Claude API Pricing 2026: Cost per Million Tokens by Model

The Claude API costs $2 per million input tokens and $10 per million output tokens on Claude Sonnet 5.5 in 2026, $4 and $20 on Opus 5.5, $10 and $50 on Fable 5.1, and $0.10 and $0.50 on Haiku 5.5 for prompts up to 100,000 tokens. Cache reads cost 5% of the input price on Sonnet and Opus 5.5, and the Batch API halves everything. Prices read on Anthropic's pricing page, October 2026.

Those four numbers per model are the whole bill for most products, and the two modifiers, cache reads and batch, are where the real savings sit. Our view: the per-token rate is the headline, but a chatbot with a long system prompt pays mostly cache-read rates, and an overnight job should never pay standard rates at all. This page prints every rate, then runs three real workloads through them.

Every price here was read on Anthropic's pricing page and its prompt caching documentation on 26 October 2026, in US dollars per million tokens (MTok). The consumer plans (Pro, Max, Team) are a different bill, covered on Claude pricing; the cross-vendor sheet of 20 models is LLM API prices 2026.

Toolradar data: the AI assistants ranking on Toolradar, the software directory we run, evaluated 1,574 tools as of October 2026; the product records are Anthropic API, OpenAI API and Google Gemini. AIpresso data: this newsletter reaches 29,000+ readers as of November 8, 2026, with an open rate of 31%.

Claude API prices per million tokens, October 2026

All figures are US dollars per million tokens.

Model	Input	Output	Cache write (5 min)	Cache write (1 hour)	Cache read	Batch input	Batch output
Claude Fable 5.1	10	50	12.50	20	0.25	5	25
Claude Opus 5.5	4	20	5	8	0.20	2	10
Claude Sonnet 5.5	2	10	2.50	4	0.10	1	5
Claude Haiku 5.5, prompts up to 100,000 tokens	0.10	0.50	0.125	0.20	0.01	0.05	0.25
Claude Haiku 5.5, prompts over 100,000 tokens	0.50	2.50	0.625	1	0.05	0.25	1.25
Claude Opus 5, 4.8, 4.7, 4.6, 4.5	5	25	6.25	10	0.50	2.50	12.50
Claude Sonnet 4.6, 4.5, 4	3	15	3.75	6	0.30	1.50	7.50
Claude Haiku 4.5	1	5	1.25	2	0.10	0.50	2.50
Claude Opus 4.1, 4	15	75	18.75	30	1.50	7.50	37.50

Fable 5, Mythos 5 and Mythos 5.1 share Fable 5.1's base and batch prices; their cache reads differ (Fable 5 and Mythos 5 at $1, Mythos 5.1 at $0.25). Sonnet 5's base price is blank on the pricing page while its caching page prints the same two figures as Sonnet 5.5, so Anthropic's own pages differ on that model; the figures above come from the pricing page.

How is the Claude API billed?

Four meters, all per token, nothing per request or per seat. Input tokens are everything you send: system prompt, conversation history, tool definitions and tool results. Output tokens are everything the model writes, including its thinking. A cache write is charged the first time you store a prompt prefix, at 1.25 times the input price for a 5-minute cache or 2 times for a 1-hour cache. A cache read is charged each time a later request reuses that prefix, at 10% of the input price on most models, 5% on Opus 5.5 and Sonnet 5.5, and 2.5% on Fable 5.1 and Mythos 5.1.

Two things that are not metered matter as much. The full 1 million token context window is billed at standard rates on Claude 4.6 and later (except Haiku 5.5, which charges five times more on any request whose prompt exceeds 100,000 tokens, cache reads included), so a 900,000-token request pays the same per-token rate as a 9,000-token one. And tool definitions are plain input: using any tool adds a system prompt of 286 tokens on the 5.5 models (406 with forced tool choice), the bash tool adds 325 tokens, the text editor 974 tokens, and the computer-use toolset about 4,500 tokens per request.

One number to carry into any cross-vendor comparison: Anthropic says Claude 4.7 and later models use a tokenizer that produces roughly 30% more tokens for the same text than Sonnet 4.6 and earlier. A million tokens on Opus 5.5 is less text than a million tokens on Sonnet 4.6, and token counts across vendors are not comparable at all; the dollar figures below assume equal counts.

