Anthropic launches Claude Haiku 5.5 at $0.10 per million tokens, and cuts Sonnet's cache reads
Anthropic launched Claude Haiku 5.5 on October 7, the third Claude 5.5 family model in about a month, with input at $0.10 per million tokens for prompts up to 100,000, and halved Sonnet cache pricing.
The decision it puts on your desk
Split your agent stack into quality and utility tiers by October 15. Route summaries, extraction, and routing work to Haiku 5.5's $0.10 tier, run front-line model calls wherever your eval says, and price each tier against your actual token count after the 30 percent tokenizer shift. Real numbers land this week, from your own logs.
Anthropic shipped a small-model release with a big pricing lever on Tuesday.
"Haiku 5.5 is available at a much lower price than Haiku 4.5," the company said in its launch post. "On average, it now costs around 75% less to run."
The price: $0.10 per million input tokens and $0.50 per million output tokens, for prompts up to 100,000 tokens, from coverage of the launch. The same coverage says pricing rises five times higher above that threshold, and the context window reaches one million tokens.
Sonnet took the companion cut. "We're halving the price of Claude Sonnet 5.5's cache reads, which means Sonnet 5.5 now runs around 20% cheaper on most agentic work," the launch post said.
Who actually saves
Follow the money at the tier where it lands.
Sub-agent work saves most: the cheap tier that routes summaries, compaction, and bounded tasks through every agent loop now bills at Haiku 5.5 rates, and the cache-read cut rewards context-heavy sessions by halving their dominant line item.
One check tempers the headline. Coverage flagged that the tokenizer behind Haiku 5.5 counts roughly 30 percent more tokens than its predecessor's, so real savings sit per task rather than per token.
Precedent check: OpenAI's GPT-6.1 Sol shipped at $2 and $10 nine days ago, and DevDay's Decisions API pushed classification work toward cheaper tiers. Anthropic's answer prices two steps under it.
Ecosystem treatment, in one line each: Sol added speed, Argon added output, and Haiku 5.5 added cost discipline, all inside a week.
The unresolved edge is quality at the new tier, since the company call sits on benchmarks and real workloads rather than list prices.
I get the sense fleet buyers will split workloads by cost-per-task now.
Source
← Previous
Mistral ships Large 4: a trillion-parameter open-weight model trained entirely in the EU
Next →
Google Cloud ships the Gemini agent: standing coworkers with identities, budgets, and audit trails
The Morning Recap
We sift the day. You make the calls.
The decision on every story, in your inbox before your first call.