Claude API pricing per million tokens in 2026, and what a month of assistant use costs

Claude API pricing per token in 2026 depends on the model, and Anthropic quotes every price per million tokens, written as MTok. As of November 2026, Anthropic's pricing page lists Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, Claude Sonnet 5.5 at $2 and $10, Claude Haiku 5.5 from $0.10 and $0.50, and Claude Fable 5.1 at $10 and $50. Batch requests cost half, and a prompt read back from the cache costs between 2.5% and 10% of the input price, depending on the model.

This guide puts the current and legacy Claude models in two tables, explains the discounts and surcharges that change a bill, says what happened to Claude 3.5 Sonnet and Claude 3.7 Sonnet, and turns tokens into a monthly cost for light, regular and heavy desktop assistant use. All prices come from Anthropic's pricing documentation and claude.com/pricing as we read them on November 6, 2026, in US dollars. We ran no tests and took no screenshots, so open the linked page before you budget.

We make Barside, a free desktop app for Windows, Linux and Mac that has no AI of its own and runs on the Claude account you connect. The last sections show where it fits into the bill. If you still need a key, how to get an Anthropic Claude API key covers the Console steps and the spend limit.

Claude API pricing per token in 2026, model by model

Claude API pricing per token in 2026 comes down to two numbers for each model: what you pay for the tokens you send (input) and what you pay for the tokens Claude writes back (output). On every model in the tables below, output costs five times as much as input. Anthropic's documentation defines the unit plainly: "MTok: Million tokens. $5 / MTok is $5 for every million tokens."

US dollars per million tokens, from Anthropic's pricing documentation, as of November 2026.

These four are the models claude.com/pricing groups under "Latest models". Anthropic's models overview describes Fable 5.1 as the model "For demanding reasoning and long-horizon agentic work", Opus 5.5 "For long-running agentic coding and knowledge work", Sonnet 5.5 as a combination of speed and intelligence, and Haiku 5.5 as the model for "high-volume, latency-sensitive tasks such as classification, extraction, and routing". Its advice if you can't decide is this: "If you're unsure which model to use, start with Claude Opus 5.5 for most workloads." All four list a 1M token context window and up to 128K tokens of output.

Claude Mythos 5.1 has the same prices as Fable 5.1, but the pricing page marks it as limited availability, "Available only to organizations verified through Anthropic's verification programs".

Older Claude models you can still call

The models overview lists nine "Legacy models (still available)", and the pricing page gives their rates.

US dollars per million tokens, as of November 2026.

Two details matter if you choose a model by name. Claude Opus 5.5 is priced below Opus 5 and Opus 4.5 to 4.8, at $4 and $20 against $5 and $25, so the newer Opus is the cheaper one per token. Claude Sonnet 4.5 also still has a row at $3 and $15, but Anthropic has deprecated it, and it retires on November 30, 2026.

Per-token prices also don't compare one for one across generations. The pricing page says Claude 4.7 and later models use a newer tokenizer that "produces approximately 30% more tokens for the same text", while Claude Sonnet 4.6 and earlier models use the previous one. The models overview puts 1M tokens at roughly 555k words on the current tokenizer. The same note is therefore likely to count as more tokens on Sonnet 5.5 than on Sonnet 4.6, and Anthropic's page says the exact increase depends on the content.

What else changes the bill: caching, batches, search and long prompts

The base price is where a bill starts. As of November 2026, several features raise or lower what a request costs, and the pricing page says the caching multipliers stack with other modifiers, including the Batch API discount and data residency.

• Prompt caching. Writing a prompt prefix to the 5-minute cache costs 1.25 times the base input price, and writing to the 1-hour cache costs 2 times. Reading it back, a cache hit, costs 0.1 times the input price on most models, 0.05 times on Claude Opus 5.5 and Sonnet 5.5, and 0.025 times on Claude Fable 5.1 and Mythos 5.1. Anthropic's page says that at the standard rates, caching "pays off after one cache read for the 5-minute duration".
• Batch processing. The Batch API takes 50% off input and output tokens. You trade time for it: Anthropic's batch processing page says most batches finish "in less than 1 hour", and a batch expires if processing does not complete within 24 hours. It suits jobs that can wait, such as summarizing a backlog of documents overnight.
• Long prompts. Claude 4.6 and later models, except Haiku 5.5, include the full 1M token context window at standard pricing, and the page gives the example that "A 900k-token request is billed at the same per-token rate as a 9k-token request." Haiku 5.5 is the exception: a prompt over 100,000 tokens pays its higher prices, even when part of it is a cache hit.
• Web search. The web search tool costs "$10 USD per 1,000 searches", plus standard token costs, because search results count as input tokens. A search that fails with an error is not billed, and the web fetch tool adds no charge beyond tokens.
• Tools. When a request includes tools, their names, descriptions and schemas count as input tokens, and the API adds a tool use system prompt, listed at 286 tokens for Opus 5.5.
• US-only inference. For Claude 4.6 and later models, setting inference_geo to "us" multiplies every token price by 1.1.
• Fast mode. Faster output from Opus costs more: $8 per MTok of input and $40 of output on Opus 5.5, and $10 and $50 on Opus 5 and Opus 4.8, where fast mode is in research preview. Fast mode is not available with the Batch API.

