Claude Haiku 5.5

anthropic/claude-haiku-5.5-20261007

This model accepts text, image, and file inputs and provides text outputs. It has a context length of 1,000,000 tokens and supports tools and reasoning. Input pricing is $0.10 per million tokens, and output pricing is $0.50 per million tokens.

  • tool use
  • structured outputs
  • reasoning

At a glance

provideranthropic
releasedOct 7, 2026
context1M
input / M$0.100

Specifications

What the API gives you

Context window
1M tokens
Max output
128K tokens
Input price
$0.100 / 1M tokens
Output price
$0.50 / 1M tokens
Cache read
$0.010 / 1M tokens
Modalities
text, image, file → text
Knowledge cutoff
—
Released
Oct 7, 2026

Measured evidence

Public benchmark record

Design Arena Coding
1277 Elo2026-11-15
Design Arena UI
1272 Elo2026-11-15
Design Arena Website
1253 Elo2026-11-15
FACTS Grounding
57%2026-10-24evaluated 2026-10-24
FrontierMath Tiers 1-3 (Epoch AI)‡
75.1% ± 2.62026-10-07max effort, evaluated 2026-10-07
GPQA Diamond (Epoch AI)‡
89.6% ± 22026-10-09max effort, evaluated 2026-10-09
LiveBench Coding
77.9%2026-10-18
LiveBench Language
62.6%2026-10-18
LiveBench Math
86.4%2026-10-18
LiveBench Reasoning
84.6%2026-10-18
LMArena Agent‡
0.52026-11-11high effort
LMArena Agent Steerability‡
2.42026-11-11high effort
LMArena Agent Task Outcome‡
1.62026-11-11high effort
LMArena Agent Tool Hallucination‡
-2.12026-11-11high effort
LMArena Creative Writing‡
1337 Elo2026-11-11high effort
LMArena Instruction Following‡
1411 Elo2026-11-11high effort
LMArena Long Query‡
1395 Elo2026-11-11high effort
LMArena Math‡
1436 Elo2026-11-11high effort
LMArena Text‡
1421 Elo2026-11-11high effort
LMArena Text Factuality‡
1422 Elo2026-11-11high effort
LMArena WebDev‡
1593 Elo2026-11-11high effort
OTIS Mock AIME (Epoch AI)‡
98.9% ± 0.52026-10-07xhigh effort, evaluated 2026-10-07
SimpleQA Verified (Epoch AI)‡
23.8% ± 1.32026-10-09max effort, evaluated 2026-10-09
Terminal-Bench (4.1)‡
40%2026-11-15Claude Code, max effort, evaluated 2026-11-13

LMArena data licensed CC-BY-4.0. Aider (Apache-2.0) and SWE-bench (MIT) are open. Design Arena uses blind human preference. ‡ Submitted scaffold plus model result. Compare the harness and effort settings, not only the model name.

Task evidence

How broad is the support?

Each status is exact to this model version and the selected task. Official release claims remain provisional.

  • Best LLM for Competition MathEstablished evidence

    100% of intended weighted task evidence across 4 independent sources. 6 of 6 configured evaluation feeds currently have results.

  • Best LLM for Business WritingEstablished evidence

    100% of intended weighted task evidence across 2 independent sources. 4 of 4 configured evaluation feeds currently have results.

  • 100% of intended weighted task evidence across 3 independent sources. 5 of 5 configured evaluation feeds currently have results.

  • Best LLM for Marketing CopyEstablished evidence

    100% of intended weighted task evidence across 2 independent sources. 4 of 4 configured evaluation feeds currently have results.

  • Best LLM for Chat & RoleplayEstablished evidence

    100% of intended weighted task evidence across 2 independent sources. 4 of 4 configured evaluation feeds currently have results.

  • Best LLM for Creative WritingEstablished evidence

    100% of intended weighted task evidence across 2 independent sources. 4 of 4 configured evaluation feeds currently have results.

Show all 43 tasks
  • Best LLM for WritingEstablished evidence

    100% of intended weighted task evidence across 2 independent sources. 4 of 4 configured evaluation feeds currently have results.

  • 100% of intended weighted task evidence across 2 independent sources. 4 of 4 configured evaluation feeds currently have results.

  • Best LLM for Math & ReasoningEstablished evidence

    100% of intended weighted task evidence across 5 independent sources. 7 of 7 configured evaluation feeds currently have results.

  • 100% of intended weighted task evidence across 3 independent sources. 6 of 6 configured evaluation feeds currently have results.

  • Best LLM for Frontend & UIEstablished evidence

    91% of intended weighted task evidence across 3 independent sources. 6 of 6 configured evaluation feeds currently have results.

  • Best LLM for Code CompletionEstablished evidence

    79% of intended weighted task evidence across 2 independent sources. 3 of 3 configured evaluation feeds currently have results.

  • Best LLM for Autonomous AgentsEstablished evidence

    67% of intended weighted task evidence across 3 independent sources. 6 of 6 configured evaluation feeds currently have results.

  • 65% of intended weighted task evidence across 3 independent sources. 5 of 6 configured evaluation feeds currently have results.

  • Best LLM for SummarizationEstablished evidence

    65% of intended weighted task evidence across 3 independent sources. 5 of 6 configured evaluation feeds currently have results.

  • Best LLM for CodingEstablished evidence

    64% of intended weighted task evidence across 4 independent sources. 7 of 7 configured evaluation feeds currently have results.

  • 64% of intended weighted task evidence across 2 independent sources. 2 of 3 configured evaluation feeds currently have results.

  • Best LLM for Grounded RAGEstablished evidence

    62% of intended weighted task evidence across 3 independent sources. 5 of 7 configured evaluation feeds currently have results.

