Last verified: 6 August 2026, with the Sonnet 5 pricing re-checked on 19 August 2026. Every price and limit below was checked against Anthropic's own pricing and rate-limit docs on those dates. Two things moved since the May version: Opus is no longer the top of the price ladder, and the usage tiers have been renamed. A third has since been corrected — the September rise on Sonnet 5, reported here as scheduled, was cancelled by Anthropic and this page previously said otherwise.
No. The Claude API isn't free, at least not in any ongoing way. New Anthropic Console accounts get a small bit of trial credit to kick the tires, and after that it's pay-as-you-go — you prepay credits and get billed per million tokens. There's no permanent free tier, nothing like the "1,500 requests a day, forever" deals you see on some other APIs. The better news is that it's cheap for light use, and there are a few legitimate ways to keep your bill near zero. Here's how it actually works in 2026.
One disclaimer up front: Anthropic changes pricing and limits from time to time. This post is about the structure — what you're paying for, how the rate limits scale — not about exact dollar figures. Always confirm current prices on Anthropic's pricing and docs pages before you budget anything off it, including this article.
The short answer
Free forever? No. Free to start? Yes — new Console accounts get a one-time-ish trial credit, enough to build a prototype, not enough to run a product. After that, it's pay-as-you-go: prepay credits, billed per million input and output tokens, with the price depending on which model you call.
Is it cheap? For light use, yes. A short completion is a fraction of a cent. For heavy agentic work — big context, long outputs — it adds up fast.
And if what you actually want is "Claude, but flat-rate instead of metered," you probably want a subscription, not the API: Claude Pro/Max, or Claude Code running on it. That's a fixed monthly fee with a usage allotment, which is what a lot of people mean when they search "is the Claude API free."
What "free" actually means here
When people ask whether the Claude API is free, they're usually asking one of a few different things. Honest breakdown:
The new-account trial credit. Create an account in the Anthropic Console and you get a small amount of credit to play with. It's real, it's free, and it's the closest thing to a free tier — but it's a one-time grant, not a recurring quota. Treat it as enough to validate an idea, not run one.
Prompt caching and the Batch API. Not "free," but the biggest lever you've got on cost. A cache read costs 10% of the standard input price — 90% off — on the parts of your prompt that repeat across calls: system prompts, long context, few-shot examples. Writing to the cache costs more than a normal input token, 1.25x for the five-minute cache and 2x for the one-hour one, so caching pays for itself after a single read on the five-minute tier. The Batch API is exactly 50% off both input and output for asynchronous work. The two discounts stack, per Anthropic's own docs, so a batched job hitting a warm cache is paying a fraction of list. (We get into cross-provider cost-cutting in our free AI APIs for developers roundup.)
A subscription instead of the API. If your use case is interactive — chatting with Claude, coding with Claude Code — you probably don't want the API at all. Claude Pro is $17/month billed annually or $20/month billed monthly; Claude Max starts at $100/month for either 5x or 20x Pro's usage. Claude Code runs on that subscription (or on an API key, your call). Worth knowing that Anthropic announced a change to how Agent SDK usage is billed against subscriptions and then paused it on 15 June 2026 — as of this check, claude -p, the Agent SDK and third-party apps still draw from your subscription allotment rather than metered API billing. That's an area they've been actively tinkering with, so it's not guaranteed to stay. For a solo dev doing day-to-day work, the subscription usually beats metering every token. This is genuinely the answer for most people asking "is it free": it isn't, but it's flat, which feels close enough.
Claude through a cloud marketplace. You can call Claude via AWS Bedrock, Google Cloud Vertex AI, and similar. Still metered usage — but if you're a new cloud customer, your provider's signup credits may apply, which is effectively free Claude until those run out.
Student and startup credit programs. Where they exist, accelerator and education programs sometimes bundle Anthropic credits. Worth checking if you qualify. Don't build a business plan on it.
What you won't find is a "use it free up to N requests a day, forever" tier. That's not how Anthropic prices the API.
How the pricing works
Three things decide what a call costs.
First, which model you pick. Here are the current per-million-token prices, checked against Anthropic's pricing docs on 6 August 2026:
| Model | Input / MTok | Output / MTok |
|---|---|---|
| Claude Haiku 4.5 | $1 | $5 |
| Claude Sonnet 5 | $2 | $10 |
| Claude Sonnet 4.6 | $3 | $15 |
| Claude Opus 5 | $5 | $25 |
| Claude Fable 5 | $10 | $50 |
One thing in that table is worth more than the rest of this article. Sonnet 5 launched at $2/$10 as introductory pricing due to end on 31 August 2026, and a rise to $3/$15 was scheduled for 1 September. That rise has been cancelled: Anthropic's pricing docs now state that the $2/$10 rate "is now the standard price" and that the increase "will not occur". Size a Sonnet budget off $2/$10.
And Opus is no longer the top of the ladder. Claude Fable 5 sits above it at $10/$50, double Opus 5. The May version of this article described a three-rung ladder — Haiku, Sonnet, Opus — and that's now a four-rung one. If you've been assuming Opus is your worst case on cost, it isn't.
The gap between cheapest and priciest is ten times on input and ten times on output. Picking the smallest model that actually does your job is the single biggest budget decision you'll make.
Second, input versus output tokens. You're billed separately for what you send (input) and what Claude generates (output), and output tokens cost more — often several times more. An app that sends a huge prompt and gets a one-line answer back is cheap. An app that sends a small prompt and gets a 2,000-word essay is not.
Third, how much context you send, which is the one that sneaks up on people. Every API call is stateless — there's no server-side memory — so if you're building a chatbot or an agent, you re-send the conversation history (and any retrieved docs, and the system prompt) on every single turn. A 20-turn conversation means turn 20 is paying for ~19 turns of history as input. Agentic coding tools are the extreme version: they shovel file contents, tool outputs, and long instructions into context, which is exactly why a serious coding session costs real money. Context size is the driver. Get a handle on it and the bill drops.
