My New Favorite Model
Thursday, Sep 3 · 55 min · about 10 min to read
Theo-brain watched all 55 min
Theo reviews Anthropic's Fable 5.1 after landing 89 PRs with it in 24 hours. Cache reads are 75% cheaper, so agent work costs less, but it writes more output, so one-shot tasks can cost more. The main lesson: it finishes whole tasks with far fewer serious review findings, so he now lets it fix and merge many PRs by itself.
Skim itWatch the design and game demos and the real-world stretch at the end; the pricing and benchmark sections are covered well enough here.
Parts
Intro: why he went deeper this time
After Opus 5 looked great early but disappointed in real use, Theo bases this review on 89 PRs landed in 24 hours.
nice to use. But considering that we landed 89 PRs in 24 hours using this model, I have a much better idea of where its strengths and weaknesses are than usual. I even took the time to do a
Fable and Mythos are the same model
Fable 5.1 is Mythos 5.1 behind extra filters, like a ghost-kitchen rebrand of the same restaurant.
responses from coming out. So, Fable and Mythos are the same model. The only difference is what happens when you send a request, not the actual weights underneath. So, if you think these are different models, they're not. They're just different doors to the same model that have different restrictions on
Price and why caching matters
Same base price, 75% cheaper cache reads, and a walkthrough of why agents re-read history on every tool call.
They claim that 5.1 will cost an estimated 25% less than Fable 5 for typical workloads whenever usage is billed by token. That's because they're reducing the price on cache reads. It's a 75% decrease on cache reads. It's a
Data retention and safeguards
A new setup lets enterprises run Anthropic's safety checks on their own servers, and false flags are down.
dangerous. So they came up with a compromise. They're introducing the new enterprise frontier safeguards, which is a system that lets them set up a provisioned box to check all the things they care about that you can run on your infra on AWS or wherever else that does the things they care about without it being on their own infra. This is very
Official benchmarks and science results
Mythos 5.1's protein binder designs hit nearly 50% across 12 targets, versus the usual 10 to 15%.
protein design competition. The hit rate for the three targets that it designed for were 10 times higher than the best design submissions that existed for this contest. It's hit rate was nearly 50% across 12 targets. Usually the hit rate expected is 10 to 15%. So that is
Safety, anti-distillation, watermarks
New accounts can no longer edit earlier turns of a chat, to stop people extracting the model's hidden reasoning.
New Claude accounts can no longer edit the prior context in a multi-turn conversation. This is an attempt to make it so any reasoning that the model did,
Independent benchmarks
Artificial Analysis scores it highest ever, but it uses more output tokens and costs more per task.
to do the eval. The model scored higher than they've ever measured before, ahead of Claude Opus 5 and Fable 5 as well, and of course ahead of 5.6 Sol and Grok 4.6. They did call out that of the
Design and game demos
Marketing pages with tasteful animations and a rebuilt Fish Slap game show a leap in front-end work.
were all good and tasteful. And I'll say just outright looking at these, this is a generational leap in home page design at the absolute least. All these marketing pages look much better than even what Fable 5 was capable of. Like
The official prompting guide
Anthropic's guide to behavior changes. On effort levels, he finds high a great default and low surprisingly capable.
But for the most part, high is a great default. Low is surprisingly capable and worth trying out for a bunch of different things.
Real-world use on T3 Code and Lakebed
He gave Fable 5.1 the PRs other agents kept looping on, and pretty much all of them landed.
stuck that for whatever reason the agent was looping on and putting too much code out and not actually completing the work as intended. I pulled a lot of those to Fable 5.1 and pretty much all of them ended up landing. This is one of the
Measured against Fable 5 and Sol
His own PR data: bigger, wider PRs with far fewer serious review findings.
Immediately, it called out that Fable 5.1 was shipping bigger and wider PRs in one day than either of the peak Fable 5 days. It had 13 PRs with an median of 489 lines of code. The most interesting piece here is that it was touching up to four packages each.
Wrap-up: handing up another layer
It will burn through usage limits faster because you let it do more, which is what he wants.
it. This model is definitely going to burn through your limits faster, not because it's less efficient, but because you're going to have it do more and let it go further. And if you let it spin up all the sub agents that it can now orchestrate better cuz it prompts itself better, too. It's going to burn more cuz it's just doing more. But I think in the
Tap any time to jump the video there.
