Part ofWhat Is Claude Code? The Complete Guide
In This Article
6 sectionsQuick answer
The full math behind our Claude Code cost calculator: what an hour burns in tokens, per-model API rates, plan hour caps, and worked API-vs-plan examples.
Key takeaways
- An active Claude Code hour runs roughly 1–4 million tokens, per community tooling like ccusage, and most of it is discounted cache reads.
- On Sonnet 5's $3/$15-per-MTok rates that prices out near $3/hour casual, $6 typical, $12 heavy — our calculator's three presets.
- Model choice scales it: Haiku 4.5 about 0.33× a Sonnet hour, Opus 5 about 1.67×, Fable 5 about 3.33× — pure ratios of their per-MTok pricing.
- Plans flat-rate those hours: $20 Pro covers roughly 40–80 Sonnet hours a week, $100 Max 5× about 140–280 plus 15–35 Opus hours, $200 Max 20× about 240–480 plus 24–40.
- Even 3 casual hours a week costs about $39 a month at API rates, nearly double Pro's $20, so break-even sits near one hour a week.
Claude Code cost per hour works out to roughly $3 for casual use, $6 for steady feature work, and $12 for heavy agentic sessions, in API-equivalent terms on Sonnet 5. Pro and Max flat-rate those hours instead — $20 to $200 a month — so a plan wins for almost anyone coding weekly.
Rates verified 21 August 2026 against Anthropic's pricing page. This is also the written-down methodology for our homepage plan calculator, so article and tool can never disagree.
Anthropic meters tokens, not time, so any Claude Code cost per hour figure is a conversion — the one our Claude Code cost calculator is built around. For the monthly picture, see our Claude Code cost guide.
Key takeaway
Claude Code cost per hour is an estimate, not a meter: an active hour pushes roughly 1–4 million tokens, near $3 casual, $6 typical or $12 heavy at Sonnet 5 API rates, while Pro and Max flat-rate those hours at $20–$200 a month.
What an hour of Claude Code actually burns
Roughly 1–4 million tokens per active hour, most of it cached input — the volume is the surprise, not the rates.
Claude Code is agentic: every turn re-sends the conversation, reads files, runs commands and feeds the output back, which is why its token volume dwarfs a chat session. Community dashboards like ccusage put an active hour between about one million tokens and four million or more on a large repo — ranges, not measurements, moving with repo size, agentic depth and session hygiene. Prompt caching keeps the dollars small: most of that traffic is project context re-read at a tenth of the input price.
Claude Code cost per hour: the API math
Apply the per-token rates to an hour's traffic and the presets fall out: $3, $6, and $12 on Sonnet 5.
The list rates the math runs on, current as of 21 August 2026 (full breakdown: our Claude API pricing guide):
| Model | Input / MTok | Output / MTok | Cost vs a Sonnet hour |
|---|---|---|---|
| Claude Haiku 4.5 | $1 | $5 | ~0.33× |
| Claude Sonnet 5 | $3 | $15 | 1× (baseline) |
| Claude Opus 5 | $5 | $25 | ~1.67× |
| Claude Fable 5 | $10 | $50 | ~3.33× |
Sonnet 5's introductory $2/$10 pricing runs to 31 August 2026; we use the standard $3/$15 so the math outlives it, so real Sonnet costs currently run below every figure here.
Blended, a Sonnet 5 session lands near $3 per million tokens of traffic: cheap cache reads and expensive-but-thin output roughly cancel. Applied to the volume ranges:
- Casual, ~1M tokens/hour — small edits, short sessions: ≈ $3/hour
- Typical, ~2M tokens/hour — steady feature work, medium repo: ≈ $6/hour
- Heavy, ~4M tokens/hour — long agentic runs, many files in context: ≈ $12/hour
Conservative round numbers, and adjustable in the calculator.
Per-hour cost by intensity and model
The full grid: your intensity's Sonnet rate times the model's price multiplier.
| Working style | Haiku 4.5 | Sonnet 5 | Opus 5 | Fable 5 |
|---|---|---|---|---|
| Casual (~1M tokens/hr) | ~$1 | ~$3 | ~$5 | ~$10 |
| Typical (~2M tokens/hr) | ~$2 | ~$6 | ~$10 | ~$20 |
| Heavy (~4M tokens/hr) | ~$4 | ~$12 | ~$20 | ~$40 |
On an API key these are real dollars; on a plan they are the rate you burn your weekly allowance at.
How plan caps translate into hours
Pro buys roughly 40–80 Sonnet hours a week; Max 5× about 140–280 plus a separate Opus pool; Max 20× roughly double that.
Anthropic publishes expected usage as weekly hour ranges, and its Claude Code cost documentation explains the metering:
| Plan | Price / month | Sonnet-class hours / week | Opus hours / week |
|---|---|---|---|
| Pro | $20 | ~40–80 | — (no Opus in Claude Code) |
| Max 5× | $100 | ~140–280 | ~15–35 |
| Max 20× | $200 | ~240–480 | ~24–40 |
Ranges are wide because caps meter tokens, not clock time: a heavy hour burns the allowance twice as fast as typical, a casual hour half. Premium models have their own pool, and a Fable 5 hour drains roughly two Opus-hours. See our Claude Pro vs Max and Claude Max guides.
The flat fee inverts it: ten hours a week on Pro makes your effective Claude Code cost per hour ≈ $0.46 ($20 ÷ 43.5), thirty hours on Max 5× about $0.77. That collapse is the whole argument for subscriptions.
Worked examples: API vs plan
Three honest scenarios, computed the way the calculator does it: hours × rate × 4.345 weeks.
- Casual, 3 hours a week on Sonnet. 3 × $3 × 4.345 ≈ $39/month; Pro's $20 covers it, window barely dented.
- Typical, 10 hours a week. 10 × $6 × 4.345 ≈ $261/month, inside Pro's floor. Shift a quarter of those hours to Opus 5 and it reaches about $304 while Pro drops out — Max 5× at $100.
- Heavy, 25 hours a week. 25 × $12 × 4.345 ≈ $1,304/month, and 25 clock-hours count as ~50 cap-hours: tight on Pro, comfortable on Max 5×.
Above roughly an hour of coding a week, the question is which tier — see our best Claude plan for coding guide.
The assumptions behind our calculator
Every number the tool uses, stated plainly — same figures, same review date: 21 August 2026.
The calculator runs on the rates in the first table (Sonnet at the standard $3/$15), the $3/$6/$12 presets from ccusage-style volumes with caching working, model multipliers that are pure price ratios, the plan caps above with burn scaled 0.5×/1×/2× by intensity, and 4.345 weeks per month.
These are planning estimates, not billing: true spend depends on your repo, your prompts and how well caching lands, so /cost and the Claude Console usage page are the source of truth.
Frequently Asked Questions
Written by
Edith
Writing about Claude and the Anthropic toolkit — models, Claude Code, pricing, features, and fixes, in clear, practical, hands-on guides tested by daily use.
View all posts →
