Building with Claude Sonnet 5.5 and handing the small repeat jobs to Claude Haiku 5.5, instead of defaulting to Claude Opus 5.5, has made my usage go a lot further this month, and the work has held up. This post covers what I changed, what Anthropic’s published prices say about why, what Sonnet 5.5 did on the Annoying Dino 2.1 update, and the parts I have not measured. The tone is happy because the result is good. The numbers are cautious because I only have one project’s worth of them.
How do I split the work between the three models?
I follow the split Anthropic describes. Its Haiku 5.5 announcement says Sonnet 5.5 and Opus 5.5 remain the better choice for complex agentic coding, and that Haiku 5.5 suits high-volume, cost-sensitive work and use as a subagent. So Sonnet 5.5 does the building, the debugging and the writing, Haiku 5.5 takes the small jobs that repeat, and Opus 5.5 stays in reserve for a problem that actually needs it. I wrote up the published figures for the newer two in what Sonnet 5.5 is and how Haiku 5.5 compares, so I will not repeat the benchmark tables here.
What do the published prices say?
| Per million tokens | Opus 5.5 | Sonnet 5.5 | Haiku 5.5 (prompt up to 100K) | Haiku 5.5 (prompt over 100K) |
|---|---|---|---|---|
| Input | $4 | $2 | $0.10 | $0.50 |
| Output | $20 | $10 | $0.50 | $2.50 |
| Cache hit | $0.20 | $0.10 | $0.01 | $0.05 |
Read straight off that table, Sonnet 5.5 costs half of Opus 5.5 on input, output and cache hits. Haiku 5.5 costs 40 times less than Opus 5.5 on input and output for prompts up to 100,000 tokens, and 8 times less above that. Those are ratios of list prices, nothing more. They say nothing about how many tokens a given task needs on each model, which is the number that decides what you actually spend.
Why don’t the list prices tell my story?
Because I do not pay them. I use Claude through a subscription, so the saving shows up as how far my plan’s usage stretches, not as a dollar figure. It stretches further. I just did not meter it, so I have no percentage to give you and I am not going to invent one. If you pay per token, you can measure the same thing properly: run one task on each model and compare tokens and cost.
What did Sonnet 5.5 do on the Annoying Dino 2.1 update?
Quite a lot, in a single working session. It built the throw-and-fetch behavior, the toys, the daily egg and four Halloween outfits, and it took the wardrobe from 33 items to 37. It ran my regression, simulation, play, store-gate, real-browser and popup suites, 154 checks across the six, against a package that started the day failing most of them. It broke the code on purpose in 14 places to prove the tests could go red. When I reported two bugs this morning, a card that would not flip when tapped and a species picker wrapping onto two lines, it found the cause of each, fixed them, re-ran every suite and handed me a new package. The full bug list is in the 2.1 devlog.
Which of the newer tools made that possible?
The model matters less than what it can reach. In this update it worked with a shell and file tools, a real headless Chromium that loads the unpacked extension and clicks with a real mouse, a visible task list, independent reviewer passes that did not see the work being written, and the WordPress.com connector that published these posts. A cheaper model that can run the browser test and read the failure is more useful to me than an expensive one that has to guess.
Where does Haiku 5.5 fit?
In the jobs Anthropic points it at: high volume, low risk, and quick. I would not hand it a hard debugging problem. Anthropic’s own figures put Haiku 5.5 at 39.2% on Terminal-Bench 4.0 and Sonnet 5.5 at 70.6%, and I take that gap seriously. What I like is that a tier exists where a small job can cost a fraction of a cent’s worth of list price and still be done well enough, which leaves more of my usage for the work that needs Sonnet.
What have I not measured?
- I kept no token counts, so I cannot say how much less I used.
- I did not run the same task on Opus 5.5 and Sonnet 5.5 side by side.
- It is one update to one extension, so it is an example, not a benchmark.
- Quality “held up” means my tests passed and I found no defect I could blame on the model. It does not mean Opus 5.5 would have done worse or better.
When I next change an extension I will record tokens or usage for the run, so the next version of this post has a number in it.
Which of my extensions are getting updates?
Annoying Dino is the one I can point to for this approach. Alongside it, Tantrum is my Chrome extension that turns any page into a toy you can wreck, mend, overgrow or doodle on, and it is on the Chrome Web Store. Dota Companion is the Dota 2 helper that moved to Chrome’s side panel in version 2.5 and is on the Chrome Web Store too. Both are worth a look if you came here for the extensions and not the model comparison.
Questions people ask
Is Claude Sonnet 5.5 cheaper than Claude Opus 5.5?
Per token, yes. Anthropic’s pricing page, read October 13, 2026, lists Sonnet 5.5 at $2 input and $10 output per million tokens, and Opus 5.5 at $4 and $20, so Sonnet is half the price. Total cost still depends on how many tokens each model uses for your task.
How much cheaper is Claude Haiku 5.5 than Opus 5.5?
For prompts up to 100,000 tokens, 40 times cheaper on list price ($0.10 against $4 input, $0.50 against $20 output per million tokens). Above 100,000 tokens it is 8 times cheaper. Anthropic recommends Haiku 5.5 for high-volume work, not for the hardest coding.
Do you pay API prices for this work?
No. I use a Claude subscription, so my saving appears as longer-lasting usage, not as a lower bill. I did not meter it, so I give no percentage.
Is Sonnet 5.5 good enough to replace Opus 5.5 for building extensions?
For the Annoying Dino 2.1 update it was, by my tests. That is one project and I did not run Opus 5.5 on it, so treat it as an example. Anthropic lists both Sonnet 5.5 and Opus 5.5 as suited to complex agentic coding.
Sources
- Pricing, Claude Platform Docs (prices above; read October 13, 2026)
- Introducing Claude Haiku 5.5, Anthropic (use-case guidance and Terminal-Bench figures, as relayed in my Haiku 5.5 comparison)
- Introducing Claude Sonnet 5.5, Anthropic (70.6% Terminal-Bench 4.0)
Leave a Reply