Claude API Pricing 2026: Anthropic Per-Token Rates & Cost Calculator
Anthropic's API has three pricing quirks that catch people out — a cache you pay extra to write to, an introductory Sonnet 5 rate with an expiry date, and an Opus range where three generations cost exactly the same. Here is all of it, with a calculator.
Prices verified July 28, 2026 against platform.claude.com/docs/en/about-claude/pricing.
Read this first: the Sonnet 5 price you are paying is temporary
Anthropic lists Claude Sonnet 5 at $2.00 input / $10.00 output as an introductory rate running through August 31, 2026. After that it moves to $3.00 / $15.00 — the rate Sonnet 4.6 and 4.5 already charge.
That is a 50% increase on both input and output, landing on a date that is already in the calendar. Every Sonnet 5 forecast you build today needs a second column. There is one further down this page.
Claude API Cost Calculator
Estimate a monthly Anthropic bill, and pull in models from OpenAI, Google, xAI or DeepSeek to see what you would pay elsewhere on the identical workload. Runs entirely in your browser.
Models to compare (5 selected)
| Model | Input cost | Output cost | Per request | Monthly total |
|---|---|---|---|---|
| Claude Haiku 4.5AnthropicCHEAPEST | $88.80 | $160.00 | $0.03 | $248.80 |
| Claude Sonnet 5 (intro rate)Anthropic | $177.60 | $320.00 | $0.06 | $497.60 |
| Claude Sonnet 4.6Anthropic | $266.40 | $480.00 | $0.09 | $746.40 |
| Claude Opus 5Anthropic | $444.00 | $800.00 | $0.16 | $1,244 |
| Claude Fable 5Anthropic | $888.00 | $1,600 | $0.31 | $2,488 |
You've sized the bill. Does Claude actually recommend your product?
The same model you're paying per token also answers “what's the best tool for this?” all day long. Our free scan asks it five different ways and scores how often your product gets named back.
On this page
The Claude model ladder, and where the price actually steps
Anthropic's range is narrower than OpenAI's and the steps are larger. There are effectively four price points, not twenty, which makes the routing decision unusually clean.
Haiku 4.5 — $1.00 / $5.00
The cheapest current Claude, and the floor of the range. Note what it is not: at $1.00 input it is 20x the price of DeepSeek V4 Flash and 10x GPT-5 nano. Anthropic does not compete at the bottom of the market. Haiku is a "cheap Claude", not a cheap model.
Sonnet 5 — $2.00 / $10.00 (introductory), then $3.00 / $15.00
The workhorse tier, and the only model in Anthropic's lineup with a dated price change attached to it. Right now it is cheaper than the older Sonnet 4.6 and 4.5, which both sit at $3.00 / $15.00 — an inversion that disappears on September 1.
Opus 5 / Opus 4.8 / Opus 4.6 — $5.00 / $25.00, all three
This is the quiet one. Three Opus generations bill at exactly the same rate. Downgrading from Opus 5 to an older Opus saves you nothing, so the only reasons to pin a version are behavioural — an eval suite you have already run, or output stability you do not want to re-validate.
Fable 5 — $10.00 / $50.00
Top of the range at 2x Opus and 5x Sonnet 5's intro rate. At $50.00 per million output tokens, an output-heavy workload on Fable 5 will dominate every other line on your invoice combined. Reserve it for the small share of traffic that genuinely fails lower down.
Claude API price table, per 1M tokens
Standard synchronous tier, USD. The cached column is the cache read (hit) price — see the cache-write section below for what it costs to put the prefix there in the first place.
