Claude Fable 5.2 Explained: Pricing, Benchmarks, and Effort Levels
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
Anthropic just released Claude Fable 5.2 and Mythos 5.2, and in this video I go through everything we know so far: the price, the benchmarks (including the two it loses), how the effort dial works, and what early-access teams built with it. If you build apps or agents with Claude and you're wondering whether to switch, this is for you.
What I cover in this video
- 0:00 The new price, and 1,000 agents formalizing Fermat's Last Theorem in 30 hours
- 0:45 What Fable 5.2 and Mythos 5.2 actually are
- 2:17 The benchmarks, the wins and the losses
- 3:51 The effort dial, the advisor and orchestrator setups, and the safety layer
- 5:24 What early-access teams built with it
- 6:55 Should you switch?
- 7:40 The caveats before calling it the best model
What Fable 5.2 is and what it costs
Anthropic released two models at once: Claude Fable 5.2 and Mythos 5.2. Most people don't realize they're the same underlying model. Fable comes with strong safeguards around cyber and biology, and anyone can use it. Mythos is for vetted organizations doing cyber or life sciences work, and they have to apply through Anthropic's verification programs. So for most of us, it's Fable.
It's out on all paid plans, in Claude Code, on the API, and on AWS, Google Cloud, and Azure. Here are the basics:
- Price: $8 input and $40 output per million tokens, down from $10 and $50 for Fable 5.1.
- Speed: up to 30% faster output than Fable 5.1.
- Fast mode: up to 2.5x faster, but it costs double, $16 in and $80 out.
- Effort levels: Low, Medium, High, Xhigh, and Max. Medium is the default in the Claude apps and Claude Code, and at Medium it uses fewer tokens than 5.1 did.
- Resets: eligible paid plans get one to three Fable resets depending on how long you've been subscribed.
The benchmarks, both sides
Keep in mind these are Anthropic's own numbers. When I recorded this, no independent leaderboard had posted its own scores yet.
On FrontierCode 2.0, their agentic coding test, Fable 5.2 at Xhigh scores 49.1%. Opus 5.5 is at 45.5%, GPT-6 Astra at 42.8%, and Gemini 4 Carbon at 41.1%. Fable 5.1 was at 32.4%, so that's a big jump in about 10 weeks. On FrontierSWE v2 it scores 69.9% vs 62.3% for Opus, and on Humanity's Last Exam with tools it's 74.2% vs 67.7%.
The robotics number stood out to me. On RoboDojo, a simulated two-arm robot, Fable 5.2 scores 64.9%, up from 27.6% for Fable 5.1. That's more than double. On the Remote Labor Index at High effort, it scores 41.3% vs about 21% for Opus.
It doesn't win everything, though:
- AutomationBench: Gemini 4 Carbon beats it, 53.9% to 51.4%.
- AA-Briefcase Elo: Opus 5.5 is still ahead, 1800 to 1785, even though Fable wins the rubric score on that same test.
So it's not a clean sweep, and I respect Anthropic for putting those in the table.
For me the cost is the real story. On FrontierCode, Fable at High effort scores 45.3% and Opus at Max scores 45.5%, and Fable gets there for roughly half the cost per task. On Humanity's Last Exam, Fable at Low effort matches Opus's top score for under a tenth of the cost. The effort dial is where the savings are.
How you're meant to use it
Think of effort as a dial. Higher effort means the model spends more tokens on the task, and more tokens means more money. So you pick the lowest setting that still gets the job done.
Anthropic describes two setups for working with other models:
- The advisor: your normal model (Haiku, Sonnet, or Opus) does the work, and when it gets stuck it asks Fable. You only pay Fable prices on the hard parts.
- The orchestrator: Fable is the boss. It breaks a big job into pieces and hands the easy ones to smaller, cheaper models.
