Here is what the Claude API costs in 2026, per 1M tokens: Claude Opus 4.7 runs $5 input and $25 output, Claude Sonnet 4.6 runs $3 input and $15 output, and Claude Haiku 4.5 runs $1 input and $5 output. The Batch API halves all of those output prices for asynchronous work, and prompt caching cuts repeated input to roughly a tenth of the list rate. A typical production workload lands anywhere from $40 a month (Haiku with batch) to several hundred (Opus agents), and the difference is almost entirely about which model you pick and which discounts you turn on.
These numbers come from our own tracked dataset, re-checked weekly against official pricing pages, so they are current as of this writing. When Anthropic moves a price, our live rankings move with it.
Claude pricing at a glance
Per 1M tokens, list price versus Batch API, as of our July 2026 check:
| Model | Input | Output | Batch output |
|---|---|---|---|
| Claude Opus 4.7 | $5 | $25 | $12.50 |
| Claude Sonnet 4.6 | $3 | $15 | $7.50 |
| Claude Haiku 4.5 | $1 | $5 | $2.50 |
Three things to internalize about this table:
- Output tokens dominate. They cost 5x input on every tier, and long completions or agentic tool loops generate a lot of them. When you estimate a bill, get the output number right first.
- The tiers are 5x apart. Opus to Sonnet is a 40 percent cut on input and output; Sonnet to Haiku is a 66 percent cut. Most teams run far more traffic on Opus than the task actually needs.
- Batch is a flat 50 percent off. Anything that can tolerate results arriving within 24 hours (evaluations, content pipelines, backfills, nightly jobs) should never pay list price.
For the live table including resellers like OpenRouter and Bedrock, see our cheapest Claude Opus 4.7 ranking.
What a real monthly bill looks like
Token math is abstract until you run it on an actual workload. Here are three realistic shapes.
Example 1: customer support chatbot on Sonnet 4.6
Say your bot handles 40,000 conversations a month, averaging 500 input tokens and 100 output tokens per turn with 10 turns per conversation. That is 200M input tokens and 40M output tokens.
- Input: 200M x $3 per 1M = $600
- Output: 40M x $15 per 1M = $600
- List total: about $1,200 a month
Now turn on prompt caching. A support bot resends the same system prompt and conversation history every turn, so in practice 70 to 90 percent of that input bills at the cached-read rate of roughly $0.30 per 1M instead of $3. If 160M of the 200M input hits cache, input drops to about $168. New total: roughly $770, a 36 percent cut from one config change.
Example 2: bulk classification on Haiku 4.5 with batch
You classify 2M documents a month at 500 input tokens and 50 output tokens each: 100M input, 10M output.
- List: 100M x $1 + 10M x $5 = $150 a month
- Batch: $50 + $25 = $75 a month
That is frontier-lab quality classification for $75. If you were running this on Opus out of habit, the same job would list at $750. Model selection is the biggest lever on this page.
Example 3: coding agent on Opus 4.7
Agents are expensive because every tool call round-trips the growing context. A team of five engineers running an Opus-backed agent might push 60M input and 8M output tokens a month.
- Input: 60M x $5 = $300
- Output: 8M x $25 = $200
- List total: about $500 a month
Agents are also where caching pays off most: the system prompt, tool definitions, and conversation prefix repeat on every single call. Well-placed cache breakpoints routinely cut agent input costs by 70 percent or more, bringing this bill closer to $270.
The two levers that halve your bill
Prompt caching
Cache reads bill at about a tenth of the input rate; cache writes cost about 1.25x for the default five-minute window. The break-even is two requests, so any prompt prefix that repeats (system prompts, few-shot examples, long documents, agent tool definitions) should carry a cache breakpoint. The catch is that caching is a strict prefix match: put stable content first, volatile content (timestamps, per-request IDs) last, or your hit rate silently drops to zero. Full mechanics in our guide to prompt caching and batch discounts.
The Batch API
Submit requests asynchronously, get results within 24 hours (usually under an hour), pay half price on both input and output. It supports the full feature set including vision and caching, and the discounts stack. There is no clever engineering here; it is a different endpoint and a polling loop. If your workload is not user-facing and you are paying list price, you are donating 50 percent margin to nobody.
How Claude compares to the field
At the flagship tier, Opus 4.7 at $25 output undercuts GPT-5.5 at $30 and sits above Gemini 2.5 Pro at $10. At the mid tier, Sonnet 4.6 matches GPT-5.4 at $15. Haiku 4.5 at $5 is pricier than Gemini 2.5 Flash but punches above its weight on coding and instruction following. The full cross-vendor picture, including open-weight models that cost cents instead of dollars, is in our LLM API pricing comparison.
Do not pay cash before you claim credits
Anthropic runs startup credit programs through accelerators, VCs, and cloud partners, and Claude is also reachable through AWS and Google Cloud, which means startup cloud credits can cover Claude usage on Bedrock and Vertex. Before you put a card down, read our Claude for Startups guide and the rundown of Claude API free credits. Credits programs regularly cover months of the example bills above.
Claude credits are one entry among 200+ hand-verified AI and startup perks we track in the Perkstack catalog, collectively worth over $5.5M. Membership pays for itself if it surfaces even one credit program you did not know about; you can browse the catalog or create an account to see what your stack qualifies for.
Bottom line
Claude costs $1 to $5 per 1M input tokens and $5 to $25 per 1M output tokens depending on tier, with batch cutting output in half and caching cutting repeated input by about 90 percent. A tuned workload (right model tier, caching on, batch where latency allows) usually runs at a quarter to a half of the naive list-price estimate. Prices move, so check the live rankings, re-checked weekly, before you commit volume, and claim your startup credits before spending real money.
Related reading: Claude for Startups guide, LLM API pricing comparison, cheapest Claude Opus 4.7 API.