The short version
omitted: you pay for reasoning that is never returned.Claude's pricing page looks simple. Two numbers per model, input and output, dollars per million tokens. Then the first real invoice arrives, it does not match the arithmetic you did, and you lose an afternoon working out why.
This page explains the gap. Every figure below comes from Anthropic's own documentation, retrieved in July 2026, with the traps marked where they exist.
Claude API pricing is per million tokens, billed separately for input and output. Cache reads cost a tenth of base input; cache writes cost more than base input. Batch halves everything. Here is the full current table.
| Model | Input | Cache write (5 min) | Cache write (1 hr) | Cache read | Output | Batch in / out |
|---|---|---|---|---|---|---|
| Claude Fable 5 | $10 | $12.50 | $20 | $1 | $50 | $5 / $25 |
| Claude Mythos 5 (limited) | $10 | $12.50 | $20 | $1 | $50 | $5 / $25 |
| Claude Opus 4.8 | $5 | $6.25 | $10 | $0.50 | $25 | $2.50 / $12.50 |
| Claude Opus 4.7 / 4.6 / 4.5 | $5 | $6.25 | $10 | $0.50 | $25 | $2.50 / $12.50 |
| Claude Sonnet 5 (intro, to 31 Aug 2026) | $2 | $2.50 | $4 | $0.20 | $10 | $1 / $5 |
| Claude Sonnet 5 (from 1 Sep 2026) | $3 | $3.75 | $6 | $0.30 | $15 | $1.50 / $7.50 |
| Claude Sonnet 4.6 / 4.5 | $3 | $3.75 | $6 | $0.30 | $15 | $1.50 / $7.50 |
| Claude Haiku 4.5 | $1 | $1.25 | $2 | $0.10 | $5 | $0.50 / $2.50 |
| Claude Opus 4.1 (deprecated, retires 5 Aug 2026) | $15 | $18.75 | $30 | $1.50 | $75 | $7.50 / $37.50 |
Opus 4.8 costs exactly 5× Haiku 4.5 on both input and output. Fable 5 costs 2× Opus 4.8 and 10× Haiku. Those multiples are the whole argument for routing: if a task can be done by Haiku, running it on Opus burns five dollars for every one it needed.
Anthropic's newer models tokenize text more finely, so the same document costs more to send. Anthropic states it plainly: Opus 4.7 and later Opus models, Fable 5, Mythos 5 and Sonnet 5 use a newer tokenizer that "produces approximately 30% more tokens for the same text". Sonnet 4.6 and earlier use the old one.
Thirty percent is the headline. The distribution is what will actually hit you.
A developer ran identical files through Anthropic's own count_tokens endpoint on Opus 4.6 and Opus 4.7. The ratio was not flat.
| Content type | Old tokenizer | New tokenizer | Ratio |
|---|---|---|---|
| Technical documentation (English) | 478 | 704 | 1.47× |
| A real CLAUDE.md file (5 KB) | 1,399 | 2,021 | 1.45× |
| Shell script | 1,033 | 1,436 | 1.39× |
| TypeScript source | 1,208 | 1,640 | 1.36× |
| Python source | 864 | 1,112 | 1.29× |
| English prose | 508 | 611 | 1.20× |
| Dense JSON | 13,939 | 15,706 | 1.13× |
| Numeric CSV | 5,044 | 5,414 | 1.07× |
| Japanese / Chinese prose | 856 / 779 | 866 / 789 | 1.01× |
Read the top two rows again. Technical English — documentation, instruction files, system prompts — inflates by 45 to 47 percent, not thirty. That is precisely the content sitting in every agent's stable prefix. The 30% figure is an average across content types you probably do not send much of.
Sonnet 5 launched at $2 / $10 against Sonnet 4.6's $3 / $15. That reads as a 33% cut. But Sonnet 5 runs the new tokenizer and Sonnet 4.6 does not, so the same text becomes roughly 30% more tokens. The intro pricing roughly cancels the tokenizer, leaving you close to flat.
Then on 1 September 2026 the intro pricing ends and Sonnet 5 moves to $3 / $15 — identical to Sonnet 4.6 per token, on 30% more tokens. On that date, for the same workload, Sonnet 5 becomes more expensive than the model it replaced. Put it in your calendar.
Prompt caching stores a stable prefix — system prompt, tool definitions, retrieved documents — so you are not charged full price to send it again. A cache read costs 0.1× base input, a 90% discount. Writing the cache costs 1.25× input on a five-minute TTL, or 2× on a one-hour TTL.