The safety layer works differently than you might expect. If a request trips the cyber filter, you don't just get a refusal. The answer comes from Opus 4.8 instead. Blocked biology questions get answered by Opus 5. Anthropic says it blocks about 55% fewer normal requests than 5.1 did, so you should hit that wall a lot less.
They also turned on something called preserved thinking for every API account. It stops people from tampering with Claude's earlier context to pull out its reasoning and train their own models on it. Anthropic says its anti-copying measures overall cut that harvested data by 90 to 99%.
What early-access teams built
Since the model just came out, most of these examples come from teams with early access:
- A prototype operating system from an empty folder. Over 1,000 subagents ran for 8 days, and it boots on a laptop with disk, keyboard, and network drivers, file storage, windows, a login screen, an installer, and Claude built in.
- Fermat's Last Theorem in Lean. 1,000 agents turned the proof into code a computer can check step by step in 30 hours. Anthropic's first run with an unreleased research model took 11 days, and the new proof came out seven times shorter. Opus 5.5 with the same setup got through 26% in 18 hours and barely progressed after 30.
- Python to Rust. A team of agents moved a production service to Rust, and the new one starts about 26x faster.
- Hardware with Diode. Three electronic devices from one paragraph: a voice assistant hub, a USB speakerphone, and a handheld virtual pet. Each board was ready to manufacture in under a day, and all three worked without a single electrical fix.
- Uber: an engineer said Fable ran a 10 to 12 hour security audit using a team of reviewer agents.
- Oracle: Fable's SQL gave the right answer while two other models summed totals 5 to 8x too high.
- Discovered Materials: simulation code ran about 36x faster on average. 13 test cases went from 3.3 hours to under 4 minutes.
If you're into multi-agent setups like these, my AI agents page has more.
Should you switch?
Anthropic themselves say that for most coding and everyday work you should start with Opus 5.5. That's their daily driver, and Fable is for big, ambitious, open-ended work. I agree. I'm keeping Opus as my daily driver. For the hard stuff, like a big migration, a bug nobody can find, or a project that would take me a week, I'm reaching for Fable. On Pro, Fable is included for up to half of your weekly limit, so I'm saving those resets for one big project instead of burning them on emails.
A couple of caveats. Anthropic says the model mentioned that it thought it was being tested more often than past models did, which makes testing it harder. They also deliberately held back a few of its AI research skills, like chip kernel tuning and model architecture work, to about Opus 5.5 level. I want to see independent scores before I call it the best model in the world.
Key takeaways
- Fable 5.2 is cheaper than 5.1: $8 in and $40 out per million tokens, with an optional fast mode at $16 and $80.
- At High effort it roughly matches Opus 5.5 at Max on FrontierCode for about half the cost per task.
- It loses on AutomationBench to Gemini 4 Carbon and on AA-Briefcase Elo to Opus 5.5.
- Use the lowest effort level that gets the job done, and consider the advisor or orchestrator setups to keep costs down.
- Keep Opus 5.5 as your daily driver and save Fable for large, hard projects. All benchmark numbers are Anthropic's own for now.
FAQ
How much does Claude Fable 5.2 cost?
Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1. Fast mode is up to 2.5x faster and costs $16 in and $80 out.
What is the difference between Fable 5.2 and Mythos 5.2?
They're the same underlying model. Fable has strong safeguards around cyber and biology and anyone can use it, while Mythos is for vetted organizations doing cyber or life sciences work that apply through Anthropic's verification programs.
Is Fable 5.2 better than Opus 5.5?
On Anthropic's own benchmarks it beats Opus 5.5 on FrontierCode 2.0, FrontierSWE v2, and Humanity's Last Exam, but Opus 5.5 still leads on AA-Briefcase Elo. Anthropic recommends starting with Opus 5.5 for most everyday coding and using Fable for big, open-ended work.
What effort level should I use with Fable 5.2?
Higher effort spends more tokens and costs more, so pick the lowest level that still gets the job done. The options are Low, Medium, High, Xhigh, and Max, and Medium is the default in the Claude apps and Claude Code.
