Skip to content
Back
Pricing briefing
AAnthropic
Anthropic
PRICING UPDATE

Anthropic launched a new Haiku model and cut its Sonnet cache price

By , Co-founder, PricingSaaS

Published

Anthropic launched Haiku 5.5 and cut Sonnet 5.5's cache read price.

Anthropic’s pricing page · Change detected

What changed

Anthropic added Haiku 5.5 to its Latest models on the pricing page Pulse saved on Oct 8, 2026, priced at $0.10 per MTok input and $0.50 per MTok output for prompts under 100K tokens, rising to $0.50 and $2.50 above that, with cache reads at $0.01 and cache writes at $0.125 per MTok ($0.05 and $0.625 for prompts over 100K). According to Anthropic's pricing page on Oct 8, 2026, Haiku 4.5 moved to the Legacy models section that same day with its prices unchanged at $1 input and $5 output per MTok, cache $0.10 and $1.25.

On the same pricing page, Sonnet 5.5's prompt-caching read price was cut from $0.20 to $0.10 per MTok as of Oct 8, 2026, while its cache write ($2.50), input ($2) and output ($10) prices were unchanged.

Why it matters

For product leaders building agent loops, a model priced at $0.10 input and $0.50 output per MTok (Anthropic's pricing page, Oct 8, 2026) changes the economics of running many small LLM calls: tasks that were too expensive to automate with a frontier model become cheap enough to run per step. Lean on Haiku 5.5 for the high-volume, low-stakes legs of a pipeline, and reserve Sonnet-class models for the steps that need more judgment.

The Sonnet 5.5 cache-read cut on Anthropic's pricing page on Oct 8, 2026 matters more for teams with long, stable system prompts and tool definitions: a cache-hit-heavy workload now pays roughly half as much on every repeated read, per that same update. Before re-architecting a prompt to chase this, check your actual cache hit rate — the saving only shows up on tokens that were already being served from cache.

Watch whether Anthropic keeps trimming cache and small-model prices ahead of its next context-window or positioning change, rather than assuming today's rates are locked in.

In context

Anthropic has been tracked since Jan 3, 2025, with 17 real pricing changes in that span, and this change lands 14 days after it added a Developer tab to the pricing page on Sep 24, 2026. The weeks before that saw a cluster of context-window and plan work — raised context windows up to 1M and revamped Pro and Max plans on Sep 17, 2026, then boosted the context window to 1M across all Claude plans on Sep 21, 2026 — on top of launching Sonnet 5 with a lower price back in July. Against that run, this release reads less like a one-off promo and more like Anthropic routinely tuning per-model and cache pricing every time it ships a new model. For a product leader, the sharper read is that cache pricing is becoming its own lever, separate from base input/output rates, and worth tracking as its own line item.

Common questions

How would this change a team's monthly API bill?

Pull your own per-model token volume from your usage dashboard, run it against the new rates, and compare that against what you're paying today on Haiku 4.5 or Sonnet 5.5 before switching any workload over.

Does the cache-read price cut help every workload?

No — it only helps if your prompts actually hit the cache; teams with short-lived or constantly-changing system prompts and tool definitions won't see Sonnet 5.5's Oct 8, 2026 saving show up on their bill.

Should we expect these rates to hold, or change again soon?

Treat them as provisional — Anthropic's own pricing-page history shows 4 changes in the last 90 days, so recheck rates before locking in long-term cost estimates.

What people are saying

@Mr_SalioX

“🚨 Haiku 5.5 Is Now live And the pricing looks really cheap • Under 100K tokens: $0.10/M input + $0.50/M output • Over 100K tokens: $0.50/M input + $2.50/M output Anthropic might’ve cooked with this one”

View original
@VelessusX

“anthropic launched haiku 5.5 yesterday: $0.10 in, $0.50 out per million tokens, about 75% cheaper than haiku 4.5. a small model that can drive a computer now costs almost nothing per step. the agent loop got cheap. what would you run in it?”

View original
@NileshblogX

“Sonnet 5.5 prompt cache reads just dropped to $0.10/Mtok (was $0.20). Cache writes stayed the same. If your agent keeps the system prompt and tool defs stable, that is the line that actually moves the bill. Check hit rate before you rewrite the prompt. #Claude #Backend”

View original

Sources