Anthropic has filled out the 5.5 family in under a month: Opus 5.5 on September 22, Sonnet 5.5 on September 28, and now the smallest sibling. On October 7, Claude Haiku 5.5 went live — billed by Anthropic as "the cheapest, fastest, and most capable small model we've ever released," with input pricing pushed down to $0.10 per million tokens, versus $1.00 for Haiku 4.5.

Pricing: a 100K-token watershed hides inside the 90% cut

The pricing is two-tier: prompts up to 100,000 tokens cost $0.10 per million input and $0.50 per million output; above that line, $0.50 and $2.50. The official footnote is blunt: 90% cheaper under 100K, 50% cheaper over it — and about 90% of requests to the previous Haiku fell under the line, averaging out to "around 75% cheaper."

Two details deserve scrutiny. First, caching: the under-100K tier prices cache reads at $0.01 per million and writes at $0.125 — a real lever for repetitive agent workloads. Second, the tokenizer change: Haiku 5.5 adopts the updated tokenizer shared with Sonnet 5.5 and Opus 5.5, consuming slightly more tokens per task. The third-party newsletter roo re-ran Anthropic's published prices on launch day and found the real cut is about 87% below the line and about 35% above it — "90% cheaper" is a story that only holds if you read the footnote too.

Benchmarks: the small model jumps harder than the big ones

On Anthropic's own table, Haiku 5.5 posts order-of-magnitude gains over its predecessor: the OSWorld 2.1 offline subset (real computer use) jumps from 15.7% to 72.4%, with GPT-6 Luna at 48.9%; Humanity's Last Exam without tools goes from 10.2% to 45.9%; Terminal-Bench 4.0 goes from 0% to 39.2%; on FrontierCode 1.1 it scores 46.4%, ahead of GPT-6 Luna's 42.4%. On knowledge work, GDPval-AA v2.1 rises from 735 to 1620, overtaking GPT-6 Luna's 1437.

The ceiling is marked clearly too: Sonnet 5.5 sits at 1840 in the same table's reference column (83.9% OSWorld, 70.6% Terminal-Bench). Anthropic itself concedes that complex agentic coding remains Sonnet/Opus territory; Haiku's home ground is narrowly scoped tasks that used to be cost-prohibitive — compaction, summarization, classification, subagent work.

Customers: cheap enough to run without thinking

Early customers use it almost uniformly as a subagent. Asana reports over 30% lower task-completion latency and up to 2.5x faster inference per agent turn; HubSpot logged 92.8% — its best score yet — on a CRM eval suite; AlphaSense's document Q&A runs 8 million calls a week and improved from 0.76 to 0.84; Box saw 11 points higher scores at about half the latency; Cognition dropped it into Devin Fusion as the sidekick, holding a top-tier FrontierCode score of 66.2. Haiku 5.5 is also the first Haiku-class model with an adjustable effort setting — the official charts label five levels from low to max — letting you tune cost versus intelligence per request. On specs, third-party pages and Artificial Analysis both list a 1M context window and 128K max output, closing the gap with Haiku 4.5's 200K.

Three more changes shipped alongside

One: Sonnet 5.5 cache reads drop from $0.20 to $0.10 per million, which Anthropic estimates makes most agentic workloads about 20% cheaper. Two: Max and Team subscribers get a new monthly API credit — $100 for Max 5x, $200 for Max 20x, up to $500 pooled for Team. Three: the Python and TypeScript SDKs add beta support for computer use and browser use — exactly the real-time, interactive workloads Haiku 5.5's speed-plus-price combination was built for.

So what? Read alone, "ten-cent input" is a price-cut headline. Read together, it's agent economics: when the marginal cost of a subagent drops by an order of magnitude, "one big model directing a fleet of small ones" stops being an architecture aesthetic and becomes financial common sense. Teams planning to migrate should first do the math on the 100K watershed and the tokenizer swap.

References: Anthropic announcement · roo pricing analysis