What changed in Claude Sonnet 5.5 pricing?
On October 7, 2026, Anthropic updated the price of prompt cache reads for Claude Sonnet 5.5. A million tokens read from an existing prompt cache now costs $0.10 instead of $0.20.
Anthropic's Claude Platform release notes describe the new price as 0.05 times the model's standard input price, down from 0.1 times the input price. The same release note explicitly says that cache writes and all other Sonnet 5.5 prices remain unchanged.
The change arrived alongside the launch of Claude Haiku 5.5, but it applies specifically to Sonnet 5.5 cached prompt reads. It is therefore a real reduction in one of Sonnet 5.5's API token rates rather than merely an efficiency claim about the model completing tasks with fewer tokens.
Official confirmation: Claude Platform release notes and Anthropic's October 7 announcement .
Claude Sonnet 5.5 pricing after the cut
The important distinction is that cache reads became cheaper. Anthropic did not halve every Sonnet 5.5 API rate.
| Usage type | Previous price | Current price | Change |
|---|---|---|---|
| Standard input | $2 / MTok | $2 / MTok | No change |
| Standard output | $10 / MTok | $10 / MTok | No change |
| 5 minute cache write | $2.50 / MTok | $2.50 / MTok | No change |
| 1 hour cache write | $4 / MTok | $4 / MTok | No change |
| Cache read | $0.20 / MTok | $0.10 / MTok | 50% cheaper |
Anthropic's current Sonnet 5.5 model page also lists a 50% Batch API discount on standard input and output. That batch discount is separate from the October 7 prompt-cache reduction.
Current model pricing: Claude Sonnet 5.5 model specifications .
What is a prompt cache read?
Prompt caching is useful when an application repeatedly sends Claude a large block of context that does not need to be processed from scratch every time. The reusable portion can be written into a prompt cache, and later requests can reference that cached material.
Examples can include a large system prompt, a codebase description, documentation, a long background file, tool definitions, operating instructions or other stable context that appears again across multiple requests.
Creating the cached content is a cache write. Reusing that stored context on a later request is a cache read. Sonnet 5.5 cache writes still cost more than ordinary input because they create reusable cached state, while subsequent reads are much cheaper.
- First request
- The reusable prompt material is written into the cache. Anthropic charges the applicable five-minute or one-hour cache-write rate.
- Later requests
- When the same cached prefix is reused successfully, those cached tokens are billed at the cache-read rate rather than the full standard input rate.
- New Sonnet 5.5 rate
- Cache reads now cost $0.10 per million tokens rather than $0.20.
Technical guidance: Anthropic prompt caching documentation .
Why Anthropic says most agentic work can cost around 20% less
The headline reduction is 50%, but that does not mean every Sonnet 5.5 API workload suddenly costs half as much. Only the cache-read component was reduced.
Anthropic says the overall effect is approximately 20% lower cost on most agentic work. The reason is that agent-style applications often reuse large blocks of context repeatedly while a task progresses through many steps.
An agent may carry the same system instructions, tool definitions, project context, repository information or reference material through many turns. If much of that content is eligible for prompt caching, cache reads can represent a significant portion of the total token bill.
Cutting that component from $0.20 to $0.10 per million tokens therefore reduces the full task cost even though standard input and output rates remain unchanged.
Important: Anthropic's roughly 20% figure describes its estimate for most agentic workloads. Your actual saving depends on how much cached context your application reuses. A workload that does not use prompt caching will not receive this particular discount.
Anthropic's cost statement: Claude Haiku 5.5 announcement, Further updates .
What did not change?
The phrase “Sonnet 5.5 pricing cut” can easily be interpreted too broadly. Anthropic did not announce a 50% reduction across every Sonnet 5.5 charge.
- Input tokens
- Standard uncached input remains $2 per million tokens.
- Output tokens
- Standard output remains $10 per million tokens.
- Five-minute cache writes
- The rate remains $2.50 per million tokens.
- One-hour cache writes
- The rate remains $4 per million tokens.
- Claude subscription prices
- The cache-read change is an API billing change. It is not a 50% discount on Claude Pro, Max, Team or Enterprise subscriptions.
- Five-hour and weekly limits
- Anthropic's October 7 API cache-price announcement does not replace or halve Claude's subscription usage limits.
Who benefits most from the lower cache-read price?
The saving matters most when the same large context is reused repeatedly. That makes the change especially relevant to developers building long-running or multi-step systems with Claude Sonnet 5.5.
