Claude API Pricing Explained (2026)
What the Claude API costs per month for a chatbot, a code reviewer and a document pipeline. List prices checked on 15-10-2026.
Published 15-10-2026. List prices last verified on 15-10-2026 and reviewed monthly.
Quick answer
This page is for the developer or small product team that is about to put Claude behind a feature and needs a monthly number before anyone approves the budget. The API bills per million tokens (MTok), with separate rates for input and output. On 15-10-2026 the Anthropic pricing page lists Claude Sonnet 5.5 at $2 input and $10 output, Claude Opus 5.5 at $4 and $20, and Claude Haiku 5.5 at $0.10 and $0.50 for prompts up to 100,000 tokens. Our three worked workloads below land between $5.50 and $450 a month, and the model you pick moves the bill far more than any other choice.
How we checked the information in this article
Our gate for this page: a price goes in only if we read it on the Anthropic pricing page on 15-10-2026, and every monthly figure we derive is labelled as calculated by us from list prices. We have not measured these workloads on a real invoice, so none of the costs below is a measured bill, and we say so each time. The gate keeps out every price that appears only on third-party blog posts, and it keeps Claude Haiku 3.5 out of the workloads, because the Anthropic rate limit page marks it as retired except on Google Cloud.
We sell skills that load into Claude Code, Cursor, Codex CLI and Copilot, and we name a few of ours where they help you control cost. No vendor paid to be listed here.
What does Claude cost per month for a chatbot, a code reviewer and a document pipeline?
Three workloads, with assumptions stated so you can swap in your own. All figures are calculated by us from the list prices on 15-10-2026 and assume the token counts shown. Models from Claude 4.7 onward use a newer tokenizer that Anthropic says produces about 30% more tokens for the same text, so count tokens with the model you plan to use.
| Workload and assumptions | Model | Monthly cost (calculated from list prices) | With the lever (calculated from list prices) |
|---|---|---|---|
| Support chatbot: 45,000 requests a month, 3,000 input tokens (2,000 of them a fixed system prompt), 400 output tokens | Haiku 5.5 | $22.50 | $16.39 with prompt caching |
| Same chatbot | Sonnet 5.5 | $450.00 | $320.47 with prompt caching |
| Code reviewer: 2,200 reviews a month, 25,000 input tokens (3,000 of them a fixed prompt), 2,000 output tokens | Sonnet 5.5 | $154.00 | $149.38 with prompt caching |
| Same code reviewer | Opus 5.5 | $308.00 | $298.76 with prompt caching |
| Document pipeline: 10,000 documents a month, 8,000 input tokens, 600 output tokens | Haiku 5.5 | $11.00 | $5.50 with the Batch API |
| Same pipeline | Sonnet 5.5 | $220.00 | $110.00 with the Batch API |
Our judgment from these rows: for a chatbot, start on Haiku 5.5 and move up only when you can show it fails your test set, because the same traffic costs twenty times more on Sonnet 5.5. For a code reviewer, caching saves only 3% here because the repeated prompt is 12% of the input, so we would spend the effort on model choice and on trimming the diff you send. Code review skills cover the review side of that workload. For a pipeline nobody waits on, the Batch API is the first thing we reach for.
If you want the same arithmetic on your own traffic, the Token Cost Estimator skill projects API cost from prompts, datasets and traffic volume, according to its listing. We have not run it against these workloads, so treat its output as a second opinion, not a check.
How does Claude API pricing work per model?
You pay for tokens, counted separately for what you send (input) and what the model writes back (output). Output costs five times input on the current models. The table shows the four current models as listed on 15-10-2026, in USD per MTok.
| Model | Input | Output | Cache read | Batch input / output |
|---|---|---|---|---|
| Claude Fable 5.1 | $10 | $50 | $0.25 | $5 / $25 |
| Claude Opus 5.5 | $4 | $20 | $0.20 | $2 / $10 |
| Claude Sonnet 5.5 | $2 | $10 | $0.10 | $1 / $5 |
| Claude Haiku 5.5 | $0.10 up to 100,000 prompt tokens, $0.50 above | $0.50 up to 100,000, $2.50 above | $0.01 up to 100,000, $0.05 above | $0.05 / $0.25 up to 100,000 |
Why these four and not the older ones: the pricing page marks them as the current line, and the older Opus 4 and 4.1 models cost $15 input and $75 output, so we would not start a new build on them. Three details change a bill without showing in the table. Haiku 5.5 charges the higher rate for the whole request once the prompt passes 100,000 tokens, and cache reads and writes count toward that length. Setting inference_geo to "us" multiplies every token category by 1.1 on Claude 4.6 and later. And web search costs $10 per 1,000 searches on top of tokens. Check the Anthropic pricing page (listed under Sources) before you commit, because we only vouch for the date above.
The pricing page also lists a fast mode for Opus models, in research preview, at $8 input and $40 output for Opus 5.5. We have not tested it, and it cannot be combined with the Batch API, so we leave it out of the workloads.
