What does Claude Sonnet 5.5 actually cost?
Anthropic launched Claude Sonnet 5.5 on September 28, 2026 with the same headline API pricing as Claude Sonnet 5.
Sonnet 5.5 input
Input tokens
$2 / 1M
Standard Claude API input-token price reported by Anthropic at launch.
Sonnet 5.5 output
Output tokens
$10 / 1M
Standard output-token price before applicable caching, batch or geography modifiers.
| Token category | Sonnet 5.5 | Sonnet 5 |
|---|---|---|
| Standard input | $2 / MTok | $2 / MTok |
| Standard output | $10 / MTok | $10 / MTok |
| Cache read | $0.20 / MTok | $0.20 / MTok |
| Cache write | $2.50 / MTok in Anthropic's launch table | $2.50 / MTok for the standard 5-minute cache write |
MTok means one million tokens.
Source: Anthropic: Introducing Claude Sonnet 5.5 .
So is Sonnet 5.5 cheaper per token than Sonnet 5?
No.
If your application sends the same number of input tokens and receives the same number of output tokens from both models, the standard token charge is the same.
The distinction that matters
Same token price does not necessarily mean same task price
API pricing tells you what one million tokens cost. It does not tell you how many tokens the model will need to solve a task. Anthropic's efficiency claim is that Sonnet 5.5 often reaches the result with less model work.
This is why two apparently contradictory statements can both be true:
Sonnet 5.5 costs the same per token as Sonnet 5.
And, according to Anthropic's testing, Sonnet 5.5 can cost less per task.
Why some people expected Sonnet 5.5 to have a lower token price
Part of the confusion comes from the pricing history of Claude Sonnet 5 itself.
Sonnet 5 launched at $2 per million input tokens and $10 per million output tokens, but Anthropic initially described that rate as introductory.
The company originally planned for Sonnet 5 to move to $3 input and $15 output per million tokens on September 1, 2026.
That increase never happened.
Anthropic updated its Sonnet 5 announcement on August 10 and made the $2/$10 introductory rate permanent.
| Stage | Input | Output | What happened |
|---|---|---|---|
| Sonnet 5 launch | $2 / MTok | $10 / MTok | Announced as introductory pricing. |
| Originally planned after Aug. 31 | $3 / MTok | $15 / MTok | Scheduled standard price. |
| August 10 change | $2 / MTok | $10 / MTok | Anthropic made the lower rate permanent. |
| Sonnet 5.5 launch | $2 / MTok | $10 / MTok | Same headline token rates as Sonnet 5. |
Source: Anthropic's Claude Sonnet 5 announcement and August 10 pricing update .
What does “up to 30% less per task” actually mean?
Anthropic says Sonnet 5.5 typically needs far fewer tokens to do the same work and, in its testing, cost up to 30% less per task than Sonnet 5.
The phrase “per task” is important.
A completed agentic task can involve more than one model response. The model may think, call tools, read results, call another tool, revise its approach and continue until the requested work is complete.
Each additional step can create more billable input and output tokens.
If Sonnet 5.5 reaches the same acceptable result with fewer steps, fewer tool calls or shorter reasoning, the total bill can fall even though every individual token still carries the same base rate.
“Up to 30%” is not a guaranteed discount. It is Anthropic's reported result from its testing. Your own savings can be smaller, larger in a particular workflow, or effectively zero if the two models consume similar amounts of billable work.
Example 1: If token usage is identical, there is no price saving
Imagine both models process exactly 1 million input tokens and generate exactly 100,000 output tokens.
Sonnet 5
Input: 1,000,000 × $2 / 1M = $2.00
Output: 100,000 × $10 / 1M = $1.00
Total = $3.00
Sonnet 5.5
Input: 1,000,000 × $2 / 1M = $2.00
Output: 100,000 × $10 / 1M = $1.00
Total = $3.00
In this scenario Sonnet 5.5 is 0% cheaper because the billable token quantities are identical.
Example 2: How efficiency can make the same token price cheaper
Now consider a purely illustrative workflow where Sonnet 5 needs 200,000 input tokens across its agent loop and produces 20,000 output tokens.
Hypothetical Sonnet 5 task
200K input = $0.40
20K output = $0.20
Total = $0.60
If a Sonnet 5.5 run happened to require 30% fewer input and output tokens on that same workload, its hypothetical usage would become 140,000 input and 14,000 output tokens.
