Per-token prices for Opus, Sonnet, and Haiku. Real-world subscription numbers. Where prompt caching saves 30-70 percent. And how to see what you're actually paying — live, by project — without waiting for a monthly invoice.
All prices in USD per million tokens. Output tokens are charged at 4-5x the input rate, so writing-heavy use cases cost more than reading-heavy ones.
| Model | Input | Output | Cache write | Cache read | Best for |
|---|---|---|---|---|---|
| Claude Opus 4.8 | $5.00 | $25.00 | $6.25 | $0.50 | Hardest reasoning, complex refactors |
| Claude Sonnet 4.6 | $3.00 | $15.00 | $3.75 | $0.30 | Daily coding work, speed/intelligence balance |
| Claude Haiku 4.5 | $1.00 | $5.00 | $1.25 | $0.10 | High-throughput / cheap classification |
Last updated: June 2026. Prices reflect current public pricing. Verify on the Anthropic pricing page before locking in a budget.
Subscription pricing is almost always cheaper than per-token API billing for individuals. The math flips at high volume or when you need the higher-tier API rate limits.
Base 5-hour rolling window. Plenty for moderate daily use. Cheapest entry point.
Individual devs, casual Claude Code use
5x the Pro allowance. Designed for developers who use Claude Code as their primary coding tool.
Daily heavy users, professional engineers
20x the Pro allowance. Effectively eliminates rate limit interruptions for solo work.
All-day coding sessions, power users
Real numbers from instrumented Claude Code sessions. Your mileage will vary depending on project size, model mix, and how often Claude has to re-read context.
Cache reads dominate; output volume small.
Most developers fall here. ~6 hours of focused coding.
Opus output tokens drive most of the cost.
Caching saved an estimated $90 vs naive replay.
Anthropic charges 25 percent more for the first request that creates a cache entry, but 90 percent less for every subsequent read. For Claude Code workflows that re-use the same context across many requests, this is the single biggest savings opportunity.
10 requests, each with 100K tokens of repeated codebase context. Sonnet pricing.
10 × 100K input × $3/M = $3.00
First request writes the cache, next 9 read it. Same 100K token context.
$0.375 cache write + 9 × $0.030 cache read = $0.65 — 78% cheaper
The Anthropic Console shows yesterday's cost. Tokemon shows what your tokens are doing right now — broken down by project, with live burn rate and a forecast of where you'll land for the month. Free, open source, runs in your menu bar.
On a Claude subscription: Pro is $20/month, Max is $100 or $200/month, Team is $30 per seat. On the API, a typical day of intensive Claude Code use ranges from $5 to $40 depending on model mix and how aggressively you use prompt caching. For most individual developers, a Pro or Max subscription is cheaper than per-token API billing.
Claude Opus 4.8 is $5 per million input tokens and $25 per million output tokens. Claude Sonnet 4.6 is $3 per million input and $15 per million output. Claude Haiku 4.5 is the cheapest at $1 per million input and $5 per million output. Output tokens are 5x more expensive than input across all three.
Yes, significantly. Prompt cache writes cost 25 percent more than a normal input token, but cache reads cost 90 percent less. For long-running coding sessions where the same context (files, conversation history) is reused across many requests, caching can reduce total spend by 30-70 percent.
Claude Pro at $20/month is almost always cheaper than the equivalent API spend for individual developers. The 5-hour rolling window gives you generous capacity, and there are no per-token charges. Move to the API only when you need programmatic access, want to bypass rate limits with higher tiers, or are building a product on top of Claude.
On the Anthropic Console you can see daily API spend after the fact. For real-time visibility — including which projects are consuming the most tokens — install Tokemon. It tracks per-project costs from your local Claude Code session logs and shows live burn rate, so you can see costs accruing as you work.
Combine three things: subscribe to Claude Pro instead of using the API, use Sonnet (not Opus) for routine work and only escalate to Opus for hard problems, and let Claude Code's automatic prompt caching do its job by working in focused sessions on a single codebase.
Deep dive into per-token math, caching, and worked examples.
Concrete cost-cutting tactics — model selection, caching, batch.
Side-by-side breakdown of subscription tiers and rate limits.