| Model | Input / 1M | Cache read / 1M | Output / 1M | Output ÷ input |
|---|---|---|---|---|
Claude Fable 5claude-fable-5 | $10.00 | $1.00 | $50.00 | 5.0x |
Claude Opus 5claude-opus-5 | $5.00 | $0.50 | $25.00 | 5.0x |
Claude Opus 4.8claude-opus-4-8 | $5.00 | $0.50 | $25.00 | 5.0x |
Claude Opus 4.6claude-opus-4-6 | $5.00 | $0.50 | $25.00 | 5.0x |
Claude Sonnet 5 (intro rate)claude-sonnet-5 | $2.00 | $0.20 | $10.00 | 5.0x |
Claude Sonnet 4.6claude-sonnet-4-6 | $3.00 | $0.30 | $15.00 | 5.0x |
Claude Sonnet 4.5claude-sonnet-4-5 | $3.00 | $0.30 | $15.00 | 5.0x |
Claude Haiku 4.5claude-haiku-4-5 | $1.00 | $0.10 | $5.00 | 5.0x |
Sonnet 5's introductory rate is on a clock — want the heads-up?
Leave an email and we'll write to you the day an Anthropic rate above actually moves, rather than making you diff the pricing docs yourself.
Already on the calendar: Claude Sonnet 5's introductory $2 / $10 per 1M tokens expires Aug 31, 2026 and goes to $3 / $15 — a 50% increase on the same workload. Prices here were last re-verified July 28, 2026.
The cache-write premium: Anthropic's biggest pricing quirk
On OpenAI and Google, prompt caching is close to free money — you pay the normal input rate the first time and a fraction of it thereafter. Anthropic charges you more than the normal rate to write the prefix into the cache, and the size of that premium depends on how long you want the cache to live.
| Cache tier | Write cost | Read cost | Break-even | Sonnet 5 write, per 1M |
|---|---|---|---|---|
| 5-minute TTL | 1.25x base input | 0.1x base input | 2 calls | $2.50 |
| 1-hour TTL | 2x base input | 0.1x base input | 3 calls | $4.00 |
The break-even column is the number that matters and it is computed, not asserted: over N calls reusing the same prefix, caching costs write + (N−1) × read against N for not caching. A 5-minute cache goes positive on the second call; a 1-hour cache on the third.
The failure mode is caching a prefix nobody reuses. A one-shot document pipeline where every request carries a different document has nothing cacheable in it — mark that prefix as cached and you have simply paid a 25% to 100% surcharge on every call for a cache that is never read. Cache the stable part (system prompt, tool definitions, reference corpus) and leave the variable part uncached, in that order, because the cacheable region has to be a prefix.
Which TTL to pick
- Interactive chat → 5-minute. Turns arrive seconds apart, so the prefix is re-hit far inside the window and you never pay the 2x write.
- Long agent sessions and spaced batches → 1-hour. Calls separated by minutes will keep letting a 5-minute cache expire, and you end up paying the 1.25x write over and over. Paying 2x once beats paying 1.25x five times.
- Traffic that genuinely arrives once → neither. Do not cache.
One more Anthropic-specific detail worth knowing: the cache read rate is 10% of base across the entire Claude range, including Fable 5 and Opus. On expensive models that makes caching a far larger absolute saving than it is on Haiku — $9.00 saved per million cached tokens on Fable 5 versus $0.90 on Haiku.
What September 1, 2026 does to a real Sonnet 5 bill
A coding-agent workload: 30,000 input tokens (repo context plus tool definitions), 4,000 output tokens of generated diff, 8,000 requests a month, 70% of input served from cache. No batch discount, because an agent is interactive. Same numbers, four models.
| Model | Input | Output | Monthly total |
|---|---|---|---|
| Claude Haiku 4.5 | $88.80 | $160.00 | $248.80 |
| Claude Sonnet 5 (intro rate) | $177.60 | $320.00 | $497.60 |
| Claude Sonnet 5 (from Sep 1, 2026) | $266.40 | $480.00 | $746.40 |
| Claude Opus 5 | $444.00 | $800.00 | $1,244 |
The same agent, unchanged, goes from $497.60 to $746.40 a month — $248.80 more, on a date nobody on your team has in their calendar. Scale that to a product doing ten times this volume and it is a five-figure annual swing that arrives with no code change and no traffic change.