Illustrative Sonnet 5.5 task
140K input = $0.28
14K output = $0.14
Total = $0.42
The rate card never changed. The total fell from $0.60 to $0.42 because the hypothetical 5.5 run consumed less billable work.
This example demonstrates the mathematics only. It is not an Anthropic benchmark result and does not imply that every Sonnet 5.5 request will use exactly 30% fewer tokens.
Why can Sonnet 5.5 use fewer tokens to finish the same work?
Anthropic's launch material points to several forms of improved efficiency rather than one single mechanism.
- Fewer steps
- Early testers reported that Sonnet 5.5 could reach successful outcomes in fewer intermediate steps than Sonnet 5 on some workflows.
- Fewer tool calls
- Anthropic says Sonnet 5.5 batched tool calls together more frequently in head-to-head coding tests, reducing the number of agent turns.
- Less unnecessary searching
- One early tester cited by Anthropic said the previous model was more likely to reach for web search unnecessarily, while 5.5 showed better judgment and used fewer output tokens.
- More efficient reasoning
- A model that finds the correct path earlier can spend fewer reasoning and output tokens exploring alternatives that do not contribute to the final result.
- Fewer retries
- In production systems, a higher first-pass success rate can reduce the number of complete API calls required before the result is accepted.
The economic effect compounds in an agent because every extra round trip can cause the conversation history and tool results to be sent back as input again.
Effort settings can matter more than the model-name price
Claude's effort level changes how much work the model does.
Anthropic says lower effort makes Sonnet 5.5 answer faster and use fewer tokens, while higher effort lets it reason for longer and check its work more thoroughly.
This means two applications using exactly the same model can have very different cost per request.
| Effort | Typical economic effect | Useful for |
|---|---|---|
| Low | Less model work, lower latency and lower token use. | Straightforward repetitive tasks where the success criteria are clear. |
| Medium | Balances quality, speed and token spend. | Normal interactive work and many routine agentic workflows. |
| High | More reasoning and checking, normally with higher task cost than lower settings. | Harder coding, analysis and agent work. |
| Xhigh / Max | Can substantially increase model work and reduce the apparent cost advantage. | Workloads where extra capability is worth higher latency and token spend. |
Anthropic says Claude Code and the Claude apps use Medium effort by default for Sonnet 5.5, while the Claude Platform defaults to High.
That means an API developer and a Claude app user can experience different efficiency even if they think they are simply comparing “Sonnet 5.5.”
Anthropic also reports that on several benchmarks, Sonnet 5.5 at Low or Medium effort exceeded Sonnet 5's best score at a fraction of the measured task cost. At higher effort settings, the cost advantage can narrow.
Source: Anthropic Sonnet 5.5 cost, speed and effort discussion .
Tool calls can make efficiency savings much larger than they first appear
An agentic application does not pay only for the final paragraph Claude produces.
Claude API tool requests can contain the tool definitions, tool-use blocks, tool-result blocks, the previous conversation and any newly generated thinking or text.
If a model takes six tool turns where another model can finish in four, the difference is not merely two small tool calls. Later requests can carry increasingly large conversation histories.
Anthropic's pricing documentation explains that tool use is billed through the normal input and output token counts, while some server-side tools can also have separate usage-based charges.
Agent economics
Fewer steps can save repeated context, not only output tokens
If an agent avoids an unnecessary tool call, it may also avoid sending that tool's result and the expanded conversation back through another model turn. This is why task-level efficiency can matter much more than comparing a single response's visible word count.
Pricing reference: Claude Platform pricing documentation .
Prompt caching can reduce repeated-input costs further
Prompt caching is separate from the Sonnet 5.5 efficiency claim.
Instead of repeatedly charging the full input rate for the same large prompt prefix, document, instructions or conversation context, a subsequent request can read previously cached content at a much lower rate.
Anthropic's Sonnet 5.5 launch table lists:
| Operation | Price per 1M tokens | Relative to ordinary input |
|---|---|---|
| Normal input | $2.00 | Base rate |
| Cache write | $2.50 | Higher initial cost to store reusable prompt content |
| Cache read | $0.20 | 10% of the ordinary $2 input-token rate |
This can be particularly useful for agents that repeatedly reuse a large system prompt, tool definitions, repository context or document set.
