Claude Haiku 5.5: 90% Price Cut, 1M Context, Adaptive Effort

Anthropic released Claude Haiku 5.5 on October 7, 2026, cutting its list price by 90% for prompts up to 100,000 tokens, expanding the context window to 1 million tokens, and adding an adaptive effort setting — the first Haiku-class model with adjustable reasoning depth. Workloads previously capped by Haiku 4.5's 200K-token window can now ingest full corpora without moving to a larger tier, and the same model that handles classification at $0.10 per million input tokens can run a heavier reasoning pass when a task demands it.

What Actually Changed

Anthropic reports the context window grew from 200K to 1M tokens and maximum output doubled from 64K to 128K tokens. The smallest cacheable prompt dropped from 4,096 to 512 tokens, which makes short, repeated prompts eligible for caching. The API model ID is claude-haiku-5-5.

The vendor-reported benchmark gains are sharp. On GDPval-AA v2.1, Haiku 5.5 scores 1,620 versus 735 for Haiku 4.5. OSWorld 2.1 offline subset moves from 15.7% to 72.4%. Humanity's Last Exam without tools rises from 10.2% to 45.9%, and Terminal-Bench 4.0 goes from 0.0% to 39.2%. Anthropic still shows Sonnet 5.5 ahead on all four benchmarks — 70.6% on Terminal-Bench 4.0 — and says complex agentic coding remains better suited to larger models.

Customer-reported figures relayed through Anthropic's launch page point the same direction: Asana reports over 30% lower task latency and up to 2.5x faster inference per agent turn; HubSpot reports 92.8% on its CRM evaluation suite averaged over three runs; AlphaSense reports a 0.84 versus 0.76 score on a 400-query test at roughly 8 million API calls per week; Box reports an 11-point gain at about half the latency.

Pricing and Competitive Position

Anthropic prices Haiku 5.5 at $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens, and $0.50 input and $2.50 output above that. Haiku 4.5 listed at $1.00 input and $5.00 output. The average real-world saving is about 75%, not the headline 90%, because Haiku 5.5's updated tokenizer — similar to Sonnet 5.5's and Opus 5.5's — consumes slightly more tokens per task, and because roughly 10% of requests exceed the 100K threshold.

As 1M context is increasingly a pricing problem rather than a capability one, the dual-tier structure matters:

Competitor prices are AI Mastery's pricing data as of 2026-09-14. Haiku 5.5 prices are Anthropic-published.

At the sub-100K tier, Haiku 5.5 input pricing undercuts every listed competitor. DeepSeek V4 Flash offers a lower output price at $0.28 per million tokens, but its input price is $0.14.

Breaking Changes

KDnuggets reports three Haiku 4.5 API behaviors now return errors: sending temperature, top_p, or top_k; ending messages with an assistant prefill; and manual thinking with budget_tokens. Two behavior shifts also require code changes: adaptive thinking is on by default, so responses may begin with a thinking block that must be read by block type, and thinking tokens count toward max_tokens. The safety classifier can now return stop_reason: "refusal", which wrappers must handle explicitly.

Anthropic says Haiku 5.5's cybersecurity safeguards are stricter than Haiku 4.5's and block penetration testing, though somewhat less restrictive than those for other recent models. Biology safeguards match Sonnet 5, Sonnet 5.5, and Opus 5.

AI Mastery analysis

The jump from $0.10 to $0.50 per million input tokens at the 100K boundary is a 5x increase, not a gentle gradient. A prompt that crosses 100K pays five times the input rate, so token efficiency becomes a first-order concern: compression, chunking, and prompt caching at the new 512-token minimum all keep workloads in the cheaper tier.

Adaptive effort is the more architecturally significant change. Teams can run one endpoint at low effort for routing and classification, then raise effort for edge cases without a model swap or a new endpoint — consistent with constrained inference beating brute-force generation on cost and speed predictability.

The Terminal-Bench gap is the sharpest tradeoff: Haiku 5.5's 39.2% versus Sonnet 5.5's 70.6%, a 31.4-point difference on complex agentic coding. Cognition's reported use of Haiku 5.5 as the sidekick in Devin Fusion — with a FrontierCode score of 66.2 and Opus 5.5 as lead — shows the practical division of labor: small models handle narrow, high-frequency steps while the large model holds strategy, the pipeline architecture pattern now driving gains.

Haiku 5.5 is available on AWS, Google Cloud, and Microsoft Azure. KDnuggets reports the earliest retirement date as October 7, 2027. Alongside the launch, Anthropic halved Sonnet 5.5 cache-read prices to $0.10 per million tokens, which it says cuts Sonnet 5.5 costs on most agentic work by about 20%. Both moves point toward inference cost, not benchmark scores, as the main competitive axis.

Sources

• Primary: Introducing Claude Haiku 5.5 \ Anthropic
• Reported by: KDnuggets: Claude Haiku 5.5: What’s New and How to Use It for Fast, Cheap AI Workflows
• Competitor prices: AI Mastery token calculator, figures as of 2026-09-14