On October 7, Anthropic officially unveiled Claude Haiku 5.5. As the most compact version in the Claude 5.5 family, this model aims to optimize costs for large-scale data processing tasks. According to the company’s announcement, Haiku 5.5 not only delivers the fastest processing speeds currently available, but also cuts costs by an average of roughly 75% compared to the previous Haiku 4.5 generation.

To understand this 75% savings, one must look at the details of Anthropic’s pricing structure:

  • 90% reduction: Applies to prompts of 100,000 tokens or fewer.
  • 50% discount: Applies to tasks exceeding the 100,000-token threshold.

Because the vast majority of Haiku 4.5’s actual production traffic (around 90%) falls into the under-100,000 token bracket, the average overall savings comes out to roughly 75%. This figure already accounts for changes to Haiku 5.5’s tokenizer, which slightly increases token consumption per request.

Listed Pricing Details

According to Anthropic’s official announcement, the pricing per million tokens for Haiku 5.5 is structured as follows:

  • Tasks ≤ 100k tokens: $0.10/M input tokens - $0.50/M output tokens - $0.01/M cache read tokens - $0.125/M cache write tokens.
  • Tasks > 100k tokens: $0.50/M input tokens - $2.50/M output tokens - $0.05/M cache read tokens - $0.625/M cache write tokens.

The reduction of Haiku 5.5’s input price to the $0.10 per million token mark drew attention from Yahoo Finance, which directly compared the model to GPT-6 Luna—OpenAI’s most lightweight offering. This step marks an important milestone in the race to optimize AI costs, particularly as organizations tighten budgets following a period of heavy token expenditure and begin weighing open-weights models from both the United States and China.

Haiku 5.5 lowering its input cost to $0.10 was directly compared by Yahoo Finance to GPT-6 Luna, OpenAI’s smallest model, signaling a new turning point in the race to drive down AI costs. The move comes as enterprises actively curb budgets following widespread spending on tokens, while also exploring the integration of open-weights models from China and the United States.

Two Accompanying Changes

Alongside lowering costs for its lightweight model tier, Anthropic introduced two other significant adjustments:

  • Slashing Sonnet 5.5 cache read pricing: Cache read costs were cut by 50%, from $0.20 to $0.10 per million tokens. Because cache reads account for a large share of the token volume processed by agents, this move reduces Sonnet 5.5 operating costs by about 20% across most agentic workflows.
  • Recurring monthly API credits: Starting this week, subscribers to Max and Team plans will receive API credits valid across all models. Specific allowances include $100 for the Max 5x plan, $200 for the Max 20x plan, and up to $500 shared among members of the Team plan.

In addition, the company added beta support for browser use and computer use to both its Python and TypeScript SDKs. Haiku 5.5 also becomes the first model in the Haiku family to support a flexible effort parameter, allowing users to easily strike a balance between processing performance and operational budgets.

Benchmark Results and Key Caveats

According to results released by Anthropic, Haiku 5.5 demonstrated across-the-board improvements over its predecessor Haiku 4.5, while also posting stronger numbers than GPT-6 Luna across categories with comparative data:

  • GDPval-AA v2.1: Scored 1,620 points (compared to 735 points for Haiku 4.5 and 1,437 points for GPT-6 Luna).
  • OSWorld 2.1: Reached 72.4% (well above the 15.7% achieved by the older version).
  • Humanity’s Last Exam (without tools): Scored 45.9% (compared to 10.2% previously).
  • Terminal-Bench 4.0: Reached 39.2% (compared to 0% for Haiku 4.5).

Even so, Anthropic candidly acknowledged that Haiku 5.5 is not an optimal solution for every scenario. For highly complex agentic coding tasks, models like Sonnet 5.5 and Opus 5.5 continue to demonstrate far superior capabilities (such as Sonnet 5.5 reaching 70.6% on the Terminal-Bench 4.0 benchmark). Haiku 5.5 remains best suited for narrow, highly repetitive workloads such as summarization, context compression, or serving as a subagent—tasks that were previously deemed too cost-prohibitive.

Furthermore, feedback from enterprise customers should be evaluated as internally compiled vendor data:

  • Asana: Reported a more than 30% reduction in task completion latency, along with up to 2.5x faster per-turn agent inference speeds.
  • HubSpot: Recorded an average score of 92.8% across three trial runs.
  • AlphaSense: Ran 400 queries and observed a statistically significant improvement over Haiku 4.5.

Notably, the figures and feedback highlighted above reflect hands-on user experiences during testing and have not been independently verified.

In terms of safety, Anthropic reported that Haiku 5.5 significantly curtails jailbreak behaviors compared to Haiku 4.5. Thanks to stronger defenses against cybersecurity threats, the model maintains the agility needed for defensive tasks compared to Sonnet 5.5.

As input costs for small models approach the $0.10 threshold and reach parity with GPT-6 Luna, the market’s competitive focus is bound to shift. The deciding factor will no longer be the headline sticker price, but the actual total cost required to finish an end-to-end task—including retry frequency and error rates.