Haiku 5.5 reads. A bigger model decides.

Haiku 5.5 is out. It costs $0.10 per million input tokens and $0.50 per million output tokens on prompts up to 100,000 tokens. Sonnet 5.5 costs $2 and $10. That is a twentieth of the price per token.

The last Aftershock promised a rerun of the eight known-answer tasks when these models landed. Here it is. Haiku scored 73 out of 80, Sonnet 78.

Here is my take. The smallest model can now do the reading in your pipeline: pulling numbers out of files, summarising, counting, auditing one document against another. Keep the deciding with a bigger model, and above all the decision to delegate. The price gap is smaller in practice than the price list says.

Five ties or wins out of eight

The tasks come in four kinds, two of each, over a folder of my own working notes: pulling numbers out of files, summarising rules, finding contradictions and counting things, and judgment calls. The answer keys were written on 8 October, before any run started. Every answer got a score from 0 to 10.

Haiku tied on all four extraction and summary tasks, 10 against 10. On the audit task it won. It found ten places where a playbook, a checklist and a script disagreed with each other or with the repo. Checked against the files, all ten were true. Sonnet found the main ones and missed a real gap, a missing clean-build step.

Sonnet won the counting task, 10 to 9. Both got the numbers right: 36 debts, 43 modules, 18 checklists. Haiku hedged, and it also wrote a scratch file after being told the task was read-only.

On the two judgment tasks, Sonnet won 9 to 8, then 10 to 6. In the answers I checked I found nothing invented, in either set.

The gap is a decision

Haiku didn’t misread anything. It chose which of two sources to trust.

The last task asked this. The parent chat sits at 80,000 tokens of context. It needs five file reads of about 4,000 tokens each, then makes about 18 more calls. Read in the chat, delegate to a Sonnet worker, or to an Opus worker?

My key said Sonnet worker. Sonnet said Sonnet worker, and it named the hole I had marked in my own key: none of my cost tables covers 80,000 tokens of context and five reads together. Haiku said read in the chat. It followed a rule of thumb from my notes, delegate only above about 150,000 tokens. In the same answer, its own estimate had the Sonnet worker cheaper, at 0.9 of the cost, and it called that inside the margin of error.

That is the call I don’t hand over. A wrong delegation call doesn’t look wrong in the answer. It shows up weeks later, on the invoice.

What moves in the routing table

Anthropic’s own wording is close. Haiku 5.5 “pairs well with Opus 5.5 and Sonnet 5.5 as a subagent”, and it is “best suited to more narrowly scoped tasks”. The test says the same, with a boundary. Three rows, updated:

1. Reading, extracting, counting, auditing one document against another: Haiku 5.5. Make it return the file and line for every claim, and check a sample.
2. Well-specified work and summaries that feed a decision: Sonnet 5.5.
3. Judgment, including whether to delegate at all: Sonnet 5.5 or Opus 5.5.

This is the same rule as before. Give the cheaper model the work where a miss shows, and file and line numbers make a miss show.

Where the cost-cutters are right

The bill falls by less than the price list suggests. Four reasons.

The 16 runs used between 77,000 and 120,000 tokens each. Most of that looks like context every subagent carries before it reads a line, though I didn’t separate it. Across all eight tasks Haiku used 734,235 tokens and Sonnet 667,147, about 10% more. Anthropic also notes an updated tokenizer that uses slightly more tokens per task.

On cached input the gap is 10 times, not 20: $0.01 against $0.10 per million tokens. A subagent that reuses a big prefix lives on cached input.

Above 100,000 tokens in the prompt, Haiku 5.5 costs $0.50 and $2.50. Against Sonnet 5.5, that is 4 times cheaper, not 20. A subagent that loads a big folder gets there fast.

Haiku was slower on five of the eight tasks. It finished all eight in 414 seconds against 474, but only because Sonnet needed 287 seconds on the audit task. Without that one, Haiku needed 269 seconds and Sonnet 187.

I measured tokens and seconds, not dollars.

What this test can’t tell you

It is one run per model per task, so a one-point gap is noise. The scorer was Sonnet 5.5, one of the two contestants, and it knew which answer was whose. That favours Sonnet, and Haiku still tied or won five tasks. The prompts are a reconstruction, because my September files were lost, so these scores don’t line up with September’s. And I didn’t run Opus 5.5 this round.

Run your own. It takes an afternoon.

1. List the subagent tasks that only read, extract, count or compare.
2. Pick five and write their answers yourself, before any run.
3. Run each on Haiku 5.5 and on your current default, with the same prompt.
4. Require file and line for every claim. Check a sample against the files.
5. Score 0 to 10 against your key. Move a row only where Haiku lands within one point.
6. Keep the decision to delegate on the bigger model.
7. Check your prompt sizes. Above 100,000 tokens the price jumps.
8. Write down the date, and rerun when the next model lands.

My eight tasks are saved with their keys. When the next small model ships, I run them the same day.