claude-haiku-5.5
Meant to excel at
-
Built for high-volume, latency-sensitive work: summaries, compaction, database queries, classification, extraction and live customer support.
Vendor's own site anthropic.com -
Intended as a cheap subagent alongside Opus 5.5 or Sonnet 5.5, with browser and computer-use beta support.
Vendor's own site anthropic.com -
First Haiku with adjustable effort and a 1M context window; $0.10 and $0.50 per million tokens up to 100K prompt.
Independent evaluator gradually.ai -
Matched Haiku 4.5 on a 24-invoice approval test with under a quarter of the output tokens, about 25x cheaper.
Independent evaluator datacamp.com -
Passed all 140 scoped subagent runs on a small repo at low and medium effort, far cheaper than Opus.
Independent evaluator dev.to -
Passed 37 of 43 business analytics questions at about $0.38 per run and the lowest median latency (14s) in Plotly's test.
Practitioner report plotly.com
Watch out for
-
Well behind Sonnet 5.5 on agentic coding (Terminal-Bench 4.0: 39.2 versus 70.6 percent); Anthropic steers complex coding elsewhere.
Vendor's own site anthropic.com -
Whole-request price jumps fivefold above 100K tokens, and the new tokenizer uses roughly 30 percent more tokens.
Vendor claim, relayed gradually.ai -
Anthropic warns xhigh effort can put the answer in thinking, returning no visible reply.
Vendor claim, relayed syntaxdispatch.com -
The haiku alias still maps to Haiku 4.5 on Bedrock, Vertex and Foundry, so pin the full model ID.
Practitioner report dev.to -
Found only 2 of 5 hidden-data-issue questions; GPT-6 Luna scored slightly higher (38/43) for less.
Practitioner report plotly.com -
Ranked 38th on Vals' Excel Modeling Benchmark v1 at 59.0 percent (about $1.15 per test) versus about 76 percent for Fable 5.1, Opus 5.5 and Sonnet 5.5.
Independent evaluator vals.ai -
Artificial Analysis shows it is very verbose (about 440M output tokens on the index at max vs 99M median) despite 232 tokens per second output.
Independent evaluator artificialanalysis.ai
Benchmarks
Agentic coding percentile 0.46
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| swe-bench pro | 64.8 | 0.67 | Vendor claim, relayed computingforgeeks.com | 2026-10-07 |
| terminal-bench 4.0 | 33 | 0.25 | Independent evaluator computingforgeeks.com | undated |
Knowledge percentile 0.23
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| humanity's last exam | 45.9 | 0.46 | Vendor claim, relayed computingforgeeks.com | 2026-10-07 |
| vals excel modeling benchmark (emb v1.0) | 59.02 | 0.00 | Independent evaluator vals.ai | 2026-10-18 |
Overall percentile 0.23
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| artificial analysis intelligence index | 43 | 0.22 | Independent evaluator deeplearning.ai | 2026-10-16 |
| gdpval-aa v2.1 | 1620 | 0.25 | Vendor claim, relayed apidog.com | 2026-10-07 |
Vision percentile 0.00
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| chartography | 46.4 | 0.00 | Vendor claim, relayed apidog.com | 2026-10-07 |
Percentile is the share of tracked models this model beats on that benchmark, averaged across the group. Only benchmarks at least four tracked models share are scored.
Practitioner reports
- +/-
Another guide says it fits read-only subagent work (codebase search, CI/log triage, diff and thread summaries) but should stay off multi-file edits, hypothesis-driven debugging and anything going straight into a commit.
Task: research | Compared with claude-sonnet-5.5 | 2026-10-14
Practitioner report prompteddev.com - +/-
Plotly's 43-question analytics test had Haiku 5.5 at 37/43 for about $0.38 a run, fastest at 14s median, but behind GPT-6 Luna (38) and Opus 5.5 (41) and weak on needle-in-haystack questions. Single run per model, so one or two questions is noise.
Task: data-analysis | Compared with gpt-6-luna | 2026-10-14
Practitioner report plotly.com - +/-
On a small repo (tinydb, 226 tests) across seven routine subagent tasks, Haiku 5.5 at low and medium effort passed all 140 runs, same as Sonnet 5.5 and Opus 5.5, at roughly 40x lower cost for codebase exploration; author says it proves little beyond short tasks on one repo.
Task: chore | Compared with claude-opus-5.5 | 2026-10-12
Practitioner report dev.to - +
Routing seven routine subagent tasks to Haiku 5.5 passed every run at about 40x lower list cost on exploration, with no speed penalty.
Compared with claude-opus-5.5 | 2026-10-12
Practitioner report dev.to - +
As an Explore subagent it located a re-exported function definition correctly in all three runs, for roughly half a cent each.
Compared with claude-sonnet-5.5 | 2026-10-09
Practitioner report developersdigest.tech - +/-
Author advises keeping Haiku off subagents that edit code, plan or review; in five headless runs only the Explore subagent pinned to Haiku found the right file, and each run cost a few tenths of a cent. One easy task per config.
Compared with claude-sonnet-5.5 | 2026-10-09
Practitioner report developersdigest.tech - +/-
Quality depends heavily on effort setting: Cursor saw CursorBench rise from 30.9% at low to 48.4% at max, and Anthropic's GDPval-AA drops from 1620 at max to 1277 at the medium default.
Task: feature | 2026-10-08
Practitioner report apidog.com - +
GitHub's early testing, cited at launch, found it matched Sonnet 5 on many coding tasks with fewer tokens and steps.
Task: feature | Compared with claude-sonnet-5 | 2026-10-07
Vendor claim, relayed prompteddev.com - +
All 24 purchase-order invoice pay, hold or return decisions were correct, faster and about 25x cheaper than Haiku 4.5, though the test was too easy.
Compared with claude-haiku-4.5 | 2026-10-07
Practitioner report datacamp.com - -
Bito's six-task coding blend (Opus 5.5 as judge) scored Opus medium 85%, Sonnet high 82%, Haiku high 79%, Haiku medium 76%; the gap to Opus widened to 8-10 points on implement-and-audit turns, and a lookup task fell from 84% (high) to 59% (medium).
Task: feature | Compared with claude-opus-5.5 undated
Practitioner report bito.ai