Claude API Pricing 2026: The Real Cost Per Token, Free Tier & Plan Breakdown

What Sonnet 5, Opus 5, and Haiku 4.5 actually cost per million tokens — plus the free tier, batch and caching discounts, and whether the API beats a Claude subscription for your workload.

By Oyekale Olawale

I get asked the same question every week by developers who just opened the Claude console for the first time: “why does my bill not match the number on the pricing page?” The honest answer is that Claude API pricing has more moving parts than most people expect — model choice, cache hits, batch mode, and even which data region you pick all change the final number. This guide breaks down every one of those levers with the current, verified rates straight from Anthropic’s pricing page, so you can actually predict your invoice before it arrives.

Claude API Pricing by Model (August 2026)

This is the table I keep bookmarked myself. All rates are per million tokens (MTok), billed separately for input and output. Anthropic charges no surcharge for long context — a request using most of the 1M-token window costs the same per-token rate as a short one.

*Sonnet 5 introductory pricing applies through August 31, 2026, then standard $3/$15 rates take effect. Cache figures use 5-minute TTL. An extended 1-hour cache is also available.

For most production apps, Sonnet 5 is the sweet spot — strong enough for coding and agentic work without Opus-tier pricing. I default new client projects to Sonnet and only reach for Opus when a task genuinely needs the extra reasoning depth. If you’re comparing this against coding-assistant billing models, it’s worth reading how GitHub Copilot’s token credit system works — the underlying per-token math is nearly identical, just wrapped in a different subscription shell.

How Token Pricing Actually Works

A token is roughly ¾ of an English word — “Hello, how are you?” tokenizes to about 6 tokens. Input tokens are everything you send: your prompt, system instructions, conversation history, and any documents attached. Output tokens are what Claude generates back. Anthropic bills these at different rates because generating text costs more compute than reading it — that’s why output is consistently 5x the input price across every Claude model.

Here’s a real bug I hit early on: I assumed a “1 million token context window” meant I’d get billed a flat rate no matter the prompt size. That’s not how it works — a 900,000-token request costs the same per-token rate as a 9,000-token one, but you’re still paying for every single token you send, not a flat window fee. Long context is free of surcharge, but it is never free of cost.

Prompt Caching: The Biggest Lever You’re Probably Not Using

If your app repeats the same system prompt, reference document, or few-shot examples across many requests, prompt caching is where the real savings live. A cache read costs just 10% of the standard input price — for Opus 5, that’s $0.50 instead of $5.00 per million tokens, a 90% saving. The catch: a 5-minute cache write costs 1.25x the base input rate, so it only pays off once you hit that cache more than once. Reorder your prompt so the stable content stops leading, though, and every “cached” segment silently reverts to full price — I’ve seen this quietly triple a client’s monthly bill because nobody noticed the prompt template got refactored.

Other Costs You’ll Hit Beyond Tokens

Is There a Claude API Free Tier?

Not an ongoing one — this trips up almost everyone searching for “Claude API pricing free.” What actually exists: new accounts created at platform.claude.com receive $5 in free credits with no credit card required, and those credits don’t expire. That’s enough to test roughly 5 million Haiku input tokens or around 1 million Opus input tokens — plenty to prototype an integration, but nowhere near enough for production traffic.

Once that $5 runs out, you’re on standard pay-as-you-go billing. There’s no free-forever API plan the way there is for the claude.ai chat interface, where the Free plan gives ongoing (rate-limited) access at $0. The Free chat plan includes web chat, code generation and data visualization, web search, memory, and file execution — but that’s a completely separate product from API access, billed and gated differently.

Claude API vs. Claude Subscription Plans

This is the comparison people actually need before spending a dollar. The subscription plans (Free, Pro, Max, Team, Enterprise) give you the claude.ai chat interface and Claude Code under one flat monthly fee. The API is metered — you pay per token with no ceiling and no subsidy.

For the full breakdown of every subscription tier — Free, Pro, Max 5x/20x, Team, and Enterprise, including seat pricing and student access — see our complete Claude AI pricing guide. This article focuses on API/token costs specifically.

Here’s the part most pricing guides skip: a flat-fee subscription and metered API access on the same model are not economically equivalent, and the gap is bigger than people expect. A Pro subscriber running heavy agentic sessions on Sonnet inside their $20/month plan can rack up compute that would cost hundreds of dollars at raw API rates — a subsidy some analysts have pegged at 15 to 30 times the subscription price for the heaviest users. That’s exactly why Anthropic has tightened usage limits on subscription plans over time: the flat fee only works if usage stays within a reasonable band.

The practical rule I give clients: if you’re a single person chatting with Claude, doing coding sessions, or running Claude Code daily — get Pro or Max. If you’re building a product other people will use, go API, because a subscription cannot be legally resold or embedded in an app; it’s licensed for the account holder’s personal use. There’s no gray area here, and no version of “share one Max plan across your SaaS users” survives a terms-of-service review.