Anthropic's own break-even, verbatim: caching "pays off after just one cache read for the 5-minute duration (1.25x write), or after two cache reads for the 1-hour duration (2x write)". Reuse a prefix even twice and you are ahead.
Cached reads do not count toward your input-tokens-per-minute limit on current models. Anthropic's own example: with a 2,000,000 ITPM limit and an 80% cache hit rate you can effectively process 10,000,000 input tokens per minute. Caching is a rate-limit lever as much as a cost lever.
Caching fails quietly. No error, no warning — just a bill that looks like caching was never switched on. These are the failure modes, straight from Anthropic's documentation.
A prefix shorter than the model's minimum is not cached, and no error is returned. The thresholds differ per model, which is the trap: 512 tokens for Fable 5 and Mythos 5; 1,024 for Opus 4.8, Sonnet 5, Sonnet 4.6 and Sonnet 4.5; 2,048 for Opus 4.7; and 4,096 for Opus 4.6, Opus 4.5 and Haiku 4.5. A 2,000-token system prompt caches on Opus 4.8 and does absolutely nothing on Haiku 4.5.
Anthropic names this as a common mistake: put a cache breakpoint on a block containing a timestamp or the user's latest message and "you pay for a fresh cache write on every request and never get a read". You are then paying 1.25× input, forever, for nothing.
Cache reads only find entries that earlier requests actually wrote. In a growing conversation, once the breakpoint drifts more than twenty blocks past the last write, you get no hit and pay full input on the entire prefix.
tool_choice, adding or removing images, changing thinking settings — invalidates messages.Since 5 February 2026, caches on the Claude API are isolated per workspace rather than per organisation. Teams running several workspaces silently lost hits they used to get. Bedrock and Google Cloud still keep org-level isolation.
The Batch API discounts input and output by 50%, and Anthropic confirms the discount stacks with prompt caching and data residency. If a workload tolerates delay — evaluations, backfills, nightly classification, document enrichment — running it synchronously doubles the bill for no benefit.
Two limits worth knowing: batch is unavailable with fast mode, and unavailable inside Managed Agents sessions.
Extended thinking is billed as output, at the full output rate. Anthropic is explicit: "You're charged for the full thinking tokens generated by the original request, not the summary tokens. The billed output token count will not match the count of tokens you see in the response."
It gets sharper. On Opus 4.8, thinking display defaults to omitted — the field comes back empty and you are billed for every token of reasoning that produced it. Anthropic's note on that setting: "You're still charged for the full thinking tokens. Omitting reduces latency, not cost."
The billed count is readable at usage.output_tokens_details.thinking_tokens. If you have never looked at that field, look today. It is frequently the largest line in an agent's bill.
| Modifier | Effect | Applies to |
|---|---|---|
| Cache write, 5-minute TTL | 1.25× input | All active models |
| Cache write, 1-hour TTL | 2× input | All active models |
| Cache read | 0.1× input (90% off) | All active models |
| Batch API | −50% input and output | All models; not with fast mode or Managed Agents |
| Long context (1M window) | No surcharge. A 900k-token request bills at the same per-token rate as a 9k one | Fable 5, Mythos 5, Opus 4.8/4.7/4.6, Sonnet 5, Sonnet 4.6 |
Data residency (inference_geo: "us") | 1.1× on all token categories | Opus 4.6, Sonnet 4.6 and later only |
| Regional endpoints on Bedrock / Vertex | +10% over global endpoints | Sonnet 4.5, Haiku 4.5, Opus 4.5 and later |
| Fast mode (Opus 4.8) | $10 / $50 — 2× standard, for up to 2.5× output speed | Research preview; not combinable with batch |
| Priority Tier | No longer purchasable. Existing commitments run to contract end | — |
| Tokenizer | ~30% more tokens for the same text | Opus 4.7+, Fable 5, Mythos 5, Sonnet 5 |
This is the one place Claude is unambiguously cheaper than its rivals. OpenAI doubles input price above 272k tokens. Google raises Gemini 3.1 Pro from $2 / $12 to $4 / $18 above a 200k prompt. Anthropic charges nothing extra for the full 1M window, and caching and batch discounts still apply across it. On long-context work, the "expensive" model is often the cheap one.