What the modifiers do to the bill

Modifier	Effect	Applies to
Prompt caching, 5-minute	Write at 1.25x input, then reads at 0.1x (0.05x on Opus and Sonnet 5.5, 0.025x on Fable 5.1)	All models
Prompt caching, 1-hour	Write at 2x input, same read rates	All models
Batch API	50% off input and output, results within 24 hours	All models; not with fast mode
US-only inference (inference_geo: "us")	1.1x on every token category	Claude 4.6 and later
Fast mode	$8 and $40 on Opus 5.5; $10 and $50 on Opus 5 and 4.8	Opus 5.5, Opus 5, Opus 4.8
Web search tool	10 per 1,000 searches plus the tokens the results add	All models
Web fetch tool	No charge beyond the fetched tokens	All models
Code execution	Free with web search or web fetch; otherwise 1,550 free hours a month, then $0.05 per container-hour	All models
Managed Agents session runtime	$0.08 per hour while the session is running, plus tokens	Managed Agents

The multipliers stack. A cached prefix read in batch on Sonnet 5.5 costs $0.05 per million tokens, our arithmetic on 0.05 times the $1 batch input rate; with US-only inference on top it is $0.055. Caching pays for itself after one read on the 5-minute cache (1.25x write against 0.1x reads) and after two reads on the 1-hour cache, so a prompt that is reused more than once a minute should always be cached, and a system prompt reused once an hour should use the 1-hour cache only if it is reused at least twice in that hour.

AIpresso readers get Anthropic's rate changes in the daily brief before the invoice reflects them; the signup is at the end of this page.

What does a chatbot cost on Claude?

Our first workload is a support assistant with a 1,500-token cached system prompt and about 500 fresh tokens per message, handling 100,000 conversations a month: 200 million input tokens, of which 150 million are cache reads, and 30 million output tokens. All figures are US dollars a month, our arithmetic on the printed rates, ignoring the cache-write premium on the first store.

Model	Fresh input (50M)	Cache reads (150M)	Output (30M)	Monthly total	Without caching
Claude Haiku 5.5	5	1.50	15	21.50	35
Claude Sonnet 5.5	100	15	300	415	700
Claude Opus 5.5	200	30	600	830	1,400
gpt-6-luna (OpenAI)	5	1.50	15	21.50	35
gpt-6.1-sol (OpenAI)	100	15	300	415	700
Gemini 3.8 Flash (Google, through 31 December 2026)	37.50	11.25	112.50	161.25	262.50
Gemini 4 Argon (Google, through 30 November 2026)	100	15	300	415	700

Three things stand out. Output is the biggest line on every row, 72% of the Sonnet 5.5 bill, so a shorter answer style saves more than any caching trick. Caching cuts the Sonnet bill by 41%, from $700 to $415, because three quarters of the input is the repeated system prompt. And the mid tier is a three-way tie: Sonnet 5.5, gpt-6.1-sol and Gemini 4 Argon land on the same $415 at the same token counts, which makes quality on your transcripts, not price, the deciding factor; Google's page prints Argon at $4 and $20 from 1 December 2026, which doubles its row.

What does a batch summarizer cost?

The second workload is overnight: 10,000 documents of 20,000 tokens each, summarized to 500 tokens, through the Batch API. That is 200 million input tokens and 5 million output tokens a month at half price. Figures are US dollars a month.

Model	Batch input (200M)	Batch output (5M)	Monthly total	Same job at standard rates
Claude Haiku 5.5	10	1.25	11.25	22.50
Claude Sonnet 5.5	200	25	225	450
Claude Opus 5.5	400	50	450	900
gpt-6-luna (OpenAI batch)	10	1.25	11.25	22.50
gpt-6.1-sol (OpenAI batch)	200	25	225	450
Gemini 3.8 Flash (Google batch)	75	9.38	84.38	168.75
Gemini 4 Argon (Google batch)	200	25	225	450

Input dominates here, 89% of the Sonnet row, because each document is 40 times longer than its summary. The decision is the model, not the modifier: Haiku 5.5 does the job for 5% of Sonnet's price, and the only question is whether its summaries are good enough on your documents. Batch results arrive within 24 hours on Anthropic's documentation, so this shape fits anything that is not a user waiting on a screen. Haiku's 100,000-token threshold matters if the documents are long: a 120,000-token document pays Haiku's higher tier, five times the small rate, on the whole request.

What does an agent loop cost?

