.6% | 67.7% | 57.2% | - |
| Chartography | Reading charts, no tools | 73.4% | 44.8% | 64.4% | 71.0% | 72.5% |
Fable 5.2 beats Fable 5.1 on every row. Two rows go to someone else: Gemini 4 Carbon leads AutomationBench, 53.9% to 51.4%, and Opus 5.5 leads the AA-Briefcase Elo, 1800 to 1785.
I've left out the agentic science research test (Terminal-Bench-Science), because its cell on Anthropic's page holds a Fable number and a Mythos number together and I can't tell you which is which.
Three more rows are for Mythos 5.2, which you can only get as a vetted organisation:
| Test | What it measures | Mythos 5.2 | Mythos 5.1 | Opus 5.5 | GPT-6 Astra |
|---|---|---|---|---|---|
| CyScenarioBench | Cybersecurity | 76.3% | 61.7% | 67.6% | 61% |
| CVE Rediscovery | Finding known security holes | 89.6% | 78.3% | 81.3% | - |
| De novo binder design | Protein design (24 hours) | 85.1% | 79.1% | 82.6% | - |
Where it beats Opus 5.5, and where it doesn't
Opus 5.5 is the model most people are on, so this is the comparison that matters.
- Big leads: controlling a robot (64.9% to 39.4%), paid remote work (41.3% to 21.3%, nearly double), agentic search (82.1% to 72.3%), reading charts (73.4% to 64.4%), business workflows (51.4% to 42.5%) and agentic coding (69.9% to 62.3%).
- Close: merge-ready code changes (49.1% to 45.5%) and using a computer (89.3% to 87.2%).
- Opus 5.5 wins: the AA-Briefcase Elo, 1800 to 1785.
The clearest gap isn't in the table at all. A team of 1,000 Fable 5.2 agents fully formalized Fermat's Last Theorem in 30 hours, which means rewriting the proof so a computer can check every step. Opus 5.5 got 26% of the way in 18 hours with the same setup.
What a job really costs
The price per token is only half of it. What you actually pay for is the whole job, and Anthropic's figures on that are where Fable 5.2 gets interesting:
- Humanity's Last Exam on Low effort: it matches Opus 5.5's top score for less than a tenth of the cost per task.
- WANDR on Medium effort: it matches Opus 5.5's top score for about 40% of the cost per task.
- FrontierCode on High effort: 45.3% against Opus 5.5's best of 45.5%, for roughly half the cost per task.
So turn the effort down before you write it off as too expensive. Medium is the default in the Claude apps and Claude Code, and on Medium it gets hard jobs done with fewer tokens than Fable 5.1 did.
My three builds in the video came to about 13 million tokens, about $20 and about 45 minutes. That includes a couple of extra asks I'd forgotten to put in the prompt.
One thing to watch on a plan: on Pro and standard Team seats, Fable can use up to 50% of your weekly usage limit.
When I'd pick Fable 5.2 over Opus 5.5
This is how I'm splitting it for coding:
- Stay on Opus 5.5 for everyday work. Small fixes, writing and quick questions. Anthropic says to start there too, and on merge-ready code changes the gap is under 4 points.
- Switch to Fable 5.2 for long agentic jobs. Work that runs across a whole codebase, or a build you want done in one go. That's where the 69.9% to 62.3% agentic coding gap shows up.
- Switch for anything visual. The website, the reactor and the pet in my video were mostly one-shot builds from a single prompt, and the design quality is the part I'd pay for.
- Switch for agents that use a browser or a computer, and for research that leans on search. Fable is ahead of Opus on using a computer, on agentic search and on reading charts.
- Start it on Low or Medium effort. That's where it matches Opus's best scores for a fraction of the cost per task. Only go up if the answer isn't good enough.
If you're on the API, set two things up first. Cyber questions Fable won't answer fall back to Opus 4.8 and biology ones to Opus 5, and you configure that with Anthropic's new Fallback API. Fable models also keep your data for 30 days by default for safety monitoring.
