Wait Which Model?
← Back to directory
Anthropic

Claude Haiku 5.5

Released Oct 7, 2026 · knowledge cutoff 2026-06

Status
Unknown
Location
United States
Type
LLM
Modality
Multimodal
Context window
1M
Max output
128K
Speed (medium effort)
137 tok/s
Price ($/MTok in / out)
$0.10 / $0.50
Cost per task (medium effort)
$0.050
Open weights
No
News1 story
Strengths & weaknesses

Strengths

  • Ten times cheaper than Haiku 4.5 per token ($0.10 / $0.50 against $1 / $5) for prompts up to 100K tokens, with a 1M-token context in place of 200K
  • First Haiku with the effort parameter, so one model can cover quick classification and routing at low effort and longer agent tasks at higher effort
  • Anthropic rates it the fastest model in its current lineup, built for high-volume, latency-sensitive work such as subagents and extraction

Weaknesses

  • Not a drop-in swap for Haiku 4.5: non-default temperature, top_p or top_k now return a 400 error, and the newer tokenizer counts the same text as about 30% more tokens
  • Prompts over 100K tokens cost five times the base rate ($0.50 input, $2.50 output per MTok), so long-context use loses most of the price advantage
  • Adaptive thinking is on by default, which adds latency before the first answer token compared with a non-thinking call

Released 2026-10-07 on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. Context 1M, max output 128K (300K on the Message Batches API beta), reliable knowledge cutoff June 2026, default effort medium with adaptive thinking on by default, per Anthropic's model overview and the Bedrock model card. Base prices apply to prompts up to 100K tokens; above that it is $0.50 input and $2.50 output per MTok. Cache reads are $0.01 and 5-minute cache writes $0.125 per MTok, and batch is 50% off. HLE 45.9% is Anthropic's no-tools figure. Anthropic's system card reports SWE-bench Pro, but its figure could not be confirmed from a primary-source excerpt this run, so sweBenchPro is null; the launch table's Terminal-Bench figure is version 4.0 and its GDPval-AA figure is v2.1, neither of which matches the tracked keys. Cost per task ($0.05) and output speed (136.8 tok/s) are Artificial Analysis' medium-effort figures; AA's time-to-first-token figures conflict between pages (12.63s vs 13.40s), so that half of speed is null. predecessorId follows Anthropic's Haiku 4.5 page, which calls Haiku 4.5 a legacy model, names Haiku 5.5 the current Haiku and advises migrating to it. ARC-AGI-2 61.3% is ARC Prize-verified at xhigh effort ($0.085/task; high 30.7%, medium 9.4%, low 2.8%), arcprize.org/results/anthropic-claude-haiku-5-5; ARC Prize left out the max-effort run because only 74 of 120 tasks completed.

Benchmarks2 of 10 reported

Humanity's Last Exam

45.9%

ARC-AGI-2

61.3%

MMLU-Proretired

—

GPQA Diamond

—

SWE-bench Verified

—

SWE-bench Pro

—

Terminal-Bench 2.1

—

AIME (latest)retired

—

LMArena Elo

—

GDPval-AA v2

—
For developers

API model strings

  • anthropicclaude-haiku-5-5
  • aws-bedrockanthropic.claude-haiku-5-5
  • google-vertexclaude-haiku-5-5

Licence

Proprietary — weights not released

Retirement

No retirement announced

Lineage

Replaces Claude Haiku 4.5