The third workload is a coding agent: 1,000 runs a month, each about 30 turns, with the growing context re-sent every turn. Per run we assume 1.5 million input tokens, 90% of them cache reads, and 40,000 output tokens; per month that is 150 million fresh input tokens, 1.35 billion cache-read tokens and 40 million output tokens. Figures are US dollars.

Model	Fresh input (150M)	Cache reads (1.35B)	Output (40M)	Monthly total	Per run
Claude Sonnet 5.5	300	135	400	835	0.84
Claude Opus 5.5	600	270	800	1,670	1.67
Claude Fable 5.1	1,500	337.50	2,000	3,837.50	3.84
gpt-6.1-sol (OpenAI)	300	135	400	835	0.84
gpt-6-astra (OpenAI)	1,500	1,350	2,000	4,850	4.85
Gemini 4 Argon (Google, through 30 November 2026)	300	135	400	835	0.84

This is the shape where cache-read rates decide the bill. Fable 5.1 and gpt-6-astra have identical base prices, but Fable's 2.5% cache read ($0.25) against OpenAI's 10% cached input ($1) is a $1,012.50 gap a month on 1.35 billion re-read tokens, our arithmetic. Without caching at all, the Sonnet 5.5 run would cost $3,400 a month instead of $835, because every one of those 1.5 billion input tokens would be charged at $2. An agent that does not cache its context is paying four times the necessary price.

Opus 5.5 at twice Sonnet's price is the trade most teams will weigh: $0.84 against $1.67 per run. If Opus finishes in fewer turns or fewer retries, it can be the cheaper model per completed task, which no per-token table shows. Which model to point at which job is the Claude review; the head-to-heads are Claude vs ChatGPT and Claude vs Gemini.

Claude vs OpenAI vs Gemini at list price

Input and output prices in US dollars per million tokens.

Tier	Claude	OpenAI	Gemini
Frontier	Fable 5.1: 10 / 50, cache read 0.25	gpt-6-astra: 10 / 50, cached input 1	No equivalent printed; Gemini 4 Argon is Google's flagship at 2 / 10
Flagship	Opus 5.5: 4 / 20, cache read 0.20	gpt-5.6-sol: 4 / 20, cached input 0.40	Gemini 4 Argon: 2 / 10, 4 / 20 from 1 December 2026
Mid	Sonnet 5.5: 2 / 10, cache read 0.10	gpt-6.1-sol: 2 / 10, cached input 0.10	Gemini 4 Argon: 2 / 10, implicit caching at 0.10
Small	Haiku 5.5: 0.10 / 0.50, cache read 0.01	gpt-6-luna: 0.10 / 0.50, cached input 0.01	Gemini 3.8 Flash: 0.75 / 3.75, 1.50 / 7.50 from 1 January 2027
Batch	50% off, all models	50% off (Flex/Batch tier)	50% off
Web search	10 per 1,000 searches	10 per 1,000 calls	5,000 free a month, then 14 per 1,000

Read on OpenAI's pricing page and Google's Gemini API pricing page on 26 October 2026. The pattern is that list prices match tier for tier between Anthropic and OpenAI, down to the cent on the small and mid models, and the differences are in the modifiers: Anthropic's cache reads are cheaper at the top (2.5% against 10%), OpenAI prints a 12-hour cache write and faster tiers at 2x and 6x, and Google's prices carry printed end dates after which they double. Gemini's consumer plans are on Gemini pricing and OpenAI's on ChatGPT pricing.

Where the Claude bill surprises people

• Thinking tokens are output tokens. On models with adaptive thinking, the reasoning is billed at the output rate whether or not you display it.
• Cache reads still cost something, and the write costs more. The first store of a 100,000-token prefix on Sonnet 5.5 is $0.25 (1.25 times $0.20), our arithmetic; every read after it is $0.01.
• Haiku 5.5 has a cliff at 100,000 tokens. One long prompt pays $0.50 and $2.50 instead of $0.10 and $0.50, on the whole request.
• US-only inference is a 10% surcharge on every category, including cache reads.
• Fast mode doubles Opus 5.5 to $8 and $40 and cannot be combined with batch.
• Tool definitions are re-sent every turn. A computer-use toolset at about 4,500 tokens per request is 135,000 input tokens across a 30-turn run before the model does anything.
• Older Opus models cost more than the new one. Opus 5, 4.8, 4.7, 4.6 and 4.5 are $5 and $25 against Opus 5.5's $4 and $20, and Opus 4.1 and 4 are $15 and $75.