Key takeaways
In Theo's own PRs, review bots flagged 0.4 serious issues per 1,000 lines with Fable 5.1, versus 2.06 with Fable 5 and 1.02 with OpenAI's Sol, even as PRs grew. The cost is speed: merges took up to 50 minutes, versus 47 for Fable 5 and 30 for Sol.
stayed strong when the PR size was going up. Review bots left 0.4 high severity findings per thousand lines of code with Fable 5.1 code. Fable 5 saw 2.06 high severity findings per thousand lines of code. Soul was lower than that at 1.02, but that means that Soul had more than 2x the high severity findings,
take longer though. It took up to 50 minutes for the PRs to merge when compared to 47 minutes with Fable 5, slightly faster, and then Soul being much faster at 30 minutes cuz the whole loop was just closed more. Let's skip
It finishes the whole task. It changes every part of the app a task touches (server, web, desktop and mobile) instead of just one, averaging 11 files per PR versus 4 for Fable 5 and Sol.
the mobile app. Fable 5.1 would make the changes in all the places it mattered instead of just focusing on one piece, which is very nice. It means that it completes the whole task instead of just completing the like isolated code change. It also writes way more code per
files per PR. On average, both Fable 5 and 5.6 Soul would touch four files per PR. Fable 5.1 was touching 11. My
Base prices are unchanged at $10 and $50 per million input and output tokens. Cache reads (reusing conversation history the model already processed) drop 75% to 25 cents. Agents re-read the history on every tool call, so agent work gets much cheaper. Plain chat barely benefits.
break down. Fable 5.1 has the same main price of $10 per million input tokens and $50 per million output tokens, but the cache reads are way cheaper with a 75% discount, only at 25 cents per million tokens for cache reads. The
costs. I was mostly talking about agentic work here. If you're just sending one message and getting back one response with maybe one or two tool calls at most, you're not going to see this big of a difference because you're not hitting cache as much. These caches
It talks more. On Artificial Analysis it used 1.7x the output tokens of Fable 5 and cost 20% more per task at max effort. Theo wouldn't use it for long one-shot jobs that cheaper models can do. The savings show up in agent work.
5.1 Max cost $3.76 per intelligence index task, which is 20% more than Fable 5 cost, because it used 1.7x the output tokens. I mentioned this earlier, I've seen this in my own usage. This model yaps a bit more, and it outputs more as a result. I think
is a little rough and I wouldn't use this model for really long one-shot tasks that could be done by other things. But again, it's the agentic work
Fable 5.1 and Mythos 5.1 are one model. Fable is the publicly available version with more safeguards. Mythos is limited to trusted-access programs for cybersecurity and life-sciences work.
don't worry. They immediately open with Fable 5.1 and Mythos 5.1 are the same model with different levels of safeguards. Fable is generally available, while Mythos is available only through the trusted access programs. Its safeguards are
Three skills here are new to you.
PR takeover skill New
- What it is
- A short skill (saved instructions for an agent) that hands a fresh agent an existing pull request. The branch is now its job: push, maintain and manage it until it lands.
- Why Theo cares
- He used to copy one agent's review feedback into another agent's thread. Letting the next agent simply take over works better. With Fable 5.1, nearly all of his stuck, looping PRs finally landed.
- How to try it
- Ask your agent to write a takeover skill: given a PR, check out its branch, push, maintain and manage it until it merges.
- When a PR keeps looping on AI review comments, hand it to a fresh Fable 5.1 agent using that skill.
- Optionally ask the original agent what it thinks of the new changes.
made?" I made this easy with a takeover skill that very simply tells the model, "Hey, here's a PR. You get the branch on your work tree and it's yours. Push it, maintain it, manage it so that it actually lands." And I let over the PR
as intended. I pulled a lot of those to Fable 5.1 and pretty much all of them ended up landing. This is one of the
Audit, then fan out to sub-agents New
- What it is
- Have the model audit a codebase for messy code and propose an ordered set of PRs. Then tell it to spin up Fable 5.1 sub-agents, watch over each PR until checks and review bots pass, and merge the simple wins.
- Why Theo cares
- On Lakebed it merged all 10 PRs, green on every check, deleting 13,000 lines net without him reading the code. He says it was risky but acceptable because nothing ships to production until he presses the button.
- How to try it
- Ask your agent for a cleanup audit with a proposed order of PRs, deletions first.
- Read the report, then tell it to use sub-agents and watch each PR until all checks and review bots pass.
- Let it merge simple wins only where a human still approves the release to production.
like real things worth cleaning up. Spin up some sub agents using Fable 5.1 and get all this work up. When you file PRs, make sure to baby sit them until they pass all the different checks and review agents that come in to leave their thoughts. Let me know when all the PRs
all 10 PRs are merged. Everyone went green on CI, Code Rabbit, Cursor, BugBot, and Macroscope when it ran. With every bot thread answered or fixed before merge. Net across the batch was 340 files. And it deleted 13K lines net, which is crazy. This is a huge cleanup
Ask for progress updates in plain words New
- What it is
- Fable 5.1 stays quiet during long runs. In one 24-minute run it made 30+ tool calls in a row with no text. You get updates back just by asking in your instructions.
- Why Theo cares
- He usually leaves and comes back when a run is done, so he doesn't mind the silence. He likes that it's a plain request rather than a hidden setting.
