The 5.5 Family Is Complete: What It Means for Buyers
With Haiku 5.5's arrival, Anthropic's 5.5 lineup is finished. Here is the complete pricing picture, what each tier is actually for, and the migration deadlines buyers should lock in now.
The 5.5 Family Is Complete: What It Means for Buyers
Three models in nineteen days. Anthropic opened the 5.5 series with Claude Opus 5.5 on September 22, added Sonnet 5.5 on September 28, and finished with Haiku 5.5 on October 7. If you buy inference by the token, the lineup is now stable enough to plan around — and the 4.x leftovers have explicit end dates.
The lineup and the numbers
Opus 5.5 — $4/$20 per million tokens, a 20% cut from Opus 5's launch pricing, with cache reads at $0.20 and cache writes at $5 (5-minute) or $8 (1-hour). It still tops the BenchLM composite at 86.3 and remains the pick for the hardest reasoning work.
Sonnet 5.5 — the balanced tier, now on public benchmarks since its GA (roughly 83.8 on BenchLM, 81.3% SWE-bench Pro). From October 7 its cache reads dropped to $0.10 per million, which quietly makes it the most cache-friendly model in the family and the natural default for coding agents and other high-repetition workloads.
Haiku 5.5 — $0.10/$0.50 per million, about 75% below Haiku 4.5. This is the high-volume tier: classification, extraction, routing, moderation, anything where cost per call dominates.
Deadlines worth putting in a calendar
The 4.x generation is closing out on a schedule: Sonnet 4.5 was deprecated on September 30 and shuts down on November 30; Haiku 4.5 left the model documentation on October 8 with a final retirement floor of October 15. Traffic still pinned to 4.5 endpoints should move before those dates — the pricing gap makes Haiku 5.5 the obvious landing spot for most Haiku 4.5 workloads, while Sonnet 4.5 traffic is cheaper after moving to Sonnet 5.5 anyway.
How the rest of the market sits
OpenAI's GPT-6 Astra ($10/$50, 1.05M context) remains the premium flagship and finished its staged rollout to all ChatGPT users on October 7. The more interesting release for buyers is GPT-6.1 Sol, announced at DevDay on September 29: a same-price successor to GPT-6 Sol with $0.10 cache input and near-Astra intelligence — the same play Anthropic just made with Haiku 5.5. Google's Gemini 3.8 Flash holds promotional pricing of $0.75/$3.75 through December 31, 2026, with standard rates of roughly $1.50/$7.50 from January 1.
The buyer's summary
Route by shape of work, not by brand: Opus 5.5 for the hardest problems, Sonnet 5.5 with caching for agentic and coding loops, Haiku 5.5 for volume. Benchmarks currently reward that split — and for the first time this year, every tier of a frontier family is priced below its predecessor.