Chat model
Claude Haiku 5.5 max · chat
max reasoning · 16 questions per request
- Accuracy, all 2,198 questions
- 57.1%55.0% to 59.1%, 95% interval; 1,254 right
- Answers kept at 99% confidence
- No answers at this thresholdnone of the 2,198 questions at 99% or more
- Cost, $ per 1,000 questionsper 1M questions
- $2.965$2,965Basis and price under Cost
- Latency
- 417.0 smedian request, 16 questions per request
How we called it
- Kind
- Chat model
- Access
- OpenRouter chat API, Anthropic's own endpoint, fallbacks off, output cap 128,000 tokens, reasoning max
- Paid as
- Provider service
- Model ID
anthropic/claude-haiku-5.5- Parameters
- not disclosed by makerpages checked: www.anthropic.com/claude-haiku-5-5, platform.claude.com/docs/en/models/overview; read 10 November 2026
- Note
- Anthropic's small chat model, added on the reading bench as a chat model
- Confidence from
- Written estimate
- Run
- 948, 10 November 2026
Settings
- Reasoning
- max
- Request format
- 16 questions per request
- Requests at once
- 4
- Temperature
- not recorded
Run record: harness seed, harness commit and prompt version
- Harness seed
- 0
- Harness commit
57f3caac4692- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 57.1% | 55.0% to 59.1% | 1,254 of 2,198 |
| Reading | 86.9% | 84.1% to 89.3% | 570 of 656 |
| Grammar | 44.4% | 41.9% to 46.8% | 684 of 1,542 |
Comparison group: RSTUVWXYZ. Compare with another setup.
Invalid answers: 784; read from broken replies: 0; empty replies asked again: 0; requests asked again after a provider error had been stored as their reply: 3 (the error is not the model's reply). Still invalid: 784 cut off at the output limit. An empty reply is asked once more; in a broken reply, any letter that is clearly readable is scored and the rest count as wrong. Invalid answers count as wrong (how).
With “Bilmiyorum” (run 950): Invalid answers: 672; read from broken replies: 0; empty replies asked again: 0. Still invalid: 672 cut off at the output limit. (how)
Accuracy by topic
English
The Text tab's 547 kept reading questions translated into English, asked with this setup's Text tab settings. Every setup: English reading; against Turkish: English and Turkish.
- Right answers
- 507 of 547, first English run (949)
- Accuracy
- 92.7%
- Non-hallucination with “I don't know”
- 7.8%: 4 “I don't know” of 51 not right (run 951)
- Answered with “I don't know”
- 96.3%: 527 of 547
- Error rate with “I don't know”
- 8.9%: 47 of 527 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- 1.2611,261at list prices
- Latency, median request, first English run
- 148.9 s; 4 requests at once
- English run
- 949
- Machine, English runs
- hosted API
- Date
- 10 November 2026
- Request format
- 16 questions per request
- Reasoning
- max
- Prompt version
5c3a4849752e- Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68
With “I don't know” (run 951): Invalid answers: 16; read from broken replies: 0; empty replies asked again: 0. Still invalid: 16 cut off at the output limit. (how)
Confidence
- Answers kept at 99% or more
- No answers at this threshold; 2,198 skipped: 1,414 with the model less than 99% sure and 784 without a valid answer
- Calibration gap
- 17.5 points
- AUROC, all questions
- 0.82 0.78 to 0.85
- AUROC, reading
- 0.89 0.85 to 0.93
- AUROC, grammar
- 0.78 0.73 to 0.82
- Brier score
- 0.117
- Accuracy of its surest answers
- most confident fifth: 99.3% right; most confident half: 97.3% right; most confident four fifths: 94.4% right
- Typical score of a right answer
- 80.0%
- Typical score of a wrong answer
- 50.0%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 417.0 s (this run)
- Latency, 90th percentile
- 551.1 s
- Requests at once
- 4
- Throughput
- 8.6 questions per minute of request time
- Request time
- 256.4 minutes, over the requests whose answers were kept
Latency: the time from sending one request to receiving its answer, measured from our computer, network and any retries within the request included. Request time: the time at least one request was in flight, over the requests whose answers were kept. It is not the run's elapsed time: it leaves out requests that failed and any pause before a run was resumed.
Cost
- Cost per 1,000 questionsper 1M questions
- $2.965$2,965
- Whole run
- $6.52
- Basis
- Estimated from published prices and recorded token usage
- Price
- $0.10 input, $0.01 cached input, $0.50 output per 1M tokens (source, 2026-11-10)
- Input tokens per question
- 541
- Output tokens per question
- 5,821