Two sane responses. Either budget Sonnet 5 at $3 / $15 from the start and treat the current rate as a windfall, or use the runway to evaluate whether Haiku 4.5 handles a meaningful share of this traffic — at $248.80 a month it is 3.0x cheaper than post-September Sonnet 5, and on an agent workload that is mostly mechanical edits the quality gap may be smaller than the price gap.
Routing between Haiku, Sonnet and Opus
Because Claude's ladder is only 10x end to end, the classic advice — "route everything cheap and escalate rarely" — pays off less here than it does on a range that spans 600x. The Claude-specific version of the advice is different, and it is about where in the request you spend.
Cache before you downgrade
Moving Sonnet 5 → Haiku 4.5 saves 50% on input. Getting cache hits on 70% of your input tokens saves 63% on input without changing model. On Claude specifically, caching is the bigger lever, and it costs you no quality. Do it first.
Escalate on a signal, not on a hunch
Route to Opus when something concrete fires — a schema validation failure, a self-reported low confidence, a retry, a flagged high-value customer. "This feels like a hard one" is not a routing rule and in practice degrades into sending everything to Opus.
Watch the 5x output multiplier
Every Claude model charges 5x for output. Any prompt change that shortens the answer — asking for JSON instead of prose, dropping the "explain your reasoning" preamble, setting a max output length — is worth five times the equivalent saving on the input side.
Batch the offline half
The Message Batches API is 50% off both directions and stacks with cached input. Evals, backfills, bulk document processing and nightly summaries all qualify. This is the only lever on this list that halves the bill outright with zero quality cost.
Costs on the Claude API that are not tokens
The per-million table is not the whole invoice. Three things sit outside it, and agent-shaped applications hit all three.
- Server-side web search — billed at $10 per 1,000 searches, separately from the tokens the results then consume. A search-augmented agent pays for the search and for the much larger prompt the retrieved content produces.
- Cache writes — covered above, but worth restating as a line item: a cached workload's input bill is not simply "10% of before". It is 10% for the reads plus 125% or 200% for every write.
- Thinking tokens — Claude's extended reasoning bills at the output rate, at the 5x multiplier, whether or not you display it. A 300-token visible answer can carry several thousand billed tokens. Budget the output field in the calculator well above what your users will see.
Claude API pricing FAQ
How much does the Claude API cost per million tokens?
As of July 28, 2026: Claude Haiku 4.5 is $1.00 input / $5.00 output, Claude Sonnet 5 is $2.00 / $10.00 on its introductory rate, Claude Sonnet 4.6 and 4.5 are $3.00 / $15.00, Claude Opus 5 (and Opus 4.8 and 4.6) are $5.00 / $25.00, and Claude Fable 5 sits at the top at $10.00 / $50.00. Billing is per token with no subscription and no minimum spend.
When does the Claude Sonnet 5 introductory price end?
Anthropic lists Sonnet 5 at $2.00 input / $10.00 output as an introductory rate that runs through August 31, 2026, after which it moves to $3.00 / $15.00 — the same rate Sonnet 4.6 and 4.5 already charge. That is a 50% increase on both directions. If you are budgeting a Sonnet 5 workload past that date, model it at $3 / $15, not the rate you are paying today. This is the single most common way a Claude bill surprises a team in late 2026.
Why does Anthropic charge extra to write to the prompt cache?
Anthropic is unusual in charging a premium for the cache write itself rather than making it free. Writing a prefix into a 5-minute cache costs 1.25x the base input rate; writing into a 1-hour cache costs 2x. Reads (cache hits) then cost 10% of base. The arithmetic is friendlier than it sounds: a 5-minute cache is cheaper than not caching from the second call that reuses the prefix, and a 1-hour cache from the third. The trap is caching a prefix that gets reused zero or one time — then you have simply paid 25% to 100% extra for nothing.
Should I use the 5-minute or the 1-hour cache TTL?
