Anthropic API Review 2026

What is Anthropic API?

A support bot that handles 10,000 chats a month, each with roughly 2,000 tokens going in and 500 coming out, costs about $4.50 on Claude Haiku 5.5 and $90 on Sonnet 5.5 at list rates. That 20x gap is the first lesson of anthropic claude api pricing per million tokens: the model you pick moves your bill more than any clever prompt trick. This review draws on Anthropic's pricing page, docs and privacy centre.

What 10,000 chats a month costs on each Claude model

Here is that same workload (20 million input tokens, 5 million output tokens) priced on the four current models. Rates are standard, uncached, in US dollars per million tokens.

Output is five times dearer than input on every model, so long answers hurt more than long prompts. Batch jobs halve both columns.

The Claude API in one paragraph

Anthropic PBC, founded in 2021 and based in San Francisco, sells its models through what its docs now call the Claude API. You sign in to the Claude Console at platform.claude.com, buy prepaid credits and send requests to the Messages API at api.anthropic.com. It is billed separately from Claude Pro or Max chat subscriptions. It suits SaaS teams, agencies building chatbots and engineers writing coding agents.

Reading anthropic claude api pricing per million tokens line by line

A few details on the pricing page change real bills more than the headline rates do:

• Thinking costs output rates. All four current models use adaptive thinking, steered with an effort setting. The older extended thinking mode with a fixed budget_tokens value is rejected with a 400 error on Claude 4.7 and later. Reasoning tokens count toward max_tokens and are billed as output.
• Haiku 5.5 has a 100K step. Once a prompt crosses 100,000 tokens, Haiku moves to $0.50 input and $2.50 output for that request. The other current models charge the same rate across their full 1M-token window.
• US-only processing costs 10% more. Setting inference_geo to "us" applies a 1.1x multiplier; the default global routing uses list prices.
• Opus 5.5 fast mode runs at $8 input and $40 output, and cannot be combined with batches.
• Server tools add fees. Web search is $10 per 1,000 searches plus tokens. Web fetch adds nothing beyond tokens. Code execution costs nothing extra when your request also carries the current web search or web fetch tool. Used alone, it is billed by container time with a 5-minute minimum; the docs give each organisation 1,550 free hours a month and the pricing page says 50 free hours a day, after which it is $0.05 per container hour.

Failed requests are not charged, but a request still bills if your client times out before it finishes.

Matching Opus, Sonnet, Haiku or Fable to the job

The model overview gives every current model a 1M-token context window, 128K maximum output, text and image input, vision and tool use. Anthropic tells developers who are unsure to start on Claude Opus 5.5, and to move up to Fable 5.1 only for hard reasoning or long agent runs that Opus still fails at higher effort settings. Claude Haiku 5.5 is aimed at classification, extraction and routing, where latency and volume matter more than depth. Sonnet 5.5 costs half as much as Opus and is a sensible default for chat products.

Where Claude API prompt caching savings come from

Prompt caching stores a long, repeated prefix (a system prompt, catalogue or code file) so later calls read it cheaply. Writing to the 5-minute cache costs 1.25x base input and the 1-hour cache costs 2x. Reads are where the money is: 0.05x on Opus 5.5 and Sonnet 5.5, 0.025x on Fable 5.1 and 0.1x on Haiku 5.5.

Take a 50,000-token system prompt sent 1,000 times inside the cache window on Sonnet 5.5. Uncached, that prefix costs $100. Cached, you pay $0.125 for one write and about $5 for 999 reads, close to a 95% cut on that portion. Cache reads also skip your input-tokens-per-minute limit, which raises throughput. The minimum cacheable block is 512 tokens on current models, and one top-level cache_control field turns on automatic caching for a growing conversation.

Batch processing for work that can wait

Batch processing suits nightly jobs such as tagging 50,000 product reviews. One batch holds up to 100,000 requests or 256 MB. Most finish within an hour; anything unfinished after 24 hours expires, and results stay downloadable for 29 days. Caching discounts stack on top of the 50% batch rate, though Anthropic says cache hits inside batches are best effort.

How to get a Claude API key and send the first request

• Create an account in the Claude Console at platform.claude.com.
• Claim the one-time $5 test credit from the dashboard by verifying a mobile number (VoIP numbers and landlines are refused), or buy credits under Settings, then Billing. Calls stop when the balance hits zero.
• Open the API keys page, click Create key, give it a name and expiry, link it to yourself or a service account, optionally scope it to a workspace, and copy it. Keys start with sk-ant- and are shown only once.
• Export it as ANTHROPIC_API_KEY. Official SDKs for Python, TypeScript, C#, Go, Java, PHP and Ruby read it automatically; raw HTTP calls send it in the x-api-key header.
• Try prompts in the Console playground (it uses the same credits) before writing code.

