Anthropic has launched Claude Haiku 5.5, introducing a 90% API price reduction for requests under 100,000 tokens to match OpenAI's GPT-6 Luna at $0.10 per million input tokens. Built for high-volume enterprise tasks like classification and summarization, the model reduces average workload costs by 75% compared to Haiku 4.5. Early adopters report significant latency improvements, though maximum benchmark performance requires adjusting the model's effort settings.
Launch of Claude Haiku 5.5
- ▪Anthropic launched Claude Haiku 5.5 ahead of a planned initial public offering (IPO)
- ▪Anthropic launched Claude Haiku 5.5 on October 7, 2026, marking the third model released in Anthropic's Claude 5.5 family within a month
Claude Haiku 5.5 pricing and cost reductions
- ▪Anthropic estimates that Claude Haiku 5.5 workloads cost approximately 75% less to run than on Claude Haiku 4.5, factoring in request sizes, token consumption, and an updated tokenizer
- ▪Claude Haiku 5.5 costs $0.10 per million input and $0.50 per million output tokens below 100,000 tokens (90% reduction from Claude Haiku 4.5), and $0.50/$2.50 per million input/output tokens above 100,000 (50% reduction from Claude Haiku 4.5)
Comparison with competitor pricing
- ▪Claude Haiku 5.5's lower-tier rates are cheaper than Google's Gemini 3.5 Flash-Lite, which is priced at $0.30 per million input tokens and $2.50 per million output tokens
- ▪Claude Haiku 5.5's lower-tier pricing matches OpenAI's GPT-6 Luna base rates of $0.10 per million input and $0.50 per million output tokens, though Luna's surcharge begins above 272,000 input tokens
Targeted applications
- ▪Anthropic positions Claude Haiku 5.5 as a supporting worker for its larger Claude Opus and Claude Sonnet models, targeting repetitive tasks like document summarization, data classification, and database querying
- ▪Claude Haiku 5.5 is built to handle tasks including extraction, live support, voice agents, and in-app assistants
Performance on benchmark tests
- ▪In vendor-reported benchmarks, Claude Haiku 5.5 scored 1,620 on GDPval-AA v2.1 and 1,578 on AA-Briefcase v1.1, outperforming both Claude Haiku 4.5 and OpenAI's GPT-6 Luna
- ▪Claude Haiku 5.5 scored 39.2% on Terminal-Bench 4.0 at maximum effort, but its score dropped to approximately 20% when run at its default medium effort level
Feedback from early adopters
- ▪Asana staff software engineer Aaron Vinh reported that task-completion latency fell by more than 30% and inference per agent turn accelerated up to 2.5 times using Claude Haiku 5.5
- ▪HubSpot software engineer Ze'ev Klapow reported that Claude Haiku 5.5 achieved a 92.8% average across three runs of HubSpot's CRM evaluation
- ▪Box VP of AI Products Yashodha Bhavnani reported that Claude Haiku 5.5 delivered an 11-point improvement over Claude Haiku 4.5 with approximately half the latency
- ▪Rogo applied AI worker Alex Wang stated that Claude Haiku 5.5 is accurate enough to trust for retrieving specific figures while being fast and cheap enough to run frequently
Cybersecurity safeguards
- ▪Claude Haiku 5.5 is Anthropic's first Haiku model with built-in safeguards for high-risk cybersecurity requests, though Anthropic stated most everyday tasks will not be impacted
- ▪Claude Haiku 5.5 introduces built-in safeguards that block penetration testing and restrict high-risk cybersecurity requests, requiring organizations to apply through verification programs for broader access
Sonnet price cuts
- ▪Alongside the launch of the Claude Haiku 5.5 model, Anthropic cut Claude Sonnet 5.5 cache read charges from $0.20 to $0.10 per million tokens and introduced monthly API credits for Max and Team subscribers
- ▪Anthropic estimates that typical agent workloads will save approximately 20% from the Claude Sonnet 5.5 caching price reduction, depending on context reuse
Debatable claims
- ▪The aggressive price cutting among AI providers is unsustainable
- ▪Anthropic's default block on AI penetration testing harms cybersecurity
Story comments
Loading comments…