Anthropic API: Pricing, Limits and Access From Russia in 2026

In short

The Anthropic API is paid programmatic access to the Claude models on a pay-per-token basis: no monthly subscription, you pay exactly for the volume of text processed. As of July 2026, prices run from $1 per million input tokens on Claude Haiku 4.5 to $10 on the flagship Fable 5, and output costs five times more than input. Rate limits depend on your account tier — Start, Build or Scale — and rise automatically as your spend grows. Anthropic does not officially serve Russia and does not accept Russian cards, so keys and payment go through aggregators like OpenRouter or through non-Russian cards — legal, but with caveats about fees and intermediary risk. Getting a key step by step is covered in the Claude API from Russia guide; this article is about the numbers: prices and limits.

What the Anthropic API costs per token

API billing is a meter, not a season ticket. A token is the smallest unit of text for a model: roughly 3–4 characters of Russian text, or three quarters of an English word. Input (your prompt, system instructions, context, conversation history) and output (the generated answer) are billed separately. The key rule: output costs 5× input across all the working models.

Current prices as of July 2026, per 1M tokens in US dollars:

*The introductory Sonnet 5 price runs until 31 August 2026 (data as of July 2026).

The cost of a request is simple arithmetic: (input tokens × input price) + (output tokens × output price). A million input tokens on Opus 4.8 is $5; a million output tokens is $25. Pick the model to match the job: Haiku for routine, Sonnet covers about 80% of the work, Opus for cases where long reasoning chains and agentic reliability matter.

What eats tokens in vibe coding

Beginners underestimate input. In agentic coding the model doesn't just read your question — it loads repository files, command output, tool definitions and the entire iteration history. A single active agent turn easily pulls 20,000–50,000 input tokens, and a session holds dozens of such turns.

Three things drive consumption fastest: long context (large files in the prompt), subagents (each runs its own conversation and multiplies tokens), and connected MCP servers (their tool descriptions sit in the input of every request). That's why your API bill correlates not with how many messages you send, but with the volume of context churning in the background.

Rate limits: where the ceilings are

A rate limit is a speed ceiling: how many requests and tokens per minute your organisation is allowed. Anthropic measures three parameters per model: RPM (requests per minute), ITPM (input tokens per minute) and OTPM (output tokens per minute). Going over returns a 429 error with a retry-after header telling you how long to wait.

Your tier is assigned automatically based on spending history and rises over time. Each tier has two ceilings — a monthly spend limit and rate limits. Data as of July 2026 (values for the Opus class):

The exact limits on your account are shown in the Anthropic console; the table gives order-of-magnitude figures for July 2026.

One important detail: on most Claude models, cached input tokens don't count toward ITPM. So prompt caching doesn't just make requests cheaper — it raises your real throughput. At an 80% cache hit rate, a 2M ITPM limit lets through up to 10M tokens per minute.

How to get a key and pay from Russia

Anthropic does not officially operate in Russia: signing up with a Russian address and paying with a Russian card both fail. There are two legal routes.

Aggregators (routers). OpenRouter is a gateway that resells access to models from Anthropic, OpenAI, Google and others through a single balance at the same per-token prices plus a small markup. You work with one key, and under the hood it routes the request to the right provider. The balance is held in dollars; from Russia it's topped up through intermediary services — as of July 2026 the fee is around 5–5.5% by card and about 5% via cryptocurrency.

A non-Russian card. If you hold a card issued outside Russia, you can pay Anthropic directly — the cardholder is responsible for complying with their own bank's terms.

A caveat about risk: intermediaries charge a fee, give no guarantee the payment arrives, and restrict refunds (usually you can only recover an unused balance within the first 24 hours). Choose established services, don't keep large sums on the balance, and check current terms before paying. Technical ways of getting around blocks are out of scope here — this is about payment routes only. Detailed instructions for setting up a key are in a separate piece, Claude API from Russia.

A separate question is which is cheaper: pay-per-token API or a subscription. For intensive vibe coding, a Claude subscription paired with Claude Code is often more predictable budget-wise than the API meter. The comparison is laid out in the Claude Code plans and payment guide. The Quest engine by qvib works on both subscription Claude Code and the API — so the payment choice is yours.

How to pay less

Two built-in mechanisms cut the bill by a multiple.

Prompt caching is the reuse of the unchanging part of a prompt: system instructions, a large document or tool definitions get cached, and subsequent requests read them at 10% of the input price (up to 90% savings). Writing to the cache costs 1.25× the input price with 5-minute retention and pays for itself from the second request with the same prefix.

The Batch API is asynchronous processing: you submit a batch of requests, the responses arrive within 24 hours, and the whole volume is billed at a 50% discount. Good for bulk jobs that don't need an instant answer — content generation, labelling, embeddings.

The third lever is choosing the right model: don't run Opus where Sonnet or Haiku will do. A well-framed task saves money too — a short, precise prompt means fewer wasted iterations.

What an hour of vibe coding really costs

Let's estimate the order of magnitude at Opus 4.8 prices. An active agentic session can churn through roughly 3M input and 300K output tokens in an hour. Without caching that's 3 × $5 + 0.3 × $25 = $22.50. With the repository context cached, the input side effectively drops several-fold, and the same hour costs a few dollars. The real figure depends on context size, cache hit rate and model — but "a few dollars an hour on Sonnet, up to a couple of dozen on Opus without optimisation" is a workable benchmark.

Want to go deeper on models, prompts and tools? Take a look at the qvib knowledge base. If a paid key is out of budget for now, you can start with free alternatives and aggregators and move to the API later.

FAQ

Do I need the API if I use Claude Code?

Not necessarily. Claude Code works both on a Claude subscription (Pro/Max) and on an API key. A subscription is a fixed monthly amount with usage limits; the API is strictly pay-per-token. For steady daily load a subscription is more predictable; the API suits irregular spikes and automation better. See the plans guide for the comparison.

Which model should I start with?

Sonnet 5. As of July 2026 it has the best price-to-quality balance ($3/$15, introductory $2/$10 until the end of August), and it covers most vibe coding tasks. Use Haiku for simple routine work and bring in Opus selectively for complex agentic chains.

Why did my bill grow when I barely wrote any prompts?

Because you mostly pay for input, and input means the entire context: repository files, conversation history, MCP tool descriptions and subagent work. Spend is driven by the volume of background context, not by the number of your messages. Turn on prompt caching and prune the history.

Is it legal to pay through intermediaries?

Paying for API models through legitimate aggregators and balance top-up services doesn't break the law — it's an ordinary payment for a foreign digital service. The risk here is commercial, not legal: fees, no guarantees and limited refunds. Choose proven services and don't keep a large balance.