  • Best LLM for Fast SummariesEstablished evidence

    58% of intended weighted task evidence across 2 independent sources. 2 of 3 configured evaluation feeds currently have results.

  • Best Value LLMEstablished evidence

    58% of intended weighted task evidence across 2 independent sources. 2 of 3 configured evaluation feeds currently have results.

  • Best LLM for Agentic CodingEstablished evidence

    51% of intended weighted task evidence across 3 independent sources. 7 of 7 configured evaluation feeds currently have results.

  • Best LLM for AI AgentsEstablished evidence

    50% of intended weighted task evidence across 3 independent sources. 9 of 9 configured evaluation feeds currently have results.

  • 50% of intended weighted task evidence across 2 independent sources. 3 of 4 configured evaluation feeds currently have results.

  • 50% of intended weighted task evidence across 2 independent sources. 3 of 4 configured evaluation feeds currently have results.

  • Best LLM for TranslationEstablished evidence

    50% of intended weighted task evidence across 2 independent sources. 3 of 4 configured evaluation feeds currently have results.

  • Best LLM for Hermes AgentEstablished evidence

    48% of intended weighted task evidence across 2 independent sources. 4 of 4 configured evaluation feeds currently have results.

  • Best LLM for OpenHandsEstablished evidence

    48% of intended weighted task evidence across 2 independent sources. 4 of 4 configured evaluation feeds currently have results.

  • Best LLM for GooseEstablished evidence

    48% of intended weighted task evidence across 2 independent sources. 4 of 4 configured evaluation feeds currently have results.

  • Best LLM for OpenClawEstablished evidence

    48% of intended weighted task evidence across 2 independent sources. 4 of 4 configured evaluation feeds currently have results.

  • Best Local LLM for CodingEstablished evidence

    47% of intended weighted task evidence across 2 independent sources. 4 of 4 configured evaluation feeds currently have results.

  • 44% of intended weighted task evidence across 2 independent sources. 5 of 6 configured evaluation feeds currently have results.

  • Best Cheap LLM for High VolumeEstablished evidence

    40% of intended weighted task evidence across 2 independent sources. 2 of 3 configured evaluation feeds currently have results.

  • Cheapest Capable LLMEstablished evidence

    40% of intended weighted task evidence across 2 independent sources. 2 of 3 configured evaluation feeds currently have results.

  • 35% of intended weighted task evidence across 2 independent sources. 3 of 4 configured evaluation feeds currently have results.

  • 35% of intended weighted task evidence across 2 independent sources. 3 of 4 configured evaluation feeds currently have results.

  • 33% of intended weighted task evidence across 2 independent sources. 5 of 5 configured evaluation feeds currently have results.

  • 38% of intended weighted task evidence across 1 independent source; broader independent confirmation is still missing from Berkeley Function Calling or LMArena Search. 6 of 7 configured evaluation feeds currently have results.

  • 21% of intended weighted task evidence across 1 independent source; broader independent confirmation is still missing from Route reliability or Structured-output eval. 1 of 3 configured evaluation feeds currently have results. Official release evidence is also tracked provisionally.

  • Best LLM for Massive ContextProvisional evidence

    20% of intended weighted task evidence across 1 independent source; broader independent confirmation is still missing from LMArena Document or LongBench v2. 2 of 3 configured evaluation feeds currently have results.

  • 17% of intended weighted task evidence across 1 independent source; broader independent confirmation is still missing from Berkeley Function Calling or OSWorld Verified. 3 of 3 configured evaluation feeds currently have results. Official release evidence is also tracked provisionally.

  • 11% of intended weighted task evidence across 1 independent source; broader independent confirmation is still missing from LMArena Vision or VLMEvalKit tasks. 2 of 3 configured evaluation feeds currently have results. Official release evidence is also tracked provisionally.

  • Best LLM for Charts & DiagramsProvisional evidence

    Official release evidence found; independent confirmation is still missing from LMArena Document or LMArena Vision.

  • Best LLM for Document ParsingProvisional evidence

    Official release evidence found; independent confirmation is still missing from LMArena Document or LMArena Vision.

Official release evidence · provisional

What the launch material actually says

Verbatim claims from the model provider. These help on launch day but do not count as independent confirmation.

  • “"Ask in Document is one of our big sources of spend, doing about 8M calls a week in production. It answers very specific questions on top of one or a few documents. We ran 400 queries, and Claude Haiku 5.5 was a statistically significant…”
    Anthropic · Introducing Claude Haiku 5.5 \ Anthropic
  • “Computer use OSWorld 2.1 Computer use OSWorld 2.1 72.4% Offline subset 15.7% Offline subset 48.9% Offline subset 83.9% Offline subset”
    Anthropic · Introducing Claude Haiku 5.5 \ Anthropic
  • “"At HubSpot, we use simulated portals to evaluate new models on CRM tasks like reporting on deals. We mostly test the smaller, more efficient models, and Claude Haiku 5.5 got the best score we've seen on this suite yet, at 92.8% averaged…”
    Anthropic · Introducing Claude Haiku 5.5 \ Anthropic
  • “"The short and high-volume work is where Claude Haiku 5.5 fits for us, like quick lookups, subagents, and summaries. While a bigger model builds the deck, a Haiku 5.5 subagent goes into the 10-K and pulls the segment revenue line the dec…”
    Anthropic · Introducing Claude Haiku 5.5 \ Anthropic

Recommendation coverage

Where this model appears

Not currently included in a published recommendation. The task evidence panel above says exactly what is confirmed and what is still missing.

Catalog data via OpenRouter and models.dev. Last checked within the past 6 hours.