Prompt length sets what Anthropic's Haiku 5.5 price cut is worth to each workload

Haiku 4.5 cost $1/$5 per million input/output tokens; Anthropic prices Haiku 5.5 at $0.10/$0.50 for requests under 100,000 tokens.

For larger requests, above 100,000 tokens, Haiku 5.5 costs $0.50/$2.50 per million input/output tokens; Haiku 4.5 costs the same regardless of the number of tokens in a request.

The new Haiku 5.5 prices are 90% and 50% cuts to per-token prices for the short and long tiers respectively.

In Anthropic's own evaluations, Haiku 5.5 scored 72.4% on the offline subset of OSWorld 2.1, which tests computer use, up from 15.7% for Haiku 4.5 and ahead of GPT-6 Luna's 48.9%.

On Terminal-Bench 4.0, which tests complex multi-step tasks in a command-line environment, Haiku 5.5 scored 39.2%, compared with 0% for Haiku 4.5, 16.4% for GPT-6 Luna and 70.6% for Sonnet 5.5.

For requests with prompts up to 100,000 tokens, prices drop by up to 90 percent; prompts longer than 100,000 tokens cost five times as much.

Anthropic notes that about 90% of requests to Haiku 4.5 fell into the lower-priced category under 100,000 tokens.

Haiku 5.5 is available on the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure; developers on Anthropic's platform can access it as claude-haiku-5-5.

Anthropic has released Claude Haiku 5.5, its fastest and most affordable small model, built for high-volume tasks like data queries and customer support.

Anthropic says Haiku 5.5 works best for narrowly scoped tasks like compaction, summarization, or sub-agent work; for complex agentic coding, Sonnet 5.5 and Opus 5.5 remain the better picks.

Anthropic is cutting Sonnet 5.5's cache-read price in half, from $0.20 to $0.10 per million tokens, and says this should make most agentic tasks about 20% cheaper.

The Sonnet 5.5 cache-read cut is rolling out across platforms on Wednesday, though some existing Azure and Google Cloud customers will have to wait a few days.

Decision models like Jev are making a splash on high-volume tasks such as summarization, classification and routing, at an even lower price than small models like Haiku.

On Alibaba's international service, Qwen3.7 Flash costs $0.03 per million input tokens and $0.13 per million output tokens for inputs up to 32,000 tokens, rising to $0.10 and $0.40 for inputs above 32,000 and up to 256,000 tokens.

Artificial Analysis gives Z.ai's GLM-5.3-Flash a score of 1,647 on GDPval-AA v2.1; Anthropic reports 1,620 for Haiku 5.5.

If Haiku 5.5's tokenizer inflated token counts by 30%, a short-tier input would cost the equivalent of $0.13 per million old-tokenizer tokens, 87% below Haiku 4.5's $1.

At 30% token inflation, a long-tier input would cost the equivalent of $0.65 per million old-tokenizer tokens, 35% below Haiku 4.5. Output ratios match because every tier on both models prices output at five times input ($2.50 x 1.3 = $3.25 vs $5 = 35% saving; $0.50 x 1.3 = $0.65 vs $5 = 87%).

If the long-tier rate applies to the whole request, a 101,000-token prompt costs about $0.05 in input against about $0.01 for a 99,000-token prompt, roughly five times as much for 2% more tokens.

Weighting the per-token cuts by request count gives 0.9 x 90% + 0.1 x 50% = 86%. Because each request above 100,000 tokens carries more tokens than any request below it, the long tier's share of a token-weighted bill exceeds its 10% share of requests.

For halving the cache-read price to cut total task cost by 20%, cache reads must have been about 40% of the task's cost before the cut; tasks with a smaller cached share save less.

Haiku 5.5's short-tier input price and Sonnet 5.5's new cache-read price are both $0.10 per million tokens.

On average, Haiku 5.5 costs about 75 percent less than Haiku 4.5.

Anthropic puts the average savings at around 75%, a figure that accounts for a mix of requests and an updated tokenizer that uses slightly more tokens per task.