- Claude Haiku 5.5 launched October 7, 2026 at $0.10 input and $0.50 output per million tokens for prompts up to 100k.
- Anthropic says it is about 75% cheaper on average than Haiku 4.5 and 90% cheaper for requests under 100k tokens.
- It scores 72.4% on OSWorld 2.1 (offline subset), up from 15.7% for Haiku 4.5.
- Sonnet 5.5 and Opus 5.5 remain the better choice for complex agentic coding.
About 90% of Haiku 4.5 requests fell under 100k tokens, and that single statistic explains why Anthropic’s newest small model matters. For most teams running high-volume workloads, the new price tier applies to nearly everything they send. Claude Haiku 5.5 arrived on October 7 with a price that makes the previous generation look expensive, and a benchmark table that tells a more nuanced story about what you trade away.
The Price Drop in Numbers
Haiku 5.5 vs Haiku 4.5 vs Sonnet 5.5
Per million tokens, Haiku 5.5 charges $0.10 for input and $0.50 for output on prompts up to 100k tokens, and $0.50 and $2.50 above that. Haiku 4.5 charged $1.00 and $5.00, and Sonnet 5.5 charges $2.00 and $10.00. Cache reads drop to $0.01 per million tokens, against $0.10 for Haiku 4.5. Anthropic also halved Sonnet 5.5 cache read pricing from $0.20 to $0.10, which it says trims costs by about 20% on most agentic work.
Versus Haiku 4.5, the model is 90% cheaper for requests up to 100k tokens and 50% cheaper above that. The model ID is claude-haiku-5-5, and it is available on all platforms including Amazon Web Services, Google Cloud, and Microsoft Azure. Anthropic calls it “the cheapest, fastest, and most capable small model we’ve ever released.”
Trend Insight — A 10x price cut on the small tier changes which tasks are worth automating at all. Classification, triage and tool-calling loops that were too costly to run on every request now become default behavior.
What You Give Up
The benchmark gap to Sonnet 5.5
The catch is capability at the top end. On Terminal-Bench 4.0, Haiku 5.5 scores 39.2% against 70.6% for Sonnet 5.5. On OSWorld 2.1 it reaches 72.4% versus 83.9%, and on Humanity’s Last Exam without tools it scores 45.9% versus 56.9%. It still beats GPT-6 Luna on the published comparisons, including 72.4% versus 48.9% on OSWorld 2.1 and 39.2% versus 16.4% on Terminal-Bench 4.0.
Safety settings are also tuned differently. Cybersecurity safeguards are stricter than Haiku 4.5 but looser than Sonnet 5.5, and they still block penetration testing and similar attacker techniques. Anthropic did not publish context window size or maximum output tokens on the launch page.
Trend Insight — The sensible pattern is routing: send the bulk of simple calls to Haiku 5.5 and escalate hard coding or long-horizon agent tasks to Sonnet or Opus.
Early Customer Results and New Controls
Speed, effort levels and agent SDKs
Asana reports over a 30% reduction in latency for task completions and up to 2.5x faster inference per agent turn. HubSpot recorded 92.8% on its CRM simulation suite, its best score on that suite. AlphaSense measured 0.84 against 0.76 for Haiku 4.5 across 400 queries, and Box saw an 11-point gain at about half the latency.
Haiku 5.5 is the first Haiku-class model with an adjustable effort setting, with levels from Low to Max, so teams can trade cost against intelligence per request. The Python and TypeScript SDKs also add beta support for computer use and browser use. Max 5x subscribers get $100 per month in API credits, Max 20x get $200, and Team plans get up to $500 pooled.
Trend Insight — Effort controls on a small model mean one model ID can cover both cheap and demanding calls, which simplifies routing code.
Related
- Satya Nadella Just Said the Quiet Part About AI
- Anthropic Just Pulled the Plug on Its Own AI
- OpenAI’s $70 Billion Number Just Lost $20 Billion
- This $0.14 Model Almost Matches GPT-5.4
Sources
AI Biz Insider · AI Trends EN · aibizinsider.com
답글 남기기