Theo-brainUnofficial. Not affiliated with Theo or t3.gg.
Sign inYou
40:14
0:00

My new favorite model.

Wednesday, Sep 23 · 40 min · about 8 min to read

Theo-brain watched all 40 min

Theo reviews Anthropic's Opus 5.5 after a day of heavy use. He says it fixes Opus 5: clearer writing, much better code, and a full day of work used only about 20% of his weekly Claude limit. Run it on medium, high or xhigh effort, not max. It is still a smaller, dumber model than Fable, so keep Fable for design work and expect occasional odd mistakes.

Skim itThe notes cover the argument; skim the video for the fish game demo, since that part is visual.

Tap any time to jump the video there.

Key takeaways

Opus 5.5 is a real fix for Opus 5. Theo calls it cheaper, less bad at writing, and much better at code and at following instructions. After a full day of use, he says it is blowing him away.

been using Opus 5.5 all day and it has been blowing me away. They made it cheaper. They made it less bad at writing. They made it way better at code and following instructions and just doing stuff every day. Is this the Fable

On the $200 Claude plan it goes a long way. A full day of pushing it hard used about 20% of his weekly limit. One job cost about 1% of his weekly: finding a bug in his own repo, updating it, opening a fix, checking it and merging it.

my weekly, except for the fact that about half that weekly usage came from me using Fable over the last 2 days. So, it's actually closer to like 20ish% of my weekly, which is crazy considering how I've been going out of my way to push this model all day. I was able tothe PR with the fixes, verify the fixes, and get it all merged for me. All of this combined was 1% of my weekly. Yeah,

Price per token dropped, but the model is not more efficient. On max effort it used almost 120,000 tokens (chunks of text it generates) per task. Opus 5 used 73k, Fable 78k and OpenAI's Astra 27k.

model is cheaper, but it's not because it's more efficient. In fact, this model is meaningfully less efficient than Opus 5. Opus 5 on Max did 73k tokens per task. Fable did 78k tokens and Opus 55 max did almost 120,000 tokens per task. For reference, GPT6 Astra did 27k tokens

Theo expects it to become the most popular coding model quickly. Opus offers zero data retention (the vendor keeps none of your data), which Fable does not. That means companies with strict rules can actually turn it on for their staff.

Fable for their employees. Opus they can absolutely enable. So this model is almost immediately going to become the most popular for coding with AI by far just from this one particular detail.

It is very strong at 3D. Its version of his fish game test, Fishlop 3D, was the first from any model that he kept playing and had to remind himself was just a demo.

gameplay loop right as well. This is the first version of Fishlop 3D I've had any model generate where I had to like remind myself it's just a demo and I have to go film. All the others are like, "Oh, cool. That looks really nice." And then I close it. This one is

Three skills here are new to you.

Picking the right effort level New

What it is
Effort, or reasoning, level is how long the model thinks before answering. Opus 5.5 offers low, medium, high, xhigh and max.
Why Theo cares
He finds low and max risky and best avoided, while medium, high and xhigh have all worked well. You can now change the level mid-session without breaking the cache, the reused earlier context that makes long sessions cheaper.
How to try it
  1. Run everyday tasks on medium.
  2. Raise it to high or xhigh for harder work.
  3. Skip low and max, and switch levels mid-session whenever you need to.
This model's interesting in this particular way. It feels like low and max are dangerous and generally best avoided, but medium, high, XH high, those all have seemed really good for my experience so far. They also report
means that you would bust cash whenever you changed your reasoning levels, which isn't great. Now you can change reasoning levels midsession without busting cash. Very nice change. Thank

Matching the model to the job New

What it is
Theo's current split: Opus 5.5 as the default for most work, Fable for front-end design, and OpenAI's models when you need a deeply thorough code review.
Why Theo cares
Fable still gets subtle design taste, animations and layout meaningfully better. Opus 5.5 also trails OpenAI's top model at combing through code to find every flaw.
How to try it
  1. Use Opus 5.5 for day-to-day coding tasks.
  2. Switch to Fable for a marketing page or a big design overhaul.
  3. For a deep audit of a codebase, try an OpenAI model.
very spiky and weird. I think Opus is probably the right default for the vast majority of stuff. But you should not be as scared to reach to Fable for front end still, especially when you're doing like a marketing page or some big design overhaul. I still think Fable is a
planning on using this model to very thoroughly review code and find every single thing wrong with it or suggest sweeping improvements to your codebase, might not be as strong at those things as OpenAI's Frontier is. But, I'm going

