Anthropic API Review 2026

What Anthropic actually sells you

The Anthropic API serves the Claude model family through the Messages API, run by Anthropic from San Francisco. In 2026 the developer surface was renamed. The Claude Console lives at platform.claude.com, the documentation at platform.claude.com/docs, support at support.claude.com, and the old console.anthropic.com address redirects. The lineup runs from Haiku 5.5 at the bottom, through Sonnet 5.5 and Opus 5.5, to Fable 5.1 at the top, with Mythos variants gated behind verification programs.

The same models are sold through four other channels, which matters for credits. Claude Platform on AWS is Anthropic-operated and billed through AWS Marketplace. Claude in Amazon Bedrock and Claude Models on Google Cloud (Google managed) are operated and invoiced by the cloud providers. Claude in Microsoft Foundry sits on Azure. Anthropic’s own grants, including plan credits and startup credits, apply only to the first-party API through the Console. Pick your door before you apply for anything.

What free credit you actually get

The new-account credit exists but is deliberately vague. Anthropic says new users receive a small amount of free credits to test the API and nothing more; no amount, no expiry. Alongside it, every new organization starts on an Evaluation tier with rate limits below the published Start tier, which Anthropic describes as fraud prevention that lifts automatically with usage history.

The credit that moves the needle is the monthly plan credit. Max subscribers get $100 a month, Ultra $200, and Team organizations pool $20 per Standard seat and $100 per Premium seat up to $500. You link one Console organization from Settings, Billing on claude.ai after seven days on the plan. Credits land shortly after each payment clears, are spent before purchased credits, and expire at the end of the billing cycle.

Scope is the catch. Plan credits cover the Claude API, the Playground, Managed Agents and the Agent SDK, but not interactive Claude Code in the terminal or IDE, not extra usage in the Claude apps, and not Bedrock, Vertex AI or Foundry. Each Console organization can receive credits from only one plan, and changing the linked organization goes through support.

How tiers and spend caps limit you

Anthropic publishes four tiers after Evaluation. Start carries a $500 monthly spend cap, Build $1,000, Scale $200,000, and Custom has none. On Start, Opus 5.5, Sonnet 5.5 and Haiku 5.5 each get 4,000 requests, 2 million input tokens and 400,000 output tokens per minute; Fable 5.x gets 4,000 requests but 500,000 input and 100,000 output. Build raises the 5.5 models to 10,000 requests, 5 million input and 1 million output, and Scale to 13,000, 10 million and 2 million.

Two details make these numbers more generous than they look. Cache reads don’t count toward input tokens per minute on current models, so a well-cached agent can process several times its nominal limit. And there’s no deposit that moves you up; tiers are assigned from usage and account standing, though a linked Team plan places an organization on at least Build. Reaching a spend cap pauses usage until the first of next month, and the Request tier increase button on the Rate limits page is the fix.

What it costs per million tokens

Haiku 5.5 costs $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens and $0.50 and $2.50 above that. Sonnet 5.5 is $2 and $10, the same as Sonnet 5, whose introductory price was made permanent instead of rising on September 1, 2026. Opus 5.5 is $4 and $20, 20% below Opus 5, and Fable 5.1 is $10 and $50. Web search adds $10 per 1,000 searches.

Modifiers stack. Batch processing halves input and output on every model. Cache writes cost 1.25x base input for the 5-minute cache and 2x for the 1-hour cache, while cache reads cost 10% of input on most models, 5% on Opus 5.5 and Sonnet 5.5, and 2.5% on Fable 5.1. Fast mode on Opus 5.5 is 2x standard, and US-only inference through inference_geo is 1.1x on all token categories for Claude 4.6 and later. Models from Claude 4.7 onward use a tokenizer that yields roughly 30% more tokens per text, so compare on task cost, not token price.

Billing is prepaid through the Billing page at platform.claude.com/settings/billing. Credits are bought with a card, available immediately, non-refundable, and expire one year from purchase. Auto-reload tops up when the balance falls under a threshold you set. Claude Platform on AWS inverts this model: usage is metered in Claude Consumption Units and invoiced monthly in arrears on the AWS bill, with no prepaid balance.

Credits for startups, researchers and labs

Claude Startups is the headline program and it’s currently closed. On October 6, 2026 Anthropic expanded it to include a free year of Claude Team with up to five Premium seats for companies new to Team, valued at $6,000, a one-time $1,000 API credit on approval, and Startup Stack partner offers worth up to $45,000. Eligibility was startups founded in the last five years or funded in the last two. An update dated October 13 says the cohort is full and applications aren’t being accepted. Founders backed by a VC in Anthropic’s partner network can still ask their investor about up to $100K in additional credits.

Research programs stay open on a monthly cycle. AI for Science applies up to $20,000 in API credits for six months to academic and nonprofit researchers on high-impact science, screened for merit, feasibility and biosecurity. The External Researcher Access Program grants $1,000 in API credits for AI safety and alignment work. Both review applications on the first Monday of each month and exclude nonpublic models and fine-tuning.

Education is plan-based rather than credit-based. Claude for Education is a campus-wide plan arranged through the education sales team, and it includes dedicated API credits for student learning without a stated amount. Separately, principal investigators at accredited universities can get Claude Team Standard seats at $0 for 12 months, with API credits available only through the AI for Science grant.

Where your credits do and don’t apply

If your company already has AWS or Google Cloud commitments, the marketplace paths are the route. Claude Platform on AWS provisions a new Anthropic organization tied to your AWS account, starts on the Start tier, and bills through AWS Marketplace. Claude in Amazon Bedrock is an AWS-native service with a default quota of 2 million input tokens per minute, regional endpoints at a 10% premium, and AWS as the data processor. Claude Models on Google Cloud (Google managed) is invoiced by Google, with multi-region and regional endpoints 10% above global.

None of these paths accept Anthropic grants. Plan credits, startup credits and research credits apply to the first-party API only. Conversely, Anthropic prepaid credits can’t pay a Bedrock or Google Cloud bill. Pick the channel by where your money already sits, and remember that Bedrock and Google managed endpoints don’t support the Message Batches API, so the 50% batch discount is first-party only.

Errors, data and other surprises

The two errors that confuse new users are both 429s. A rate limit 429 carries a retry-after header; a spend cap 429 doesn’t, and its error_code is enforced_spend_limit_reached, meaning usage is paused until 00:00 UTC on the first of next month. A self-set spend limit returns a 400 instead. Acceleration limits can also trigger 429s when traffic spikes, so ramp gradually rather than flipping a switch.

Data handling is stricter than most platforms. Prompts and outputs aren’t retained by default and are never used for training without express permission. The exception is Covered Models, including Fable 5.1, which require 30-day retention and aren’t available under zero data retention. Inputs and outputs are deleted within 30 days, with flagged content kept up to two years. Zero data retention is arranged through sales. Support runs through support.claude.com, with no published response times for self-serve accounts.

Who it’s worth it for

Yes, at 4.2, for anyone with a Claude plan or a caching-friendly workload. A Max subscriber effectively gets the API free up to $100 a month, Sonnet 5.5 at $2 and $10 with 5% cache reads is cheap for agents, and batch plus cache discounts compound. We hold back from higher because of the unquantified starter credit, Evaluation tier limits, a closed startup program and one-year credit expiry. Founders hoping for startup credits this quarter should subscribe to the program newsletter and build on plan credits in the meantime.