Match the TTL to how often the prefix is genuinely reused. Interactive chat, where turns arrive seconds apart, is a 5-minute case — the cheaper write premium wins and the prefix is hit again well inside the window. A coding agent working through a long session, or a batch of documents all answered against the same large reference corpus with gaps between calls, is a 1-hour case: the 2x write is amortised over many more hits and you avoid repeatedly paying the write premium as a 5-minute cache keeps expiring. The wrong choice is a 1-hour cache on traffic that arrives once an hour anyway.
Is Claude Opus worth 2.5x the price of Sonnet?
On raw rates Opus 5 is $5.00 / $25.00 against Sonnet 5's introductory $2.00 / $10.00 — so 2.5x on both directions, and still 1.67x once Sonnet 5 moves to $3 / $15. The honest answer is that it depends entirely on whether the task fails at Sonnet. If Sonnet gets the task right first time, Opus is pure overspend. If Sonnet needs two attempts plus a human correction, Opus is cheaper in total cost of ownership even at 2.5x. The productive pattern in production is routing: Sonnet or Haiku by default, escalate to Opus only on a failed validation or a low-confidence signal.
Is there any reason to run Claude Opus 4.8 or 4.6 instead of Opus 5?
Not on price. Opus 5, Opus 4.8 and Opus 4.6 all bill identically at $5.00 input / $25.00 output as of July 28, 2026. Older Opus generations are not a budget option — they are the same money. The only reasons to pin an older Opus are behavioural: a prompt suite you have already evaluated against it, or output-format stability you do not want to re-validate.
Does a Claude Pro or Max subscription include API credit?
No. Claude Pro and Claude Max cover the Claude apps and Claude Code under those plans; the API is a separate product billed per token through the Anthropic Console, with its own payment method and its own spend limits. Buying Max does not reduce your API bill by a cent, and API usage does not consume your Max limits. Teams routinely discover this when a prototype built on a personal Max plan is moved to an API key and starts generating an invoice.
Does Claude have a free API tier?
No. Anthropic does not publish a perpetually free API tier the way Google does for Gemini. New Console accounts may receive promotional starting credit, but there is no standing free allowance you can build a product on. If a zero-cost tier is a hard requirement for your prototype, the Gemini API is the one to look at.
Do server-side tools like web search cost extra on the Claude API?
Yes, and they are billed outside the token rates. Anthropic's server-side web search is charged at $10 per 1,000 searches on top of the tokens the results consume once they enter the context window. So a search-augmented agent pays twice: once for the search call, and again for the substantially larger prompt the retrieved content produces. If you are budgeting an agent, count both.
Can I stack the Claude batch discount with prompt caching?
Yes. The Message Batches API takes 50% off both input and output for work returned within 24 hours, and it applies on top of cached-input rates. A batched job over a cached prefix lands at roughly 5% of the naive sticker price on the input line. The constraint is only that nobody can be waiting on the answer — which describes evals, backfills, bulk document processing and nightly pipelines, but not a chat product.
Per-token pricing for the other providers
Every page in this set reads from the same verified price table, so the numbers never disagree with each other.
OpenAI API pricing
GPT-5.6, GPT-5, mini and nano rates, plus the cost calculator.
Gemini API pricing
Google's free tier, the 200k context price cliff and the Batch/Priority service tiers.
Grok API pricing
xAI's unusually flat input/output gap and what it means for output-heavy apps.
DeepSeek API pricing
The cheapest rates on the board and a 50x cache-hit discount.
LLM API pricing comparison
All five providers head to head, with a recommendation per workload shape.
More on Claude
Claude Pro & Max pricing
The consumer subscriptions — separate product, separate bill from the API.
Claude review
Capabilities, context window and who Claude is actually best for.
ChatGPT vs Claude
Head-to-head on the assistants rather than the APIs.
Best AI coding tools
The agents and editors built on top of the Claude API.
ChatGPT already recommends your Claude-powered app. Does it recommend yours?
If you're building in LLM APIs, run a free AI-visibility scan on your own product — we ask ChatGPT across 5 prompt angles and score how often you get named. ~30 seconds, no signup, no card.
📬 Get the best new AI tools delivered weekly
One concise email with fresh launches, trending picks, and featured standouts.