Anthropic Releases Claude Haiku 5.5: 1M Context, Adaptive Thinking, and Tiered Pricing

On October 7, 2026, Anthropic officially released Claude Haiku 5.5, its next-generation lightweight, high-efficiency language model. Arriving one year after Claude 4.5 Haiku, the new model combines a massively expanded context window, fine-grained reasoning control, and a tiered pricing architecture.

Image source: Artificial Analysis (@ArtificialAnlys)

In independent evaluations conducted by Artificial Analysis, Claude Haiku 5.5 demonstrated a substantial performance leap over previous generations, significantly outpacing existing compact models across frontier knowledge tasks and terminal manipulation.

Adaptive Thinking and 1M Token Context Window

The defining architectural advancement in Claude Haiku 5.5 is the introduction of 'Effort settings' and 'Adaptive thinking', marking the first time these reasoning features have appeared in the Haiku product tier.

Rather than remaining confined to standard single-pass generation, the model can adaptively scale its reasoning depth based on task complexity or follow user-configured effort levels.

• 1 Million (1M) Token Context: Expanded fivefold from the 200k token ceiling of Claude 4.5 Haiku to a full 1M token window.
• Multimodal Input Support: Handles massive text documents alongside high-resolution image inputs within a single unified context.
• Dynamic Reasoning Controls: Inherits the reasoning mechanisms previously exclusive to the Sonnet and Opus tiers, enhancing structured handling of complex multi-step workflows.

Benchmark Analysis: Intelligence Index 43 and Agent Performance

According to benchmark data published by Artificial Analysis, Claude Haiku 5.5 achieved a composite score of 43 on the Intelligence Index. This represents a 26-point increase over the prior Haiku generation and surpasses key peers at max effort, including GLM-5.3 Flash (42), Gemini 3.8 Flash (41), and GPT-6 Luna (38).

• AA-Briefcase Knowledge Work: Reached 1578 Elo on the private frontier knowledge benchmark AA-Briefcase, placing within the confidence intervals of GPT-6 Astra (max) and Claude Fable 5.1 (high).
• Terminal-Bench 4.0 Manipulation: Scored 33% on terminal interaction tasks, a notable jump from the 0% baseline of Haiku 4.5.
• Output Token Consumption Profile: Under max effort settings, Haiku 5.5 consumed approximately 162k output tokens per task—exceeding Opus 5.5 (max) and generating roughly triple the token volume of GPT-6 Luna (max, ~50k).

Tiered Pricing Structure and Production Considerations

Alongside the release, Anthropic introduced a tiered pricing model that adjusts rates based on prompt length.

For prompts up to 100k tokens, the base pricing is set at $0.10 per 1M input tokens and $0.50 per 1M output tokens—roughly 10% the cost of its predecessor and matching GPT-6 Luna. However, for prompts exceeding 100k tokens, rates scale by 5x to $0.50 (input) and $2.50 (output).

Key considerations for production deployment include:

1. Step-up Costs on Long Prompts: Exceeding the 100k token threshold increases input, output, and caching unit costs fivefold, necessitating careful chunking strategies and context budget management.
2. Safety Over-Refusal Sensitivity: Provisional evaluation noted an over-refusal issue that temporarily depressed its AutomationBench-AA score to 35%, pending safety adjustments and re-evaluation by Anthropic.
3. Task-Level Cost Profiling: Because adaptive thinking increases output token volume, total operational cost must be evaluated per task rather than solely on per-million token list prices.

Sources

• Artificial Analysis Official X (@ArtificialAnlys): Claude Haiku 5.5 Release & Benchmark Analysis
• Artificial Analysis Model Overview: Claude Haiku 5.5 Comparison & Specs