Anthropic released Claude Haiku 5.5 on October 7, 2026, dropping the cost of its frontier small-model tier by an average of 75% compared to Haiku 4.5. Alongside the new model, Anthropic also cut the price of Claude Sonnet 5.5’s cache reads in half, bringing the cost of most agentic work down by around 20%.
For small and mid-sized businesses automating their operations, model releases at the "small and fast" tier are often more consequential than flagship heavyweight models. Heavyweight models write the strategy; lightweight models do the daily chores. Haiku 5.5 changes the return on investment for high-volume, repetitive business processes.
A Restructured Cost Baseline
Haiku 5.5 introduces a two-tiered pricing structure based on prompt length. For prompts under 100,000 tokens—which Anthropic notes account for roughly 90% of requests to the previous Haiku model—input tokens cost $0.10 per million and output tokens cost $0.50 per million. If a prompt exceeds the 100,000-token threshold, those prices jump to $0.50 for input and $2.50 for output.
This pricing heavily incentivizes using the model as intended: for short, focused tasks rather than massive document ingestion. When kept within that 100,000-token window, the cost floor for basic automation is remarkably low.
Anthropic also introduced an adjustable effort setting for the first time in a Haiku-class model. This allows operations teams to dial the model's compute up or down depending on whether a specific task requires strict cost optimization or higher intelligence.
Concrete Applications for SMBs
In benchmark testing, Haiku 5.5 outperformed both its predecessor and GPT-6 Luna across multidisciplinary reasoning, agentic coding, and knowledge work. But practically, this model is built to act as the routing and extraction engine within larger automation workflows.
The Subagent Pattern Complex automations often fail or become prohibitively expensive when a single large model is asked to do everything. Haiku 5.5 is designed to function as a "subagent" that pairs with heavier models like Opus 5.5 or Sonnet 5.5. For example, a financial automation might use Haiku 5.5 to scan a 10-K and extract a single segment revenue line, passing that exact figure back to a larger model assembling a client deck. Because it is highly accurate on narrow tasks and extremely cheap, it can be triggered hundreds of times an hour without ballooning API costs.
High-Speed Triage and Classification Latency is the primary bottleneck for live customer support or tasks requiring browser use. Early testing from enterprise users highlights a steep drop in response times. Asana reported up to 2.5x faster inference per agent turn and a 30% reduction in latency for task completions. Box noted that the model scored 11 points higher on internal benchmarks than Haiku 4.5 at roughly half the latency. For an SMB managing inbound ticketing, Haiku 5.5 can accurately classify, tag, and route requests almost instantly.
CRM Auditing and Data Clean-up Keeping databases accurate is a notorious drain on human capital. HubSpot’s testing of Haiku 5.5 on CRM tasks—such as identifying stale but ambiguous records—yielded a 92.8% average score. More importantly, the model was the fastest to complete the task, with the highest hit rate and lowest false-positive rate. For operations teams, this makes routine database deduplication and compaction a prime candidate for immediate automation.
A Push Toward Agentic Workflows
Alongside the Haiku 5.5 release, Anthropic’s decision to halve Sonnet 5.5 cache reads directly targets agentic workflows. By making it 20% cheaper to run Sonnet 5.5 for complex, multi-step agentic work, and introducing Haiku 5.5 as a hyper-fast, low-cost subagent, the overall cost of running fully automated, multi-model systems has dropped significantly.
For businesses that have hesitated to deploy automations due to API costs or slow execution times, this release resets the baseline. The underlying tools for building fast, reliable, and cost-effective operations are now lighter and sharper than before.