Anthropic's Claude 5.5 family: cheaper flagships, a smarter small model
Most model releases are a single point on a timeline. Anthropic just did something rarer: over three weeks it replaced its entire lineup — Opus 5.5 on September 22, Sonnet 5.5 on September 28, and Haiku 5.5 on October 7. Three releases usually stay separate news items. Read together, they tell a single story about where frontier AI is going, and it is not a story about raw capability. It is a story about price.
Start with the flagship. Claude Opus 5.5 performs at the level of the previous flagship, Claude Fable 5.1, on most work but costs 40% less to run than Opus 5. Its API pricing is $4 per million input tokens and $20 per million output tokens — a 20% cut from Opus 5's $5 and $25. The bigger number hides in the cache reads, which make up most of the cost of agentic and coding work. They dropped about 60%, to $0.20 per million tokens. Anthropic's message is plain: the same frontier quality at two-fifths the operating cost.
The middle tier follows the logic. Sonnet 5.5 keeps its predecessor's price — $2 and $10 per million — but needs far fewer tokens per task, which Anthropic says nets out to up to 30% lower cost per task than Sonnet 5. The company calls it a faster, lower-cost complement to Opus. Notably, it beats Opus 5.5 on some coding benchmarks — Terminal-Bench 4.0 at 70.6% versus 66.4%, and AutomationBench — despite costing half the list price. The "one model for everything" pattern is already eroding.
Then the small model, where the real change happens. Claude Haiku 5.5 is Anthropic's cheapest and fastest model, and its price is the headline: $0.10 input and $0.50 output per million tokens for prompts up to 100k tokens, compared with $1 and $5 for Haiku 4.5. That is roughly 90% cheaper on the short-prompt traffic that made up most previous Haiku demand; Anthropic puts the average saving at about 75%. It carries a 1M-token context window, up to 128k output, and an adjustable effort setting that lets developers trade depth for speed and cost.
The pricing would be easy to dismiss as a commodity squeeze if Haiku 5.5 had stayed a lightweight. It did not. The agentic numbers are what make the release meaningful. On the OSWorld 2.1 offline subset, Haiku 5.5 scores 72.4% against Haiku 4.5's 15.7%. On Terminal-Bench 4.0 it lands 39.2% where its predecessor scored zero. Nothing about that is flagship-grade yet — Sonnet 5.5 sits at 70.6% on the same coding test — but the gap between a $0.10 model and a $4 model is no longer a capability cliff. A year ago this kind of agentic work demanded the most expensive thing Anthropic sold. Today a model at one-fortieth the input price mostly keeps up.
That is the strategic shift underneath the three releases. Anthropic's framing of the family — the larger 5.5 model plans and judges while Haiku handles reading, context compaction, and narrow sub-tasks — describes an architecture of delegation, not a ladder of all-purpose replacements. Haiku is built for summarization, classification, and the high-volume sub-agent work that backs larger models. The flagship still sets the ceiling; the small model sets the scale.
For developers the takeaway is economic. When the cheapest model in a lineup can do real agentic work, the cost-per-task curve of building automation collapses. Anthropic has now shown the whole tier ladder moving down at once — flagships 20–40% cheaper, the mid-tier effectively cheaper per task, and a small model 90% cheaper that suddenly knows how to use tools. The interesting frontier of 2026 is less about which model is smartest and more about which work is now cheap enough to automate. That gap is closing fast.