finally a good small model
Thursday, Oct 8 · 41 min · about 10 min to read
Theo-brain watched all 41 min
Haiku 5.5 is very cheap and fast below 100k tokens of context, but every price jumps 5x once it passes that. Theo won't write code with it and stays on Opus. Its value is as a cheap helper: Opus plans the work and hands Haiku the context gathering, checking and many parallel attempts.
Skim itSkim the notes for the pricing; watch the stretches with the sub-agent demo, his cost argument and the live test that went wrong.
Tap any time to jump the video there.
Key takeaways
Haiku 5.5 is very cheap below 100k tokens of context. It costs 1 cent per million cached tokens read (reusing context already sent), 10 cents per million input and 50 cents per million output. Haiku 4.5 charged $5 per million output. That makes it about 20x cheaper than Sonnet 5.5.
5.5 is a 10th the cash read price. It's 1 cent per mil tokens read. 12.5 cents per million cash token writes, 10 cents for normal 1 mil read, and 50 cents per mil out, where previously it was 5 bucks per mil out and Sonnet 5.5 was 10 bucks per mil out. This makes it 20x cheaper in output tokens and input tokens than
The catch: past 100k tokens of context, every price goes up 5x. Theo reads this as Anthropic steering Haiku toward short, quick jobs, so it doesn't lose money when people run long sessions in Claude Code.
advance here, if you go over 100k tokens, the price doesn't double, it 5x's. So, you go from 50 cents per mil out to 250 per mil out. You go from 10 cents per mil in to 50 cents per mil in,
code. This is a way of steering people towards the intended use case by making it really, really cheap if they use it on short, quick response things. But
Theo's core point is that writing the code is cheap. The expensive parts are gathering context beforehand and verifying afterward. A cheap model's value is doing those parts for the smart model, not writing the code itself.
I'm trying to say here is I don't get why people are looking for a dumb model to do the code part. The code is cheap. The expensive part is all of the prep before you write the code file and all the verification after. And Haiku is a tool your models can use to do those parts more cheaply is valuable. And if
On pure value per dollar at API prices, his stack is GPT-6 Luna, GPT-6.1 Sol and Opus 5.5. Haiku is now neck and neck, and still worth it if you want to stay with one provider, already pay for Claude Code, or want models that work well with Opus.
So, if you're trying to get the absolute most value and you're paying API prices, 6 Luna 61 Soul Opus gets you pretty much all of that. That said, Haiku 55 is neck and neck now. Are Soul and Luna still better per dollar? Slightly, usually.
However, if you are trying to do everything through one provider, like you want to just use your Anthropic inference cuz you have a crazy spend target with them and you just use their APIs, or you're subscribed to Claude Code and don't want to also subscribe to Codex, or you want models that
Claude subscribers now get monthly API credit for building their own apps: $200 on the $200 plan (20x) and $100 on the 5x plan. Theo calls it genuinely nice of Anthropic.
really cool thing, where if you're subscribed as a 20x user, so you're on the $200 plan, you get $200 in API credit every month. 5x users get $100 in credit. And this is for the API. So
Two skills here are new to you.
Opus as boss, Haiku 5.5 as sub-agents
- What it is
- A smart model (Opus) plans the work and hands many small jobs to cheap Haiku helpers running in parallel. In Anthropic's demo, Opus alone made 25 attempts for $0.47. Opus with 10 Haiku helpers made 86 attempts in under a minute for $0.14.
- Why Theo cares
- Haiku is cheap and fast enough to run many attempts at once without huge bills. His own rerun of a PR audit used exactly this setup.
- How to try it
- Run the main task on Opus and tell it to use lots of sub-agents, with Haiku 5.5 for all of them.
- Tell it to do little itself beyond getting the data and directing the sub-agents.
- Remove any rule in your CLAUDE.md that forces all sub-agents to be Opus.
only designed 25 attempts and it cost $0.47 to do. But with Opus 5.5 commanding 10 Haiku instances, it was able to do it in under a minute with 86 attempts. Way more attempts because it's able to get more renditions out when it's that much cheaper and faster and only cost $0.14. So, for the types of
First, I told it to use lots of sub-agents and workflows to review these PRs and to use Haiku 55 for all sub-agents throughout the work. To say it in one more sentence, minimize work that you do yourself outside of actually getting the data and orchestrating the
Cap Haiku's context at 100k New
- What it is
- An Anthropic employee's tip: set Haiku 5.5's auto-compact window to 100k in Claude Code. The session then summarizes itself before passing the price threshold. The setting applies only to Haiku, sub-agents included.
- Why Theo cares
- It guarantees Haiku sub-agents never get expensive. The trade-off showed in his game test: 47 of 51 requests went over 100k. A cap would have made it cheaper but probably worse.
- How to try it
- Ask your agent to set Haiku 5.5's auto-compact window to 100k in Claude Code.
- Use it when you pay API prices, not just a subscription.
come out publicly and said, "If you want to keep your bill cheap and not worry about this, you can manually set Haiku 5.5's auto compact window to 100k so that you're always working within that cheaper token pricing. Apparently it's a save per model, so this will only apply to Haiku including sub-agents. So if you
requests exceeded 100k input tokens, so they used Haiku's higher price tier. If I had capped the context, the game probably would have come out much worse, but it would have been way cheaper. Look
Opus writes, GPT-6.1 Sol audits
- What it is
- His agents use Opus for coding. Before opening a pull request (a proposed code change), they ask OpenAI's GPT-6.1 Sol for a review.