| Feature | Price | Watch out for |
|---|---|---|
| Web search | $10 per 1,000 searches | Search results bill as input tokens in that turn and in every subsequent turn |
| Web fetch | No surcharge | A 500 KB PDF is roughly 125,000 input tokens |
| Code execution | 1,550 free hours/month per org, then $0.05 per container-hour | Minimum 5-minute billing. Attaching files bills container time even if the tool is never called |
| Managed Agents | Tokens + $0.08 per session-hour | Batch, fast mode and residency modifiers do not apply here |
| Computer use | No fee; adds ~466–499 tokens to the system prompt | Screenshots bill as image tokens |
| Any tool present | Adds 290–804 tokens of system prompt, depending on model | Opus 4.7 is heaviest at 675–804 tokens |
Anthropic replaced the old numbered Tier 1–4 system with Start, Build, Scale and Custom. Each carries a monthly spend cap that acts as a hard stop: hit it and API usage pauses until the next calendar month unless you request a raise.
| Tier | Monthly spend cap | Opus 4.x RPM / ITPM / OTPM | Fable 5 RPM / ITPM / OTPM |
|---|---|---|---|
| Start | $500 | 1,000 / 2M / 400k | 1,000 / 500k / 100k |
| Build | $1,000 | 5,000 / 5M / 1M | 2,000 / 1.5M / 300k |
| Scale | $200,000 | 10,000 / 10M / 2M | 4,000 / 4M / 800k |
| Custom | No cap | Negotiated with the account team | |
Two details catch people out. The Opus limit is a combined bucket across Opus 4.8, 4.7, 4.6 and 4.5 — switching versions does not hand you a fresh allowance. And Sonnet 5 has its own bucket, separate from Sonnet 4.x, so migrating genuinely doubles your Sonnet-class headroom.
Anthropic no longer publishes a deposit ladder. The current documentation says only that organisations move up "over time as you use the API", and that you can request an increase from the Console limits page. Any article quoting a specific dollar threshold for Start → Build is repeating pre-June-2026 information.
Not a real one. Anthropic's own wording: "New users receive a small amount of free credits to test the API." No dollar figure is published. The widely-repeated $5 figure appears only in third-party blogs, so treat it as unconfirmed.
Of the three major vendors only Google offers a genuine ongoing free API tier, and only on Flash-class models — Gemini 3.1 Pro is paid-only. If a free tier is what you need for prototyping, that is where it lives.
Claude is sold through three cloud marketplaces, and the billing mechanics differ from the direct API in ways worth understanding before procurement routes through them.
Claude Platform on AWS and Claude in Microsoft Foundry bill in Claude Consumption Units at $0.01 per CCU — 100 CCU equals $1.00. Anthropic rates your usage in dollars at standard rates, applies any negotiated discount, converts to CCUs and reports hourly. Your cloud bill shows a single CCU line item.
The consequences are real: billing is arrears-only, there are no prepaid credits, discounts show up as fewer CCUs metered rather than a cheaper CCU, and fast mode is unavailable. Organisations on Claude Platform on AWS are also pinned to the Start tier and do not auto-advance.
On Amazon Bedrock and Google Vertex, regional and multi-region endpoints carry a 10% premium over global endpoints, for Sonnet 4.5, Haiku 4.5, Opus 4.5 and every model since. Bedrock also does not support automatic caching, and its minimum cacheable prefix for Fable 5 is 1,024 tokens rather than 512.
An AI gateway resells Claude through an OpenAI-compatible endpoint, so a single key and balance cover Claude, GPT and Gemini together. Prices vary more than people expect: in our own Fable 5 comparison the spread runs from 40% below Anthropic's list to 10% above it for the identical model.
That is the whole reason to check rather than assume. The side-by-side table sits on our Fable 5 API page and the full ranking on our comparison page; how gateways actually charge is covered in our guide to AI API gateways.
At list price Claude sits between the two. Opus 4.8 at $5 / $25 undercuts GPT-5.5 at $5 / $30 on output, and both are roughly double Gemini 3.1 Pro at $2 / $12.
| Model | Input | Cached input | Output | Long-context penalty |
|---|---|---|---|---|
| Claude Opus 4.8 | $5 | $0.50 | $25 | None — full 1M at standard rates |
| OpenAI GPT-5.5 | $5 | $0.50 | $30 | Input 2×, output 1.5× above 272k |
| Google Gemini 3.1 Pro | $2 | $0.20 | $12 | $4 / $18 above a 200k prompt |
| Claude Haiku 4.5 | $1 | $0.10 | $5 | None |
| Google Gemini 3.1 Flash-Lite | $0.25 | $0.025 | $1.50 | — |
But per-token parity is not per-task parity, and the tokenizer comes back here. Artificial Analysis measures the real cost of running the same evaluation suite on each model: Claude Opus 4.8 cost $4,011, GPT-5.5 cost $2,159 and Gemini 3.1 Pro cost $815. Opus and GPT-5.5 carry nearly the same sticker price and a 1.9× gap in what they actually cost to run. We pull that apart in our real-cost comparison.