New accounts get a little room to try all of this. The pricing page's FAQ says "New users receive a small amount of free credits to test the API."

Were Claude API prices different in October 2026?

If you're looking up Anthropic Claude API pricing for input and output in October 2026, start from the tables above, with one caveat: Anthropic's pricing pages show today's prices and no dated price history. We could not confirm from Anthropic's pages what each model listed in October, so we don't reproduce an October table.

The pricing page does record one dated change. It says the $2/$10 per million input/output token pricing for Claude Sonnet 5, "announced at launch as introductory pricing through August 31, 2026, is now the standard price", and that "The previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur." If you planned an October budget for Sonnet 5 at $3 and $15, Anthropic's page says that rise never took effect.

The deprecations page adds a date from the same weeks. On September 30, 2026, Anthropic told developers using Claude Sonnet 4.5 that it retires on the Claude API on November 30, 2026, and named claude-sonnet-5-5 as the replacement. For what you used in a past month, the Usage page in the Claude Console is your own record, and the deprecations page says the Usage page's Export button downloads a CSV with usage broken down by API key and model.

What happened to Claude 3.5 Sonnet and Claude 3.7 Sonnet?

Both are retired on the Claude API, and the model Anthropic now recommends in their place is Claude Sonnet 5.5. Anthropic's model deprecations page defines retired as "no longer available for use. Requests to retired models fail."

Claude 3.5 Sonnet

Claude 3.5 Sonnet pricing per million tokens no longer appears on Anthropic's pricing pages, and both Claude 3.5 Sonnet models are retired. The deprecations page, which now calls them the Claude Sonnet 3.5 models, says Anthropic notified developers on August 13, 2025 and retired claude-3-5-sonnet-20240620 and claude-3-5-sonnet-20241022 on October 28, 2025. Its recommended replacement for both is claude-sonnet-5-5, listed at $2 per million input tokens and $10 per million output tokens as of November 2026. We found no price for Claude 3.5 Sonnet on Anthropic's current pricing pages, so we don't quote an old one here.

Claude 3.7 Sonnet

Anthropic's Claude 3.7 Sonnet pricing per million tokens has also left its pricing pages, so Claude 3.7 Sonnet pricing today means the price of its replacement. The deprecations page lists the model as Claude Sonnet 3.7: Anthropic announced on October 28, 2025 that claude-3-7-sonnet-20250219 would retire, and it was retired on February 19, 2026. The recommended replacement is again claude-sonnet-5-5, at $2 and $10 per MTok as of November 2026, and a request to claude-3-7-sonnet-20250219 on the Claude API fails.

Old models and their replacements

A few more points help if you are moving off an old model:

• Partner clouds keep their own dates. The deprecations page says its dates apply to Anthropic-operated platforms, and that Claude Models on Google Cloud (Google managed) and Amazon Bedrock "set their own retirement schedules". We did not read those platforms' pages, so we could not confirm whether Claude 3.5 or 3.7 Sonnet is offered there today, or at what price.
• Some retired models keep a price row. Anthropic's pricing page still lists Claude Opus 4.1 and Opus 4 at $15 and $75, and Claude Sonnet 4 at $3 and $15, each marked retired except on certain partner clouds. We found no row for Claude 3.5 Sonnet or Claude 3.7 Sonnet.
• Count tokens on the new model. Sonnet 5.5 uses the newer tokenizer, which the pricing page says produces approximately 30 percent more tokens for the same text than earlier models. Count a few of your real prompts on it before you compare old and new costs.
• Expect notice next time. The deprecations page says Anthropic notifies customers with active deployments at least 60 days before it retires a publicly released model, and that impacted customers are notified by email and in the documentation.

What a month of desktop assistant use costs

A desktop AI assistant usually calls the model several times for each thing you ask, because it reads your request, picks a tool, looks at the result and writes again. In Barside, the assistant works in a loop of up to 40 steps per request, and each step is another exchange with the model that carries Barside's list of tools and any matching note excerpts. That makes the cost of one request larger than the length of your question suggests.

On Sonnet 5.5, one such request costs 24,000 × $2 / 1,000,000 = $0.048 for input and 1,600 × $10 / 1,000,000 = $0.016 for output, or $0.064 in total. The table repeats that arithmetic for three made-up months.