Claude API Pricing vs. ChatGPT (OpenAI) API

On raw list price, OpenAI usually undercuts Claude at the flagship tier, but the two companies structure their lineups differently, which makes head-to-head comparisons tricky. Here’s where things stand as of August 2026:

*OpenAI’s Terra and Luna tiers received price cuts in late July 2026 — Terra dropped 20% and Luna dropped 80% from launch pricing, while Sol held its rate.

The number that rarely makes it into these comparisons: Anthropic’s 1M-token context window on Sonnet and Opus 5 comes at no premium, which removes a lot of the chunking and re-summarization work you’d otherwise pay for indirectly. If your workload leans on long documents, large codebases, or extended agent memory, that architectural difference tends to matter more than the sticker price gap. If you’re weighing model quality rather than just cost, our breakdown on whether ChatGPT is actually good for coding and our Claude vs. GPT-4 comparison both dig into that side of the decision.

Real-World Cost Examples

Numbers on a pricing page don’t mean much until you see them applied. Here are three scenarios I actually calculate for clients:

What Developers on Reddit Actually Say About Claude API Pricing

The recurring theme across developer forums isn’t that Claude is overpriced — it’s that agentic workflows consume tokens far faster than people expect until they’ve watched a real invoice. The most common complaints fall into three buckets: surprise bills from long agent sessions that weren’t using caching correctly, confusion about subscription vs. API billing being separate systems, and frustration that fast mode and long-context surcharges aren’t obvious until the bill lands. The fix for all three is the same: enable prompt caching deliberately, default to Sonnet or Haiku unless a task truly needs Opus-level reasoning, and set a hard spending cap in the console before you scale traffic.

How to Get Started With the Claude API: Step-by-Step

1. Create an account at platform.claude.com — no credit card required to claim your starting credits.
2. Generate an API key from the Console under API Keys, and store it as an environment variable, never hard-coded.
3. Pick a starting model — Haiku 4.5 for testing and simple tasks, Sonnet 5 for most production work.
4. Send your first request to the /v1/messages endpoint with your model, max_tokens, and messages array.
5. Enable prompt caching on any repeated system prompt or reference document before you scale traffic.
6. Set a spend limit in Console settings so a runaway loop can’t drain your balance overnight.
7. Move batch-eligible work — summarization, classification, evaluation — to the Batch API for the automatic 50% discount.

7 Ways to Cut Your Claude API Bill

FAQ

How much does the Claude API cost per token?

Prices are quoted per million tokens (MTok), not per single token. Sonnet 5 costs $2 input / $10 output per MTok on introductory pricing (rising to $3/$15 after August 31, 2026); Opus 5 is $5/$25; Haiku 4.5 is $1/$5.

Is there a Claude API free tier?

There’s no ongoing free API tier. New accounts get roughly $5 in one-time free credits with no credit card required, which don’t expire but are limited compared to production usage.

Is Claude API pricing different from Claude Pro or Max subscriptions?

Yes, completely separate. Subscriptions are flat monthly fees for the chat interface and Claude Code with usage limits; the API is metered per token with no subscription and no usage cap.

Is Claude API pricing cheaper than ChatGPT’s API?

It depends on the tier. OpenAI’s flagship GPT-5.6 Sol ($5/$30) costs slightly more on output than Claude Opus 5 ($5/$25). At the budget tier, OpenAI’s Luna model currently undercuts Haiku 4.5. Neither provider is cheaper across every tier.

Should I use the API or a subscription plan?

If you’re an individual chatting with Claude or coding solo, a Pro or Max subscription is almost always cheaper. If you’re building a product for other users, you need the API — subscriptions can’t legally be embedded in or resold through an app.

Does prompt caching actually save money?

Yes — cache reads cost 10% of the standard input price, a 90% saving on repeated content. It only pays off once you hit the same cached prefix more than once, since the initial write costs slightly more than a normal request.

What’s the cheapest way to run high-volume Claude workloads?

Combine Haiku 4.5, the Batch API’s 50% discount, and prompt caching’s 90% cache-read discount together — these three stack and can cut effective costs by well over 90% versus real-time Opus calls.

Conclusion

Claude API pricing is genuinely competitive once you use the levers Anthropic built in — caching, batching, and smart model routing. The mistake I see most often isn’t picking the “wrong” model, it’s ignoring caching entirely and letting every request hit full input price. Start with Haiku or Sonnet, cache anything repeated, batch anything that isn’t real-time, and check your Console spend weekly. Do that, and your bill will track your actual usage instead of surprising you at the end of the month. If you’re deciding between a subscription and metered API access, start with our full Claude AI pricing guide covering every plan tier in detail. If content generation and AI tooling costs are part of your broader research, our guide on generating code with generative AI and our review of whether Claude AI is safe and trustworthy are good next reads. For teams tracking how AI visibility affects organic traffic, see our piece on how Google AI Overviews are reshaping SEO and our roundup of the best LLM SEO tracking software. And if you’re comparing coding assistants more broadly, check our Verdent alternatives roundup alongside our Grok 4 vs. GPT-5 and Gemma 4 vs. Gemma 3 comparisons.