For agent work, the Claude Agent SDK (the renamed Claude Code SDK) packages the tools, hooks, subagents and MCP support behind Claude Code as a Python or TypeScript library, authenticated with the same API key.

Claude API rate limits and how the tiers step up

Limits apply per organisation and are counted as requests per minute (RPM), input tokens per minute (ITPM) and output tokens per minute (OTPM). New accounts may begin in an Evaluation tier with lower limits that rise automatically as usage history builds. Above that sit Start, Build and Scale tiers, with monthly spend caps of $500, $1,000 and $200,000, plus Custom deals with no cap. Anthropic places organisations on a tier automatically from usage history and account standing, and you can apply for more through Request tier increase on the Console's rate limits page.

• Start: 1,000 RPM, 2M ITPM and 400K OTPM each for Opus 5.5, Sonnet 5.5 and Haiku 5.5.
• Build: 5,000 RPM, 5M ITPM and 1M OTPM.
• Scale: 10,000 RPM, 10M ITPM and 2M OTPM.

Fable 5.1 gets tighter ceilings, shared with Fable 5: 1,000 RPM, 500K ITPM and 100K OTPM at Start, rising to 4,000 RPM, 4M ITPM and 800K OTPM at Scale.

Capacity refills continuously, so 60 RPM may be enforced as one request per second. When a 429 carries a retry-after header, wait that many seconds. A 429 whose error_code reads enforced_spend_limit_reached is different: you have hit the monthly spend cap, and retries (including the SDK's automatic ones) fail until 00:00 UTC on the 1st of next month or until your tier changes.

Handling the 529 overloaded error

A 529 overloaded_error means Anthropic's servers are busy for everyone, not that your key or request is wrong. Retry with exponential backoff. Official SDKs retry 429s and 5xx responses twice by default; raise max_retries for background jobs. For user-facing features, add a fallback model and watch status.claude.com, which tracks the Claude API and Console separately.

Claude API vs OpenAI API on cost and coding

On list price the two are now close. OpenAI's API pricing page lists gpt-6.1-sol at $2 input and $10 output, the same as Sonnet 5.5; gpt-6-astra at $10 and $50, matching Fable 5.1; and gpt-6-luna at $0.10 and $0.50, level with Haiku 5.5 below 100K tokens. Caching is close too: both charge 1.25x input to write a cache, and Sonnet 5.5 reads match gpt-6.1-sol's $0.10 cached rate. At the top end Fable 5.1 reads cost $0.25 against $1 for gpt-6-astra, and Anthropic adds a 1-hour cache option. Both firms halve prices for batch jobs. Long prompts split them: OpenAI charges double for input and 1.5x for output once a prompt passes 272K tokens, while Sonnet 5.5, Opus 5.5 and Fable 5.1 keep one rate up to 1M tokens.

For coding, the deciding factor is usually the agent stack rather than raw model rates. Anthropic's tool use, server-side code execution and the Agent SDK come from the team behind Claude Code, so developers who already use that tool keep one mental model. Run your own evals on real tickets or pull requests before deciding.

Paying for the Claude API from India

India appears on Anthropic's Supported Regions Policy, a single list that covers the apps, the API and Claude models sold through cloud partners. Prices are in US dollars, and the billing article describes paying by card through the Console; Anthropic does not publish rupee prices or explain GST handling for API credits. Credits expire after a year and are non-refundable, so small top-ups with auto-reload beat one big purchase. Companies with AWS, Google Cloud or Microsoft accounts can also reach Claude models through Amazon Bedrock, Google Cloud or Microsoft Foundry and pay through that cloud bill. Max subscribers get $100 to $200 of monthly API credit and Team plans up to $500 pooled, but those credits expire each cycle and do not apply to Claude used through Bedrock, Google Cloud or Foundry.

What Anthropic does with your API prompts

Anthropic's commercial terms say it may not train models on customer content from its services, and that you own the outputs. The privacy centre adds two exceptions: feedback you send, and data you explicitly agree to share. API inputs and outputs are deleted from Anthropic's back end within 30 days by default. Content flagged by safety systems can be kept for up to two years, and zero data retention is offered only through a separate agreement with Anthropic.

Where the Claude API falls short

• No ongoing free tier beyond the one-time $5 credit.
• Output is text only. There is no image or audio generation endpoint.
• Evaluation-tier limits on a brand-new account can throttle a product launch.
• Cache writes cost extra, so rarely reused prefixes lose money.

Other APIs to price before you commit

The OpenAI API is the closest like-for-like rival and adds image and audio models. Google's Gemini API has a free tier for its Flash models through AI Studio, which suits prototypes. DeepSeek's API is among the cheapest per token, though its data sits with a Chinese company.