Claude API Pricing (2026): Every Model's Token Cost

Claude API pricing is pay-as-you-go by the token, with no monthly fee: Claude Haiku 4.5 costs $1/$5 per million input/output tokens, Claude Sonnet 5 costs $3/$15, and Claude Opus 5 costs $5/$25. Output is billed at 5× input across every model, and batching or prompt caching can cut effective costs by up to 95%.

Rates verified 9 August 2026 against Anthropic's published pricing. Prices change when new models launch — re-check before you budget.

We run this site's content pipeline on the Claude API, so we watch these numbers closely. The good news for anyone budgeting a project: the whole system is per-token and transparent — no seat license, no minimum spend, no surprise tiers. This guide breaks down every model's per-token rate, the three discount levers that matter most, worked cost examples you can copy, and a simple rule for picking a model by budget.

How Claude API pricing works

Every request bills input and output tokens separately, per million tokens, with output at 5× input. There is no monthly floor — you pay only for what you consume.

Input tokens are your prompt, system message, and any documents; output tokens are everything the model generates. Anthropic publishes the current rates per model, and they change only when a new model launches.

Roughly four characters equal one token, so a 750-word page is about 1,000 tokens. To estimate a bill before you're charged, multiply expected input and output tokens by the model's rate; to track real spend, read the usage object returned in every API response, which reports exact input and output counts per call. That habit — estimate, then reconcile against usage — is how you keep costs from surprising you.

Here is the full table across the current lineup, including both discount columns:

Two things stand out. First, across the three mainstream tiers the spread is only 5× — Haiku 4.5 to Opus 5 — so choosing among them is about fit, not affordability; the premium Claude Fable 5 sits above at $10/$50 for the rare task that outruns Opus, stretching the top-to-bottom range to 10×. Second, Opus 5 costs exactly what Opus 4.8, 4.7, and 4.6 cost before it — four consecutive releases at the same rate, so upgrading the Opus line has been free every time. For a deeper tour of what each tier is built for, our Claude models explained guide covers the whole lineup.

The three levers that cut your Claude API cost

Batch processing takes 50% off, cached input reads cost 10% of standard rate, and the 1M context window carries no surcharge. Stacked, they matter far more than the base rates.

Batch processing takes 50% off every token when you submit jobs asynchronously and accept results within 24 hours. Any pipeline that isn't interactive — nightly summarization, bulk classification, report generation — should default to the batch API. Half price for changing one endpoint is the easiest win available.

Prompt caching is the bigger lever for chat and agent apps. When requests share a long, stable prefix — a system prompt, a codebase snapshot, a document under discussion — prompt caching lets you reuse it. Cache reads cost 10% of the standard input price, a 90% discount, while the one-time cache write costs 1.25× input for a 5-minute cache or 2× for a 1-hour cache. Agents re-send that prefix on every turn, so the saving compounds fast. One detail worth checking on Opus 5: the minimum cacheable prefix drops to 512 tokens, half what Opus 4.8 required, so short prompts you had written off as uncacheable may now qualify with no code change.

The 1M-token context window carries no surcharge. Claude Opus 5, Opus 4.8, and Sonnet 5 support the full million-token context at flat standard rates — a 900K-token request pays the same per-token price as a 9K-token request. You still pay for the tokens you send, so filling a 1M window with Opus costs about $5 in input before output; caching exists precisely so you don't pay that repeatedly for the same material.

Worked Claude API pricing examples

Five representative workloads with the arithmetic done, at list prices with no discounts unless the row says otherwise.

The pattern to internalize: input dominates on document-heavy tasks, output dominates on generation-heavy ones, and caching flips the math on anything conversational. The chatbot row shows a cached 20K-token prefix costing $0.006 to read instead of $0.06 to reprocess cold — a 90% cut on the largest part of the request. Run those same batches asynchronously and every figure in the table halves again.

One budgeting caveat specific to Opus 5: thinking is on by default, and max_tokens caps thinking plus response text together. A route that never set thinking explicitly will generate more output tokens than the same request did on Opus 4.8, so re-baseline your output estimates rather than carrying them across.

How to pick a Claude model by budget

Ask what a slightly wrong answer costs you. If mistakes are cheap to catch, buy speed and volume. If they compound invisibly, buy reasoning.

• Tight budget, verifiable outputs — classification, extraction, tagging, routing — use Claude Haiku 4.5 at $1/$5 and batch it. This is the cheapest token pricing on offer, and errors surface immediately.
• Everyday coding and writing — use Claude Sonnet 5 at $3/$15, against Opus at $5/$25. It delivers most of Opus's coding quality for 60% of the price, and the introductory rate makes it cheaper still through 31 August 2026.
• Hardest reasoning and large refactors — use Claude Opus 5. Reserve it for tasks where the failure cost justifies the premium.

Our standing rule is start on Sonnet, escalate to Opus only when Sonnet measurably fails, and downgrade to Haiku wherever outputs are mechanically verifiable. The Claude Sonnet vs Opus breakdown has the benchmark data and full cost math behind that call. If you weigh Claude against rival assistants on value rather than sticker rate, our Claude vs ChatGPT comparison shows where each one earns its price. Note that Claude Code — Anthropic's agentic coding tool — bills differently: it needs a paid subscription or API credits, as covered in What is Claude Code?.

Add-on and tool costs

Server-side tools bill on top of standard token rates. For most projects they're a rounding error; at scale they're the line teams most often forget to forecast.

Web search costs $10 per 1,000 searches, with the tokens for processing results billed separately at the model's normal input rate. Code execution includes roughly 1,550 free hours monthly per organization — and is free entirely when paired with web search or web fetch — then $0.05 per hour beyond that. Fast mode on Opus 5 runs the same model at higher output speed for double the token rate, and is a Claude API research preview only, unavailable on Bedrock, Google Cloud, or Microsoft Foundry.

Claude pricing at a glance

Subscriptions and API billing are separate systems — a Pro or Max plan never includes API credits, and API spend never touches your plan.

For the full breakdown of every plan, see our how much Claude costs guide.