-
Coding agents. A coding workflow can repeatedly reuse repository context, tool definitions and operating instructions while it inspects, edits and tests files.
-
Research agents. Long background documents or stable instructions can stay cached while the agent performs multiple searches, comparisons and synthesis steps.
-
Customer support systems. A large policy manual, product documentation or support framework can be reused across many turns instead of being billed as fresh uncached input each time.
-
Tool-heavy agents. Complex tool definitions and persistent system instructions can be good candidates for caching when they remain stable between requests.
-
Long document workflows. Applications that repeatedly work against the same large source material can benefit when the reusable prefix qualifies for cache hits.
Short one-off requests may see little or no difference because there may be no reusable context worth caching. The value of this change therefore increases with repeated context reuse, not simply with the number of API requests.
Does this make Claude Pro or Max 50% cheaper?
No. The October 7 change concerns Claude Platform API prompt cache reads. It should not be confused with the price of a Claude consumer subscription or the amount of usage included in that subscription.
Claude's app subscriptions and Claude Platform API billing are
different systems. A person using Sonnet 5.5 through the Claude
website or app normally consumes their plan's included usage
allowance. A developer calling
claude-sonnet-5-5
through the API is billed according to Claude Platform token
pricing and any applicable credits.
| Usage route | How this price cut applies |
|---|---|
| Claude web and apps | The cache-read API price does not make the subscription itself 50% cheaper. |
| Claude Code using plan allowance | Subscription usage limits remain a separate system from API token billing. |
| Claude API | Eligible Sonnet 5.5 cache reads are now billed at $0.10 per million tokens. |
| Claude Console and API agents | The new price can reduce spending when the workload produces Sonnet 5.5 prompt cache hits. |
Anthropic also added monthly API credits for Max and Team
The cache-read reduction was not the only developer-cost change Anthropic announced around October 7. Max and Team subscribers can now receive monthly Claude Platform API credits after linking an eligible Claude Console organization.
| Eligible plan | Monthly API credit |
|---|---|
| Max 5x | $100 per billing cycle |
| Max 20x | $200 per billing cycle |
| Team Standard | $20 per eligible seat |
| Team Premium | $100 per eligible seat |
Anthropic says Team credits are pooled and currently capped at $500 per month. The credits can be used for the Claude API, Playground, Claude Managed Agents and Claude Agent SDK. They are separate from the plan's ordinary Claude, Claude Code or Cowork usage limits.
Free, Pro and Enterprise plans are not currently eligible for this monthly API-credit program. Anthropic also says unused credits expire at the end of the billing cycle rather than rolling over.
Separate changes: the Max and Team API credit program does not cause the Sonnet 5.5 cache-read price cut. Both changes improve API economics, but they work independently.
Current eligibility and credit amounts: Anthropic Help Center .
What developers should do now
If you already use Sonnet 5.5 through the Claude API, you do not need to change the model ID to receive the new published cache-read rate. The bigger question is whether your workload is actually structured to generate useful prompt cache hits.
-
Identify stable repeated context. Look for large instructions, tools, documents or other prefixes reused across many calls.
-
Check whether prompt caching is appropriate. Use Anthropic's current caching documentation rather than assuming every repeated token automatically becomes a cached read.
-
Measure cache hits. Compare actual cache-read token volume with normal input and output usage to understand where your bill is coming from.
-
Recalculate agent economics. Workloads previously rejected as too expensive may deserve another look if repeated cached context represents a large share of their token usage.
-
Keep model quality in the decision. A lower cache price does not automatically make Sonnet 5.5 the cheapest model for every workload. Smaller models can remain more economical for narrow, high-volume tasks.
Bottom line
Claude Sonnet 5.5 now has a genuine lower API rate than it did at launch: prompt cache reads cost $0.10 per million tokens instead of $0.20. That component is exactly 50% cheaper.
Standard Sonnet 5.5 input remains $2 per million tokens and output remains $10. Cache writes also remain unchanged. Therefore, the model is not universally “50% cheaper.” Anthropic's broader estimate is that the new cache-read rate makes Sonnet 5.5 roughly 20% cheaper on most agentic work.
For developers building long-running coding, research, support or tool-using agents, the practical impact can be meaningful because the same large context is often reused many times. For ordinary Claude subscription users, however, this should be understood as an API pricing update rather than a reduction in subscription prices or plan usage limits.
Latest official references: Anthropic announcement , Claude Platform release notes , and Sonnet 5.5 model page .