Which Claude model should you pay for?

Haiku 5.5 for anything high-volume with short prompts: classification, extraction, routing, summaries of documents under 100,000 tokens. At $21.50 a month for the chatbot workload and $11.25 for the batch job, it is a hundredth of the frontier price and ties the cheapest OpenAI model.

Sonnet 5.5 as the default for products: the mid tier with the 5% cache read, which puts the chatbot at $415 and the agent at $0.84 a run. Opus 5.5 when the task is long-running agentic work and a higher completion rate per run would pay back the 2x. Fable 5.1 only for the hardest reasoning, priced at five times Sonnet; its 2.5% cache read is the one place it is cheaper to operate than its OpenAI equivalent. The consumer plans are the wrong bill for any of this; is Claude Max worth it covers when a $100 or $200 subscription beats the API for one person.

Methodology and sources

Per-model prices, cache multipliers, batch rates, the long-context and Haiku 5.5 threshold rules, tool-use token overheads, the data-residency multiplier, fast mode and the Managed Agents runtime rate were read on Anthropic's pricing page on 26 October 2026, with the per-model cache-read prices cross-checked on the prompt caching page the same day. OpenAI's rates come from its API pricing page and Google's from the Gemini API pricing page, both read on 26 October 2026, at their standard tiers.

The three workloads are ours: the token counts are assumptions stated in each section, the dollar figures are those counts multiplied by the printed rates, cache-write premiums on the first store are ignored, and token counts are treated as equal across vendors even though tokenizers differ. The Toolradar count is from its AI assistants category as of October 2026.

Louis Corneloup, founder of Toolradar and Dupple, wrote this page; we run no test lab and give no rating. The method is on how we rate, corrections go through the editorial policy, and the author note is on about. Re-read the pricing page before you commit a budget: Anthropic prices change per model release, and Google's current rates carry printed end dates.

FAQ

How much does the Claude API cost per million tokens in 2026?

Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, Sonnet 5.5 $2 and $10, Opus 5.5 $4 and $20, and Fable 5.1 $10 and $50. Cache reads cost 2.5% to 10% of the input price depending on the model, and the Batch API halves both input and output rates.

Is there a free tier for the Claude API?

Anthropic's pricing page prints no free tier for the API; every model is billed per token from the first request. Code execution includes 1,550 free container-hours a month, and web fetch adds no charge beyond the tokens it brings in. The free plan on claude.ai is a consumer product, not API access.

How does Claude API pricing compare with OpenAI?

Tier for tier the list prices match: $0.10 and $0.50 on Haiku 5.5 and gpt-6-luna, the same mid-tier rate on Sonnet 5.5 and gpt-6.1-sol, $10 and $50 on Fable 5.1 and gpt-6-astra. The difference is cached input: Anthropic charges $0.25 per million cache-read tokens on Fable 5.1 against OpenAI's $1 on gpt-6-astra, which on our agent workload is about $1,012 a month in Anthropic's favour.

How much does prompt caching save on Claude?

On the chatbot workload, caching takes Sonnet 5.5 from $700 to $415 a month, a 41% cut, and on the agent workload from $3,400 to $835. A cache read costs 10% of the input price on most models, 5% on Opus 5.5 and Sonnet 5.5, and 2.5% on Fable 5.1; the first write costs 1.25 times the input price for a 5-minute cache or 2 times for a 1-hour cache.

Does the Batch API work with caching?

Yes. Anthropic states the caching multipliers stack with the Batch discount and data residency, so a cached prefix read in a batch request on Sonnet 5.5 is $0.05 per million tokens, our arithmetic on 0.05 times the $1 batch input price. Batch is not available with fast mode.

Does the Claude API charge more for long context?

Not on Claude 4.6 and later models other than Haiku 5.5: the 1 million token window is billed at standard rates. Haiku 5.5 is the exception, charging $0.50 and $2.50 per million tokens on any request whose prompt, cache reads included, exceeds 100,000 tokens.

Cite this: AIpresso, "Claude API Pricing 2026: Cost per Million Tokens by Model", October 2026.

AIpresso is the short daily brief on which AI prices moved and by how much. Subscribe at aipresso.com and skip re-reading three pricing pages every month.

Some offers on this page may be paid placements or contain affiliate links.