Rough intuition: one short Q&A on a small model is a fraction of a cent. A typical Sonnet request with moderate context is in the cents. A long agentic session with lots of files in context and long generated diffs can be a few dollars. Your actual numbers depend entirely on your workload — the only way to really know is to run a realistic week and look at the dashboard.
Rate limits, explained
Anthropic uses usage tiers. As of August 2026 they're named Evaluation, Start, Build, Scale and Custom — if you've read an older guide referring to "Tier 1" through "Tier 4," that naming is gone. New organisations land in Evaluation, whose limits Anthropic doesn't publish ("limits below the standard limits shown on this page"). Everything above it is documented. For Claude Sonnet 5, per the rate-limits docs:
| Tier | Requests/min | Input tokens/min | Output tokens/min | Monthly spend cap |
|---|---|---|---|---|
| Start | 1,000 | 2,000,000 | 400,000 | $500 |
| Build | 5,000 | 5,000,000 | 1,000,000 | $1,000 |
| Scale | 10,000 | 10,000,000 | 2,000,000 | $200,000 |
Note the last column, because it surprises people: the Start tier caps you at $500 of API spend per month whether you want the ceiling or not. Limits are also per-model, not one blanket number for your org — Claude Fable 5 runs lower limits at the same tier (1,000 RPM but only 500,000 input tokens/min).
One genuinely useful detail that isn't widely known: cached input doesn't count toward your input-tokens-per-minute limit on any current model. Anthropic's own worked example is a 2,000,000 ITPM limit with an 80% cache hit rate effectively processing 10,000,000 input tokens per minute. If you're rate-limited rather than cost-limited, prompt caching is the fix for both.
You can request increases through the Console if you've got a legitimate need outgrowing the automatic tiering.
What that means in practice: don't load-test from a fresh account, because you'll hit the per-minute caps immediately and conclude the API is "rate-limited" when really you're just on tier one. Build retry-with-backoff from day one — when you do hit a limit, the API tells you, and a well-behaved client backs off rather than hammering. And plan your launch around tiering: if you expect a spike, get your account spending (and aging) ahead of time, or request a limit increase before launch, not the morning of.
Exact rate-limit numbers per tier change, so check the current docs rather than trusting a number you read in a blog post. (Yes, including this one — which is why the date at the top of this page is the date every number on it was last checked.)
Models retire, too, and faster than most people plan for. Claude Opus 4.1 was retired on 5 August 2026, the day before this refresh. Opus 4 and Sonnet 4 went on 15 June 2026; Sonnet 3.7 and Haiku 3.5 on 19 February 2026. Anthropic commits to at least 60 days' notice, and requests to a retired model ID fail outright on the first-party API even where the same model lingers on Bedrock or Google Cloud. If you hardcoded a dated model string a year ago, go and check it.
How to use Claude without burning money
A few habits do most of the work. Turn on prompt caching if any meaningful chunk of your prompt repeats across calls — it's the big, easy win. Run non-urgent work through the Batch API for roughly half off: classification jobs, bulk summarization, overnight processing. Pick the smallest model that works — Haiku-class for extraction, classification, routing; Sonnet for most coding and reasoning; Opus only when the task genuinely needs it. Test down the ladder, not just up. Trim your context — don't re-send the entire conversation when a summary will do, don't dump whole files when a function will do. Set a monthly spend cap in the Console so a runaway loop or a bad deploy can't drain the account. Watch the usage dashboard, especially after you ship something new — costs that surprise you in week one are cheap to fix; costs that surprise you in month three aren't. And for interactive use, lean on a subscription: if you're personally coding with Claude all day, Claude Code on Pro/Max is almost certainly cheaper than the API. Our Claude Code tips post covers getting the most out of it, and the Claude Code vs Cursor vs Copilot comparison puts it next to the alternatives.
API or subscription?
Quick way to decide. If you're building an app or automation that calls Claude programmatically, you need the API — there's no subscription path for that. You'll pay per token; optimize with caching, batching, and model choice. If you're personally coding, chatting, or doing day-to-day work with Claude, get a subscription (Claude Pro/Max, Claude Code). Flat fee, big allotment, no per-token anxiety — and this is what most "is the API free" askers actually want. If you're just prototyping or exploring, start with the new-account trial credit on the API, or the free tier of the Claude consumer app, then decide. And if you're already deep in AWS or Google Cloud, consider Claude via Bedrock or Vertex AI so it rolls into your existing cloud bill, possibly against your cloud credits.
TL;DR
The Claude API isn't free on an ongoing basis. New accounts get a small one-time trial credit; after that it's pay-as-you-go — prepaid credits, billed per million tokens. Cost comes down to model choice times (input plus output tokens) times how much context you send; output costs more than input, and context size is the main driver. Rate limits scale by usage tier — new accounts start low and get promoted automatically as they age and spend — so build backoff/retry and don't load-test from a fresh account. The cheapest way to use it: prompt caching, the Batch API, the smallest model that works, trimmed context, a spend cap, and an eye on the usage dashboard. And if you want flat-rate instead of metered Claude, that's a subscription (Claude Pro/Max, Claude Code), not the API.
Pricing and limits change. Confirm the current numbers on Anthropic's official pages before you budget.
Want the wider view — which AI APIs actually have usable free tiers, and which "free" offers come with a catch? Read Best Free AI APIs for Developers (2026) — With Real Rate Limits. For non-AI freebies, Every Free Credit and Promo Code for Developer Tools is the companion list. More in the AI Tools section.
Top comments (0)