Anthropic's general pricing documentation currently defines a standard five-minute cache write at 1.25× base input pricing, a one-hour write at 2×, and normal cache reads at 0.1× for most models.
At this article's September 29 review time, Anthropic's Sonnet 5.5 launch page already lists the model's cache write and read prices, while some of its broader Platform pricing tables had not yet added Sonnet 5.5 as a separate model row. Verify the current pricing page before budgeting for less-common cache configurations.
What about the 50% Batch API discount?
Anthropic's general Message Batches documentation says batch processing charges usage at 50% of standard API prices for supported models.
That feature can be important for workloads that do not require an immediate response, such as large offline evaluations, document processing or bulk analysis.
However, there is an important launch-day documentation caveat for Sonnet 5.5.
At the time this article was reviewed, Anthropic's detailed Batch API model table still listed Sonnet 5 rather than a dedicated Sonnet 5.5 row.
Do not assume silently
Verify Sonnet 5.5 Batch support before building your budget around it
The general Batch API discount is 50% for supported usage, but Anthropic's detailed documentation was still catching up with the newly released Sonnet 5.5 model when BeeVaults reviewed it on September 29.
Reference: Claude Message Batches documentation .
Faster output does not directly lower the token price, but it can still matter
Anthropic says Sonnet 5.5 generates output 30%+ faster than Sonnet 5 and calls it its fastest Sonnet model to date.
A faster token stream does not itself make each token cheaper. If the token counts remain identical, the bill remains identical.
But speed can still affect application economics.
Faster completion can allow interactive products to use a stronger effort setting while remaining responsive, reduce waiting during multi-turn agent loops and improve how quickly high-volume workloads move through a system.
For a commercial product, developer time and user waiting time can also matter even though those costs do not appear directly on the Claude API invoice.
What did early Sonnet 5.5 testers report about token efficiency?
Anthropic's announcement includes several third-party observations that help explain why task cost can diverge from the unchanged rate card.
These are individual tester results rather than guarantees for every application, but they show the kinds of savings Anthropic is referring to.
| Tester | Reported observation | Why it affects cost |
|---|---|---|
| Slack | About 14% fewer output tokens across its offline Slackbot evaluations while improving on almost all of those evaluations. | Lower output-token volume directly lowers output spend at the same $10/MTok rate. |
| Box | Reported 12% fewer total tokens and 2.4× faster performance than Sonnet 5 in its evaluation. | Fewer total tokens reduce billed model usage even without a rate cut. |
| Balyasny Asset Management | Reported around 121K tokens per answer on one private finance suite versus 497K for Sonnet 5. | A large reduction in tokens can dominate the economics of high-volume workloads. |
| Lovable | Reported roughly one-third fewer tool calls and about half as many shell runs in its coding evaluations. | Fewer agent steps can reduce repeated context and tool-result tokens. |
These results should be interpreted as workload examples, not universal percentage discounts. Different prompts, tools, contexts and acceptance criteria can produce very different economics.
All tester observations above are attributed in Anthropic's Sonnet 5.5 launch announcement .
When might Sonnet 5.5 not be noticeably cheaper?
There are several situations where the headline efficiency claim may not translate into a large invoice difference.
- Very short requests
- A tiny one-turn prompt leaves little room for step-count or tool-use efficiency to create a meaningful absolute saving.
- Identical token usage
- If both models consume essentially the same input and output volume, the normal token price is the same.
- Higher effort
- Raising effort can cause Sonnet 5.5 to reason longer, which can absorb some of the efficiency advantage.
- Output-dominated tasks
- If you explicitly require a fixed very long output, both models may need to generate nearly the same number of expensive output tokens.
- Tool-heavy fixed workflows
- If the workflow requires exactly the same tool sequence regardless of model intelligence, there may be fewer opportunities for 5.5 to eliminate steps.
How to tell whether Sonnet 5.5 is actually cheaper for your application
Do not compare one impressive response from each model. Use a repeatable evaluation set that reflects the work your application actually performs.
-
Use the same task set. Run Sonnet 5 and Sonnet 5.5 against the same representative prompts and data.
-
Keep the acceptance criteria constant. A cheaper response that fails the task is not a meaningful saving.