Dropping your anti-slop writing rules New

What it is
Theo removed the 'unslop' rules from his agent instruction files (the CLAUDE.md file the agent reads every session) to see how Opus 5.5 writes on its own.
Why Theo cares
Its plain writing was good enough that he does not plan to add the rules back. Clearer agent reports are his favorite change in this release.
How to try it
  1. Ask your agent to remove the writing-style rules from your CLAUDE.md or agents file.
  2. Work a normal session and judge whether its updates are easy to follow.
  3. Only add rules back if the writing gets worse.
it. I actually took the time to go delete my unslop stuff from my agents in CloudMD in order to see how this model just organically talks and I don't think I'm going to bother readding it cuz it's doing a great job. One last thing from

Advice

Don't throw out Fable in favor of Opus 5.5. Theo says Opus is smaller and dumber and will make more mistakes. On longer jobs it can get stuck repairing things that are fine, or wrongly panic that its context window (working memory) is about to run out.

I want to make sure we aren't getting too excited here in just outright throwing out Fable in favor of Opus because this model is smaller. It is dumber. It will make mistakes. So will Fable, just less so. And as we're doingsee Opus 5.5 get stuck repairing things it doesn't need to when you send it off on some performance work or trapping itself thinking its context window is about to close and it's going to die when it's not. Opus will have these

Hot takes

For now, the $200 Claude subscription is better value than OpenAI's Codex plan.

His reasoning: Codex plans are too easy to burn through. The options there are Astra, which is capable but spiky and too expensive, or the new Soul.

sub. It is a little bit too easy to burn through your codec plans right now. And since the only options you have are Astra, which is capable but spiky and way too expensive, or Soul, which we'll be talking about in a new video soon, this is a way better value for your 200 bucks right now. Kind of crazy that this

Opus 5.5's big lead over Fable on the Artificial Analysis index shows how bad these benchmarks are.

His reasoning: The Opus–Fable gap is about the size of the Astra–Muse Spark gap. Fable really is way better than Muse Spark, so by that logic Opus would be just as far ahead of Fable, which he finds telling of the benchmark rather than the model.

bigger. And the gap between Astra and Muse Spark is roughly the same. So if you think Fable is way better than Uspark, which it is, allegedly the gap from Fable to Opus is similarly sized in favor of Opus, which I think more so showcases how bad these benchmarks are.

Anthropic are the pre-training kings and OpenAI the post-training kings, but Anthropic caught up a lot on post-training with this release.

His reasoning: OpenAI has historically been better at squeezing a big model's ability into cheaper, smaller models.

historically, OpenAI has been better at taking the capability of the high-end expensive model and forcing it into the cheaper, smaller models. Anthropic's never been the best at that. They're the pre-training kings and OpenAI is the post-training kings. But it seems like Anthropic caught up a lot here on the

How Theo thought about this

He tested max effort instead of trusting it. On Skatebench, moving from xhigh to max raised tokens per question from about 300 to almost 5,000, over 10x, for a 1% score gain. Earlier, a max-effort run went 6.5 hours writing a plan and was only halfway done.

to max, which is just a one little bump inside of your slider, it went from 300ish tokens per task to almost 5,000 per task. I'm not exaggerating. That's over a 10x gap for that tiny little jump. And you get almost no improvement in score here at all. It's a 1% bump.earlier. I had a thread go for 6 hours and 30 minutes writing a markdown plan because I had it on max and it just ran in loops forever. I ended up having to stop it and then switch to XH high and

His own code-review benchmark showed Opus's known weakness, even though Opus 5.5 nearly doubled Opus 5's accuracy. He weighted that weakness lightly, because deep codebase reviews are something he does only every couple of weeks, mostly to test new models.

So, I will say the gap from Opus 5 to 5.5 in this bench is absolutely hilarious, nearly doubling in the accuracy and reliability of itsto be real with you guys. This is not a task I do every day. This is a task I do once every two or so weeks, mostly as a way to test new models. It is super

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.