• Light. Example: Hana, a freelance translator, asks the assistant for about five things a workday, such as turning a note from a client call into dated tasks. In this example her month costs $6.40 on Sonnet 5.5 and $0.32 on Haiku 5.5.
• Regular. Example: Tom, a project manager, sends about 20 requests a workday, from meeting notes to follow-up drafts. His month comes to $25.60 on Sonnet 5.5 and $51.20 on Opus 5.5.
• Heavy. Example: Ada, a solo founder, runs research and agents through about 60 requests a workday. Her month is $76.80 on Sonnet 5.5 and $153.60 on Opus 5.5. If her requests also run 200 web searches, add $2 in search fees at $10 per 1,000, plus the search results as input tokens.

Caching could lower these numbers in an app that uses it. If 4,000 of the 6,000 input tokens in each call were cache hits on Sonnet 5.5, those 4,000 tokens would cost $0.0004 per call instead of $0.008, after a first write at $2.50 per MTok. Whether an app caches its prompts is up to that app, so the table leaves caching out. The batch discount fits an assistant poorly, since you wait on its answer and a batch can take up to 24 hours.

A Claude plan is the flat-price way to pay. As of November 2026, claude.com/pricing lists Pro at $17 per month "with annual subscription discount ($200 billed up front)" or $20 if billed monthly, Max at $100 per month and Ultra at $200 per month, with tax not included. Our guide to AI research assistant pricing shows how to find the break-even point between a flat plan and paying per use, and AI note-taking app pricing in 2026 does the same for notes apps.

How Barside fits into a Claude bill

Barside charges nothing for AI. It's free on Windows, Linux and Mac, it has no AI model of its own, and it includes no free AI use: the assistant, AI search, agents and Quick AI all run on the account you connect in Settings → AI. You can connect Claude in two ways.

• A Claude API key. Barside opens Anthropic's key page, picks the new key up from the clipboard and saves it encrypted on your computer. Anthropic then bills each request at the prices above. The Claude API key guide walks through every step.
• My plan. From Barside 1.59, the Claude card offers My plan, which runs Claude Code on your computer, signed in with your Claude plan, so that plan's limits apply instead of a bill per use. Anthropic hasn't said that apps may use a plan this way, so Barside's guides call an API key the safer choice. Web search inside Barside and cloud agents still need an API key.

Barside's guides say a Claude connection starts on Claude Opus 5, and Choose a model under the buttons switches it. As of November 2026, Anthropic lists Opus 5 at $5 and $25 per MTok and Opus 5.5 at $4 and $20, so check which model is selected before you compare your bill with the tables here. The line under the assistant's name shows the AI and the model in use. With a Claude key, the assistant searches the web through Anthropic's web search, which Anthropic bills on your account at the rate in the list above.

1. Download Barside for Windows, Linux or Mac, and open it.
2. Open Settings → AI and choose Claude. Make a key on the page Barside opens and copy it, or choose My plan on the Claude card to use your Claude plan.
3. Pick a model with Choose a model. A smaller model keeps quick requests cheap, as the tables show.
4. Under AI spending limit, turn on Limit what Barside's AI spends and set an amount for a day, a week (from Monday) or a month.
5. Press Ctrl+Shift+A, type what you need in the box and press Run.
6. After a few days, compare what Settings shows you spent with the Usage page in the Claude Console.

Set a spending limit, and know where it stops

Barside's AI spending limit caps what its AI spends on your computer. You choose an amount from US$0.50 to US$10,000 for a day, a week or a month, and Barside adds up an estimate of each answer's cost from the AI company's list prices for your model. Settings shows what was spent, how many requests that took and when the count starts again, and Reset the count starts it from zero.

• At 80%. Barside shows a notification, and the assistant and Library show a banner.
• At the limit. The assistant, agents, Ask library, Quick AI and the Space helper say You reached your AI spending limit instead of sending a request, until you raise the limit, turn it off or the count starts again.
• What it doesn't see. The count is an estimate, each computer counts only its own use, and cloud agents aren't counted. Set a spend limit in the Claude Console as well, and the key guide has those steps.

Barside Plus is optional, at US$20 a month or US$200 a year plus sales tax. It adds cloud sync, encrypted cloud backup and cloud agents, and it has no AI credits inside it, so your Claude usage is billed by Anthropic either way.

On privacy, Barside keeps your Claude key encrypted on your computer and sends it only to Anthropic. For work on your computer, requests go straight to Anthropic with no Barside server in between. Cloud agents are the one exception: turning them on sends your key to Barside's server, which seals it and uses it only to run your scheduled agents. Read the AI you connect and your data for the details, and AI notes apps where you bring your own key for what stays put if you later switch to another model or company.

Works well with

These guides cover the key itself, the assistant that spends it, and the wider question of paying per use or per month.