- How to try it
- Add to your instructions: before you start, say in a line what you're about to do.
- Ask for brief updates while it works and a short recap at the end that stands on its own.
And if you find yourself wanting more updates, you can simply say exactly that. The example they gave is, before you start, say in a line what you're about to do. Brief updates while you work help the user follow along. Close with a short recap that stands on its
Advice
Revisit your CLAUDE.md (the instruction file your agent reads). Try deleting everything, see how the model behaves, then add rules back only where needed.
been like. This is definitely one of those model releases where you should go look at your Claude MD and see what is deletable. Try deleting everything, see how it behaves, and then add parts back as it makes sense to.
Keep the default high effort, but try low and medium. They handle a surprising amount. The catch: low misses the hidden piece that makes a simple-looking task complex.
evaluations. I've honestly been surprised at how many things I can get done with low and medium. They do miss things, so if you have a task that you think is simple, but there's some tiny piece you forgot about that makes it complex, low will miss that. Medium,
Remove old instructions like "hold all findings for the final response." This model behaves differently enough that they may hide things you actually want.
it's spamming you with stuff. For example, stuff like hold all findings for the final response, it might be worth removing those lines if you have them because this model's behavior is different enough that you might end up suppressing things you actually do want.
Hot takes
Benchmarks no longer tell the story, and Opus 5 sucks.
His reasoning: Opus 5 benchmarked well and seemed fine in small test work. Once you merged its code in real projects, it was nowhere near as nice to use.
bit different. Benchmarks don't tell even close to the whole story anymore. If they did, then I would actually use Opus 5 willingly. But, Opus 5 sucks. We all hopefully understand that now. So,
because I don't want to make the mistake I made in the Opus 5 video where the model seemed totally good as you were using it in a little bit of test type work and it killed it in the benchmarks, but then when you actually started merging the code, it was nowhere near as nice to use. But considering that we
Anthropic optimizes models to always be accurate; OpenAI optimizes them to find and create new information.
His reasoning: Anthropic's models top the hallucination benchmark while OpenAI's Sol leads physics reasoning. He thinks the push to kill hallucinations makes Claude quicker to say "I don't know" and less willing to experiment.
They focus so much on killing hallucinations that it seems to result in the model being less able to experiment and try things that aren't in its weights. Because if it's not a fact,
with their models. Anthropic really wants the model to always be accurate when it gives information. OpenAI wants the model to be able to find and create new information a little more readily.
Blocking edits to earlier chat turns is annoying, and applying it only to new accounts is stupid.
His reasoning: It makes building on Claude's API much harder. The team behind the Pi agent has struggled with it for features like branching a conversation.
way harder to build on top of the Claude APIs. I know the Pi team in particular has been struggling a lot with this with things like their branching features. So, fingers crossed this gets smoothed out in the future, but for now, very annoying, and restricting it only to new accounts is stupid.
Anthropic's front-end design skill is worth using again.
His reasoning: He had mostly stopped using it because it hurt as often as it helped. With 5.1, designs made without it were much uglier, and those made with a separate "taste" skill were bad and boring.
from. Because recently, I feel like the design scale has hurt as much as it helped, and I've pretty much entirely stopped using it. But when I switch off of the design scale, it's so much uglier. It's still better
also think the ones with the taste scale were all pretty bad and boring. But I guess that design scale is really back cuz god damn, it did a great job with these. I'm actually really impressed with the front-end design capability.
How Theo thought about this
Rather than trust benchmarks, he had the model find his single best 24-hour shipping window with Fable 5 and with Sol from real PRs. He then compared those with 5.1 on time to results, merge speed and changes needed. He flags the bias: 5.1's first day against months-tested peaks.
So, I went a different angle. I had it find the best 24-hour window of Fable 5 usage and 5.6 Soul usage by going through my real pull requests across T3 Code and Lakebed, which are my two main projects I'm working on right now. So,
model is benefiting you or not. But remember, this was just my first 24 hours of Fable 5.1 against my best with these two models I've had for months.
He hands each new model the next layer of his work. He now lets 5.1 merge PRs because it is picky about what counts as ready, and dozens of merges haven't burned him. He lets it merge anything that isn't a product change he'd have opinions on, especially bug fixes and speed-ups.
found that I'm telling this model more often to merge the PR when it decides it's ready, and I've yet to be burned by that despite it doing it dozens of times for me already. 5.1 is picky enough about what it thinks is good enough that I trust it to do that, and I haven't
the model to autonomously find PRs, confirm that they are good, confirm that they're fixing things that matter, and if they're not making like meaningful product changes I might have opinions on, just let it merge them, especially for bug fixes and performance improvements. Just don't make it my
Sign in to mark what you've watched and keep notes.
Have a question about this one? Ask about this video →
Theo Daily is unofficial and not affiliated with Theo or t3.gg. Written by Claude Opus 5.5 from each video's transcript.