-
Record input and output usage. Capture the token counts returned in the Claude API usage fields.
-
Count tool calls and retries. These are often where agentic cost differences accumulate.
-
Measure latency. Sonnet 5.5's reported speed improvement may be valuable even when token spend is close.
-
Calculate cost only for successful tasks. Divide total spend by the number of accepted completions, not merely by API requests.
Useful production metric
Cost per accepted task
This captures tokens, retries and success rate in one useful number. It is much more informative than comparing the published $2 and $10 token prices, because those prices are identical for Sonnet 5 and Sonnet 5.5.
Sonnet 5.5 vs Sonnet 4.6 is a different pricing comparison
The “same price” conclusion applies specifically to Sonnet 5.5 versus Sonnet 5.
Sonnet 4.6 was priced at $3 per million input tokens and $15 per million output tokens, so Sonnet 5 and 5.5 have lower headline per-token rates than that older generation.
There is another wrinkle: Anthropic's pricing documentation says Claude 4.7 and later models use a newer tokenizer that can produce approximately 30% more tokens for the same text than models using the older tokenizer, including Sonnet 4.6.
Therefore a Sonnet 4.6-to-5.5 cost comparison should not assume that a 33% reduction in the published token rate produces an identical 33% reduction in the bill for the same text.
Source: Anthropic pricing documentation .
Frequently asked questions
How much does Claude Sonnet 5.5 cost?
Anthropic launched Sonnet 5.5 at $2 per million standard input tokens and $10 per million output tokens. Its launch table also lists $0.20 per million cache-read tokens and $2.50 per million cache-write tokens.
Is Sonnet 5.5 cheaper per token than Sonnet 5?
No. Sonnet 5 and Sonnet 5.5 both use the same standard $2 input and $10 output price per million tokens.
Then why does Anthropic say Sonnet 5.5 costs less?
Anthropic says Sonnet 5.5 typically requires fewer tokens to complete the same work. In its testing, that translated to up to 30% lower cost per task. The saving comes from lower total usage, not a lower basic token rate.
Does every Sonnet 5.5 task cost 30% less?
No. “Up to 30% less” is an Anthropic test result, not a fixed discount. Your actual difference depends on prompts, effort, tool usage, retries, caching and workload complexity.
Was Sonnet 5 supposed to become more expensive?
Yes. Anthropic originally planned to move Sonnet 5 from its introductory $2/$10 rate to $3/$15 on September 1, 2026. On August 10 it cancelled that increase and made $2/$10 permanent.
How much does a million cached Sonnet 5.5 tokens cost to read?
Anthropic's Sonnet 5.5 launch table lists cache reads at $0.20 per million tokens.
Does higher effort make Sonnet 5.5 more expensive?
It can. Anthropic says higher effort makes Claude reason longer and check its work more thoroughly, while lower effort uses fewer tokens and answers faster. Total task cost therefore depends on the selected effort level.
What effort level does Sonnet 5.5 use by default?
Anthropic says Claude Code and its apps default Sonnet 5.5 to Medium effort, while the Claude Platform defaults to High.
Is Sonnet 5.5 faster than Sonnet 5?
Anthropic says Sonnet 5.5 generates output more than 30% faster than Sonnet 5, making it its fastest Sonnet model at launch.
Should I migrate just to save money?
The best answer depends on your own workload. Compare accepted-task rate, token usage, tool calls, retries and latency across a representative evaluation set. The published rate card alone cannot tell you which model is cheaper per successful task because the rate is the same.
The answer: same rate card, potentially lower bill
Claude Sonnet 5.5 did not launch with a cheaper standard API token price than Sonnet 5.
Both models cost $2 per million input tokens and $10 per million output tokens.
The economic improvement Anthropic is advertising is efficiency. Sonnet 5.5 is intended to use fewer tokens, fewer agent steps and fewer unnecessary tool calls to reach a successful result.
In Anthropic's testing that translated to up to 30% lower cost per completed task, while early testers reported a range of token and step reductions on their own workloads.
Whether that saving appears in your bill depends on the task. For a short request where both models consume the same tokens, there is effectively no pricing difference. For a long agent workflow where 5.5 avoids several reasoning or tool turns, the difference can become meaningful.
The useful metric is therefore not price per million tokens. For an application deciding whether to migrate, measure cost per successful task.