The prompt I used for the three builds
Paste this into one fresh chat with Fable 5.2 selected. It asks for three separate artifacts, one at a time: a redesign of my website, a nuclear power plant you can control, and an original virtual pet. If Claude stops partway through, type continue.
Show me how good your design is. Build three separate pieces. Make each one its own artifact: a single self-contained HTML page I can open, click and use right here in the chat. Put all CSS and JavaScript inline, load any library from cdnjs.cloudflare.com, need nothing installed or built, and don't rely on browser storage. Do them one at a time, finish each one completely before you start the next, and give each your full effort. If you run out of room, stop at a clean point and I'll say "continue".
1. A redesign of my website, abaisolutions.com.
Read the live site and keep its real copy, offer and numbers. Do not invent clients, testimonials, stats, logos, an address or company details. ABAI is my AI automation practice, an "AI growth partner" for service businesses. Never call it an agency or a company.
Brand, exactly: black #000000, surface #100E20, white #FFFFFF, purple #5600FF, hover #7C3AED. No other purple, no lavender or pale tints. Secondary text is white at lower opacity, never grey. The wordmark is "ABAI" with "AB" in white and "AI" in #5600FF.
Keep the six steps: Strategy call (free, 30 minutes), Blueprint (fixed fee in writing), Build (around the tools you already use), Launch (most live in 2 to 4 weeks; the workflows, accounts and data are yours), Train (your team), Improve (alerts catch breaks before you notice). Keep the four real results: "569 leads in one pipeline. Zero duplicates." / "Deal to cash, with nobody chasing." / "A signed-off ticket becomes a paid invoice, end to end." / "Each client's report ready before the call starts." Add an FAQ and one call to action: book a free 30-minute strategy call.
It should look like a top design studio built it, not an AI template. One strong art direction, bold type, generous spacing, and motion only where it means something. No gradient blobs, glows, emoji or "AI brain" imagery. It must work perfectly on a phone. Tell me your art direction in three lines before you build it.
2. The most impressive interactive visualisation you can make of how a nuclear power plant works.
Use a pressurised water reactor. One living cutaway shows the whole chain: fuel rods, a neutron splitting U-235, the chain reaction, the pressurised water, the steam generator, the turbine, the generator, the cooling tower and the water looping back. Add a control rod slider that visibly slows or speeds the chain reaction, with live gauges for reactor power, core temperature and megawatts out. Add a zoom from the whole plant into one fuel pellet and then a single fission event, a SCRAM button that drops every rod at once, and a two-sentence plain-English explanation when I click any part.
It should look like a premium science museum exhibit, with real depth, lighting and motion (Three.js or WebGL is fine), the blue Cherenkov glow and moving steam. Keep the physics honest, with real figures for a typical 1,000 MW plant. No mushroom clouds, because a reactor can't explode like a nuclear bomb.
3. An original handheld virtual pet, not based on any existing toy or game.
The device is a 3D model I can rotate, with a real industrial-design feel: materials, a soft-touch finish, a lanyard loop, and three buttons that press in when clicked. Give it a product name and a colourway. Its screen runs the game: an original pixel-art creature with a name and a personality. With the three buttons I can feed it, play a small mini-game, clean up after it and put it to sleep. It gets hungry and bored over time and grows through three stages. Below the device, add a product box design and a one-page spec sheet, like a real launch.
It should feel like a product you'd want to buy, with a game that's actually fun for five minutes.Part 1 is my own site. To redesign yours, swap part 1 for this and fill in the brackets:
1. A redesign of my website, [your site].
Read the live site and keep its real copy, offer and numbers. Do not invent clients, testimonials, stats, logos, an address or company details.
Brand, exactly: [your colours as hex codes, and what each one is for].
Keep these sections and facts: [the steps, results and offer that must stay word for word].
It should look like a top design studio built it, not an AI template. One strong art direction, bold type, generous spacing, and motion only where it means something. No gradient blobs, glows, emoji or "AI brain" imagery. It must work perfectly on a phone. Tell me your art direction in three lines before you build it.Every number on this page comes from Anthropic's launch post, "Introducing Claude Fable 5.2 and Claude Mythos 5.2", and the Fable product page on anthropic.com, read on 10 November 2026.