Chapters
Transcript
Read the full transcript
0:00Anthropic just dropped Claude Fable 5.2 and it's cheaper than the model it replaces. $8 in, $40 out per million tokens, down from 10 and 50. And on their hardest coding benchmark, Fable 5.2 at high effort scores about the same as Opus 5.5 at max effort for roughly half the cost per task. Then they posted this. A team of 1,000 Fable 5.2 2 agents took Fermott's last theorem and turned the whole proof into code a computer can check step by step in 30 hours. Anthropic's first time doing that with an unreleased research model took 11 days. So, in this video, we're going to be breaking down everything we know about Fable 5.2. What the benchmarks
0:45actually say, the good and the bad, what people are already building with it, and if you should even switch. Let's jump right into it. So, first, what is it? Anthropic released two models today. Fable 5.2 and Mythos 5.2. And here's something most people don't think about. They're the same underlying model. Fable has strong safeguards around cyber and biology, and anyone can use it. Mythos is for vetted organizations doing cyber or life sciences work, and they have to apply through anthropics verification programs. So, for you and me, it's Fable. It's out today on all paid plans in Cloud Code, on the API, and on AWS, Google Cloud, and Azure. They say it writes up to 30% faster than Fable 5.1.
1:31And there's a fast mode that's up to 2 and 1/2 times faster, but it costs double, 16 in, and 80 out. You also get effort levels: low, medium, high, X high, and max. Medium is the default in the Claude apps and in Claude Code. And at medium, it uses fewer tokens than 5.1 did. and eligible paid plans get one to three fable resets depending on how long you've been subscribed. This is where I run my coding agents. Beckin puts cloud code, codecs, and cursor side by side on one canvas. I hold a key and just talk to them, and Becken talks back when they finish. Every project keeps its own billboard and memory. If you build with agents, try [music] Beckan. Hey, Beckin.ai. So, let's get into the numbers. And remember, these are Anthropic's own
2:17numbers. As I'm recording this, no independent leaderboard has posted its own scores yet. On Frontier Code 2.0, their agentic coding test, Fable 5.2 at XHigh scores 49.1%. Opus 5.5 is at 45.5. GPT6 Astra 42.8 and Gemini 4 Carbon 41.1. Fable 5.1 was at 32.4. So that is a massive jump in about 10 weeks. On Frontier Sui, 69.9 versus 62.3 for Opus. And on humanity's last exam with tools, 74.2 versus 67.7. Now look at robots. On Robo-Dogjo, a simulated two-arm robot, Fable 5.2
3:02scores 64.9. Fable 5.1 was 27.6. That's more than double. And on the remote labor index at high effort, 41.3 versus about 21 for Opus, which is insane, but it doesn't win everything. On automation bench, Gemini 4 Carbon beats it, 53.9 to 51.4. And on the AA briefcase ELO score, Opus 5.5 is still ahead, 1,800 to 1,785, even though Fable wins the rubric score on that same test. So, this is not a clean sweep, and I respect them for putting those in the table. Now, here's what this is all about. It's the cost. On Frontier Code, Fable at high effort scores 45.3. Opus at max scores 45.5,
3:51and Fable does it for roughly half the cost per task. On humanity's last exam, Fable at low effort matches Opus' top score for under a tenth of the cost. So, the effort dial is where the money is. So, how are you actually supposed to use this thing? Here's how it actually works. The effort setting is a dial. Higher effort means the model spends more tokens on the task, and more tokens means more money. So, you pick the lowest setting that still gets the job done. [clears throat] Then, Anthropic talks about two setups. The first one is the adviser. Your normal model, Haiku, Sonnet, or Opus, does the work. And when it gets stuck, it asks fable. So, you only pay fable prices on the hard parts. The second one is the orchestrator.