How much do prompt caching and the Batch API save?
Caching charges 1.25 times the input price to write a prefix for five minutes, or 2 times for one hour, and then a fraction of the input price on every read. On Sonnet 5.5 a read costs $0.10 against $2 for fresh input, which is 5%. The pricing page says a five-minute write pays for itself after one read. The Batch API takes 50% off both input and output, and the page states that these discounts stack with caching. Our prompt caching guide goes deeper into when a cache write pays off.
What that means in our calculated rows: 27.2% off the Haiku chatbot and 28.8% off the Sonnet chatbot from caching a 2,000-token prompt, 50% off the pipeline from batching, and 3% off the code reviewer. These are calculated savings, not measured ones. We assumed that the cache is rewritten at most once every five minutes of the day, which is the worst case for a bot with steady traffic. If your traffic is bursty, your hit rate will be lower and your saving smaller. For a deeper cost-engineering walk through, our guide on reducing Claude Code token usage covers the same levers inside a coding agent, and skills that reduce Claude Code costs lists the coding-session levers.
If you run Claude through an agent for your team, the Usage Budget skill adds loop and retry caps plus a weekly usage report, per its listing, which addresses the failure we worry about most: an agent that retries itself into a large bill. Browse more cost and evaluation tooling in our LLM ops skills.
How do billing, spend limits and alerts work?
You are billed on usage, and the Anthropic rate limit page sets a monthly spend cap per tier: $500 for Start, $1,000 for Build and $200,000 for Scale, with no cap on the Custom tier. You can view the cap and set your own lower limit on the Billing page in the Claude Console, the same place where you create a key (see how to get a Claude API key). When you hit the cap, requests return HTTP 429 until 00:00 UTC on the first day of the next month, and the response has no retry-after header, so automatic retries just fail. For us that is the detail to build around: treat enforced_spend_limit_reached as an outage, not a rate limit. We did not verify email alerts or per-workspace budgets on 15-10-2026, so check the Console for those.
New organizations may start in an Evaluation tier with limits below the standard ones. We could not confirm from the page how long that lasts. If you route Claude through AWS or Microsoft Foundry, the invoice comes from them in Claude Consumption Units at $0.01 each, per the pricing page. Bedrock and Google Cloud have their own regional pricing, and the pricing page says regional endpoints carry a 10% premium over global ones, so compare your cloud contract before you assume the first-party rates apply. Our Bedrock, Vertex AI and Foundry comparison covers the cloud routes. When the question is how to book and report the spend, see finance and accounting skills. Teams that already own a cloud bill often belong with DevOps and cloud skills that keep that spend visible.
Is the API cheaper than a Claude subscription?
It depends on who is using it. The Claude pricing page lists subscription prices in euros, while the API prices on this page are in US dollars. On 15-10-2026 it lists Pro at €15 a month on the annual plan (€18 billed monthly) and Max from €90 a month, and says Claude Code is included in paid plans. Max 5x and Max 20x also list $100 and $200 a month in API credits. When a paid plan hits its limit you can turn on usage credits to continue at standard API rates. Our rule: one person working by hand in Claude Code belongs on a subscription, and anything that runs without a person, such as the three workloads above, belongs on the API. We have not compared the two on identical work, so we do not claim a break-even point. See our comparison of AI coding agent costs and the Claude Code pricing explainer for the subscription side.
What is the cheapest way to run production traffic?
Our ranking, cheapest effective first, for a team that can accept some latency:
-
Batch everything nobody waits for. It is a flat 50% on input and output, which is why it sits first. It moves down if a user is waiting on the answer, because batches are asynchronous. Workflow automation skills help with the batch jobs themselves.
-
Pick the smallest model that passes your test set. The same chatbot is $22.50 on Haiku 5.5 and $450 on Sonnet 5.5. It moves up if you have no test set, since then you cannot prove the small model is good enough. An evaluation skill such as LLM Eval Harness, which runs repeatable test suites against prompts and models per its listing, is how we would build that set.
-
Cache the fixed prefix. It saved 27.2% to 28.8% in our chatbot rows, but only 3% on the reviewer. It moves up when your system prompt or tool definitions are long relative to the user message.
Work that runs through an API often belongs to a wider build, so look at the APIs and backend skills for the plumbing. In coding sessions, the token optimizer skill keeps raw test and build logs out of your context by returning compact summaries, per its listing.
What to do next
Take the workload closest to yours, replace our token counts with counts from your own logs, and recompute with the rates on the Anthropic pricing page. Then set a lower monthly limit in the Console than the tier cap. Prices change, so verify the figures here before you put them in a budget; this is cost guidance, not financial advice. If you are weighing Claude inside a desktop workflow instead of the API, our Claude Cowork pricing guide is the next read.
Written by Samuel Rose, Founder of Agensi. Next scheduled review: 15-11-2026.