thinking_tokens. You are billed for reasoning you cannot see. Measure it before optimising anything else.As of July 2026: Claude Fable 5 is $10 input / $50 output, Opus 4.8 is $5 / $25, Sonnet 5 is $2 / $10 on intro pricing (rising to $3 / $15 on 1 September 2026), and Haiku 4.5 is $1 / $5. Cache reads cost a tenth of the input rate on every model.
At list price Claude Opus 4.8 is slightly cheaper than GPT-5.5 on output ($25 vs $30) and identical on input. In practice Opus costs more to run because it emits more tokens: Artificial Analysis measured $4,011 to run its evaluation suite on Opus 4.8 against $2,159 on GPT-5.5.
No permanent one. Anthropic says new users get "a small amount of free credits" but publishes no figure. Of the major vendors only Google offers an ongoing free API tier, and only on Flash-class models.
Anthropic's newer models — Opus 4.7 and later, Fable 5, Mythos 5 and Sonnet 5 — use a tokenizer producing about 30% more tokens for the same text. Independent measurement puts it at 1.20× on English prose and up to 1.47× on technical documentation. Sonnet 4.6 and earlier are unaffected.
Cache reads cost 0.1× base input, a 90% discount. Writing the cache costs 1.25× on a five-minute TTL. Anthropic states caching pays off after a single read at that TTL, or two reads on the one-hour TTL.
The usual causes are a prefix shorter than the model's minimum (4,096 tokens on Haiku 4.5, 1,024 on Opus 4.8), a breakpoint on a block that changes every request, a changed tool definition, or the breakpoint drifting more than twenty blocks past the last write. None of these produce an error.
Yes, at the full output rate, including tokens you never receive. On Opus 4.8 the thinking display defaults to omitted — the field returns empty and you are still billed. Check usage.output_tokens_details.thinking_tokens.
No. Fable 5, Mythos 5, Opus 4.8/4.7/4.6, Sonnet 5 and Sonnet 4.6 include the full 1M-token window at standard pricing. A 900k-token request bills at the same per-token rate as a 9k one. OpenAI and Google both charge a premium above their thresholds.
Fifty percent off both input and output. It stacks with prompt caching and data residency, but is unavailable with fast mode or inside Managed Agents sessions.
Request an increase from the Console limits page, or contact sales. Anthropic no longer publishes spend thresholds for advancing between the Start, Build and Scale tiers.
API usage pauses until the next calendar month unless you request a raise. Caps are $500 on Start, $1,000 on Build and $200,000 on Scale. Custom has no cap.
Not inherently. Token rates track Anthropic's list price, and regional or multi-region endpoints carry a 10% premium over global ones. Bedrock also does not support automatic caching, and billing on Claude Platform on AWS is arrears-only in Claude Consumption Units at $0.01 each.
It depends entirely on the gateway. In our own Fable 5 comparison, prices run from 40% below Anthropic's list to 10% above it for the identical model. A gateway is not automatically cheaper — check the landed price with top-up fees included.
The usual suspects, in order: you switched to a model on the new tokenizer, your cache stopped being hit after a tool-definition or model change, extended thinking is consuming more output than you realise, or a long session was never cleared.
No. Anthropic states that Priority Tier capacity commitments are no longer available for purchase. Existing commitments run to the end of their contract.
Claude's list price is honest and easy to read. Its effective price is a function of three things the price table does not show: how your text tokenizes, whether your cache is actually being hit, and how many invisible thinking tokens the model burns before it answers.
Get those three right and Claude is competitive with anything on the market. On long-context work, where OpenAI doubles its input rate and Google raises its own, Claude's flat 1M-token pricing makes it the cheapest flagship available. Get them wrong and you will pay a premium for a model you could have run on Haiku.
If you do only one thing: turn on prompt caching, then go and read thinking_tokens across your last hundred requests. Almost everyone is surprised by what they find.
count_tokens endpoint, April 2026.Copyright © 2026 Best AI Gateways
Independent ranking. We may be rewarded for recommending the service we rate best and sending users to it — that reward pays for the research behind this comparison and never buys a ranking position, at no extra cost to you. Information is provided “as is” without warranty. See our Disclaimer & Terms.