4:36Fable is the boss. It breaks the big job into pieces and hands the easy ones to the smaller, cheaper models. And then there's the safety layer, which is kind of clever. If you ask Fable something that trips the cyber filter, you don't just get a refusal. The answer comes from Opus 4.8 instead. Blocked biology questions get answered by Opus 5, and they say it blocks about 55% fewer normal requests than 5.1 did. So, you should hit that wall a lot less. One more thing, they turned on something called preserved thinking for every API account. It stops people from messing with Claude's earlier context to pull its reasoning out and train their own models on it. And Anthropic says its anti-copying tricks overall cut that harvested data by 90 to 99%.
5:24So what are people actually building with it? It just came out, so most of this comes from the teams that had early access, and some of it is wild. One engineer had Fable 5.2 build a prototype operating system from an empty folder. It ran over a,000 sub aents for 8 days, and it boots on a laptop with disk, keyboard, and network drivers. file storage, Windows, a login screen, an installer, and Claude built right in. Come on now. Then the Fermont run from the beginning. 1,000 agents, 30 hours, and the final proof came out seven times shorter than that first run. Opus 5.5 with the same setup got through 26% in 18 hours and barely made any more progress after 30. A team of agents
6:10running an early build of Fable 5.2 2 moved a production service from Python to Rust. And the new one starts about 26 times faster. With a company called Diode, it designed three electronic devices from one paragraph with very little human help. A voice assistant hub, a USB speaker phone, and a handheld virtual PET. Each board was ready to manufacture in under a day, and all three worked without a single electrical fix. An engineer at Uber said Fable ran a 10 to 12-hour security audit and sent out about 70 reviewer agents. Oracle said Fable's SQL gave the right answer, while two other models summed the totals 5 to eight times too high. And a company called Discovered Materials got their material simulation code running about
6:5536 times faster on average. 13 test cases went from 3.3 hours to under 4 minutes. So, should you switch? Here's the funny part. Anthropic themselves say for most coding and everyday work, start with Opus 5.5. That's their daily driver. Fable is for the big, ambitious, open-ended stuff. And I agree with that. I'm keeping Opus as my daily driver. And yes, I know slopus. But for the gnarly stuff, a big migration, a bug nobody can find, a project that would take me a week, that's where I'm reaching for Fable. And on Pro, Fable is included for up to half of your weekly limit. So, I'm saving those resets for one big project instead of burning them on emails. Now,
7:40both sides, these are anthropics numbers. They say the model also mentioned it thought it was being tested more often than past models, which makes testing it harder. And they held back a couple of its AI research skills, like chip kernel tuning and model architecture work to about opus 5.5 level on purpose. So, I want to see the independent scores before I call it the best model in the world. So, what's the first thing you're throwing at Fable 5.2? A big refactor? a side project, maybe a robot. Tell me in the comments. And if you got something out of this, hit like. It helps more builders find it. Have a blessed day.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
Claude Fable 5 for Builders: Pricing, Benchmarks, and Model Routing
Claude Fable 5 is Anthropic's strongest widely available model. It has a 1M-token context window and costs $10/$50 per million input/output tokens. I'd use it for hard work like migrations, deep debugging, and code review, and send everyday tasks to Opus 4.8, Sonnet, or Haiku.
Jev by TypeSafe Explained: Pricing, Benchmarks, and Claude Code
Jev is TypeSafe's System One model: you send it a state plus typed questions and it returns a choice, a score, or a probability instead of text. I'd keep Claude Code and Opus for planning and building, and put Jev on the small yes or no calls inside the app.
Gemini 4 Carbon vs Claude Opus 5.5: Benchmarks and Speed Test
Gemini 4 Carbon tied Claude Opus 5.5 on both of my hidden-test builds (12/12 and 28/32 each), but in my runs through OpenRouter it took up to 7.6 times longer and twice saved no file. Carbon costs $2 in / $10 out until November 30, then the same as Opus.
