Claude API Pricing (2026): Cost per 1M Tokens for Every Model

Claude API pricing runs from $1 per 1M input tokens on Claude Haiku 4.5 up to $10
on Claude Fable 5. Anthropic bills input and output separately, and output costs five
times input across the entire range — so the headline input price tells you very little about
what you will actually pay. The table below gives every current model, and the blended column
is the number worth comparing.

Which Claude model should you use?

Anthropic’s line is a straightforward capability ladder, and most teams overpay by starting
at the top of it.

Haiku 4.5 is the one to treat as the default rather than the fallback. At
$1 in / $5 out it absorbs an order of magnitude more traffic than an Opus-tier model for the
same money, and the vast majority of what a production application does — classification,
extraction, routing, short-form generation — never needed a frontier model. Its 200K context
is the one real limit; the rest of the line reaches 1M.

Sonnet 5 is the workhorse. Its benchmark profile is coding-heavy: 85.2% on
SWE-Bench Verified, 80.4% on Terminal-Bench 2.1, 81.2% on OSWorld-Verified. Those numbers sit
close to models costing several times more, which is the whole argument for putting Sonnet
rather than Opus at the centre of a coding agent.

Opus 5 is the flagship, ranked first on the Artificial Analysis Intelligence
Index at 61. It arrived at unchanged Opus pricing, so anyone already on Opus 4.8 can migrate
with a model-string change and no budget consequence — worth doing promptly, because a
generational capability jump at flat pricing is rare.

Fable 5 only makes sense at the top of the difficulty curve. At $50 per 1M
output tokens it costs roughly five times Opus on the output side, so routing ordinary requests
to it burns budget for no measurable gain. Reserve it for work where a wrong answer is
expensive: unattended multi-step agents, deep research, refactors across a large codebase.

The pattern that works is a router — Haiku in front, escalating to Sonnet or Opus only when
a confidence check fails. Teams that instrument their token mix usually discover output is
where the bill actually lives, and that capping max_tokens and moving first drafts
to a cheaper tier saves more than any prompt-compression trick.

Three things about Claude pricing that catch people out

Sonnet 5’s price is temporary. The $2 / $10 rate is introductory and runs
through 31 August 2026, after which standard pricing of $3 / $15 applies. That is a 50%
increase, and it will land silently on any workload sized against the current number. If you
are modelling unit economics for something launching this year, model it at $3 / $15 and treat
today’s rate as a discount.

Opus 5’s Fast mode doubles both rates. The API-only Fast tier bills at
$10 / $50 rather than $5 / $25. It earns that only where latency is the product — interactive
coding assistants, live agent loops where a user is waiting. For batch work, overnight agents
and anything queued, standard mode delivers the same intelligence at half the price.

There is no long-context premium. This is the quiet advantage over rivals.
Claude Opus 4.8 bills its 1M-token context at the same rate from the first token to the
millionth. Several competing models step the input rate up once a prompt passes a threshold —
GPT-5.6 Sol charges 2x input and 1.5x output above 272K tokens, and Gemini 3.1 Pro roughly
doubles input above 200K. That makes their long-context features unpredictable to budget,
because the same feature costs a different amount depending on how much a user pasted in. With
Claude you can size prompts around what produces the best answer rather than around a pricing
cliff.