- Why Theo cares
- Sol is a different model family with a different view, like asking a friend at another company. He says it catches the dumb things Opus misses, so he lands code much faster.
- How to try it
- Have your agent set up a skill that sends finished work to GPT-6.1 Sol for review before any PR is opened.
- Keep Opus doing the actual coding.
Soul as an auditor. I have set up skills for all my agents where they use Opus as the thing that does the coding, and then they consult with Soul for opinions before they put up the PR, and I'm landing code so much faster simply because Soul catches all the dumb things Opus might miss. And by the time the PR
Claude's design skill New
- What it is
- A design skill for Claude models that improves front-end and visual output. With it off, Haiku's demo sites went back to generic AI-looking design.
- Why Theo cares
- He had uninstalled it and now plans to bring it back. He says it helps a lot right now.
- How to try it
- Ask your agent to install the Claude design skill before front-end work.
- Compare a page built with and without it.
Yeah, now now we're back into classic LM design. It seems like the new Claude design skill helps their models a shitload right now. Yeah, the gap to that to this is hilarious. I had uninstalled the design skill. I definitely need to bring it back cuz it is it is much better now than it was
Advice
Don't pick Haiku as your coding model. He stays on Opus with his $200 plan. If Opus solves a problem in 100k tokens and Haiku burns a million before failing, Haiku ends up costing more.
well, but I would not select this model as a tool I use in my coding agents, because I'm on the $200 tier, and I would rather just use Opus. And as big
using more money. Cuz if Opus can solve a problem in 100k tokens, and Haiku burns a million before it figures out it can't, Haiku ends up more expensive. So,
Use Haiku when the smart model is checking a theory, comparing against several open-source projects, or reading lots of documentation. That is where it shines.
confirm it or wants to go look at eight apps in the world and like random open-source projects to compare against or read through a bunch of documentation to figure out some facts about some idea you have. That is when Haiku is incredible. But the execution of the
Don't run Sonnet 5.5 on Max reasoning. It costs $7.60 per task versus $2.75 on X-High, and Theo says it makes no sense at all.
fine value at $2.75 per task. But, then Sonnet 5.5 Max makes literally no sense at $7.60 per task. Do not if you're using Max reasoning levels, I have questions. But, I also will say very confidently, do not use Sonnet 5.5 on Max, it makes no sense
Hot takes
You can skip Sonnet entirely and use just Opus and Haiku from Anthropic.
His reasoning: On his value-per-dollar chart, Sonnet now adds almost nothing, maybe just barely on high reasoning.
whole lot. Now, Sonnet, I showed you guys Sonnet just doesn't offer much with the Pareto line. Actually, it offers nothing now. Maybe just barely on high there, interesting. But, yeah, I I think you can avoid Sonnet entirely and just use Opus and Haiku and have a very, very good time with your Anthropic models.
Small, dumb models corner themselves and make things harder not just for themselves but for you.
His reasoning: In his test, Haiku said it would wait for the GitHub rate limit to reset but never triggered a wait, so the thread simply stopped. It ruined his T3 Code experience for at least the next 50 minutes.
Except for the fact that it trigger anything to spin itself back up. It said that it's going to wait until the rate limits up, but it didn't trigger a wait. We would show that in the UI if it did. So, this thread is just done until I
Small models will corner themselves like this, make something harder for not just themselves, but for me. Like, this just ruined my T3 code experience for at least the next 50 minutes until my quota
Anthropic is dominating, and OpenAI should be scared.
His reasoning: He struggles to use his Codex subscriptions because his Claude subscriptions are so much more useful.
line. Anthropic is dominating across the field now, and if I was OpenAI, I would be very, very scared, and I would be praying that the next training run for Astra comes out unbelievable. Because right now I'm honestly struggling a bit to use my Codex subs because my Anthropic and Claude subs are just so
How Theo thought about this
When ThePrimeagen reported Haiku doing worse and slower than GPT-6 Luna, Theo checked the setup before accepting the result. Prime had turned reasoning off, which Anthropic trains its models to be bad at. With reasoning on, Theo saw about 180 tokens per second.
that he did here is he turned off reasoning. He has both 6 Luna and Haiku 5.5 on no reasoning. OpenAI has trained their models to be good at that. Anthropic has trained their models to be bad at that. With
is worth, when I tried out Haiku 55 on my machines with Claude code and my normal Claude code sub, I was seeing around 180 tokens per second for my just traditional prompting working on things
After Haiku's PR audit stalled, he reran the same prompt on Opus. His reasoning: Opus is a smart model, so it should avoid his rate limits and find ways around them. Opus then planned the job itself, about 15 PRs per Haiku agent across roughly 100 agents.
I'm going to run this on Opus, because unlike Haiku, Opus is a smart model, and it should not only be able to not hit my rate limits, it should find ways to work around them, too. Here is where it's
tackle with sub-agents. Maybe batches of around 15 PRs per Haiku agent will mean roughly 100 agents running. Beyond the typical small-scale workflow guidelines, but the user explicitly wants, quote, lots of sub-agents. Look at that. It's
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.
