Claude Code pricing 2026 — subscription vs API, per-model rates, and what a month costs

Claude Code has two entirely separate prices, and most of the confusion about what it costs comes from mixing them up. A claude.ai subscription bundles it into a flat monthly fee with usage limits. An API key bills per token with no monthly fee and no limits. This page has the current numbers for both, the arithmetic for deciding which one you want, and what a real month actually costs.

Rates verified August 5, 2026 against claude.com/pricing and Anthropic's API pricing documentation. Kunavo's per-token prices are read live from the model catalog.

TL;DR — which one is for you

They do not stack. While a credential variable is set, Claude Code uses the key and your subscription sits parked — its limits stop applying and usage bills to the key instead. Run /status to see which one is live.

Route 1 — subscription pricing

Claude Code requires a Pro, Max, Team, Enterprise or Console account; Anthropic's setup documentation states the free Claude.ai plan does not include Claude Code access. Plan fees as published on claude.com/pricing:

The important mechanic is the limit, not the fee. Usage on a plan is metered against a rolling five-hour window and a weekly window, both shared with Claude chat — so an afternoon in the web app eats into the same allowance as your coding session. Switching models with /model does not reset a window, because the windows span all models.

Route 2 — API pricing, per model

On a key there is no plan and no window: Claude Code is billed as ordinary Claude API traffic. Files and tool output count as input tokens; edits, explanations and thinking tokens count as output. Rates per 1M tokens:

One caveat stated plainly: Anthropic is running an introductory rate of $2 / $10 per 1M on Claude Sonnet 5 through August 31, 2026, which is below Kunavo's $2.10 / $10.50. From September 1 it reverts to the $3.00 / $15.00 list price above and Kunavo is about 30% under again. Full per-model detail is in the Claude API pricing guide.

Caching is part of the price

Claude Code re-sends the whole conversation on every step, so most of what it bills is repeated context — which is exactly what prompt caching is for. Anthropic's published multipliers apply to Claude Code traffic like any other:

In a warm session most input tokens are cache reads at a tenth of the rate, which is why a naive tokens × price estimate overstates the bill badly. It also explains the single worst cost pattern in Claude Code: coming back after a long break misses the cache and reprocesses the entire context at full price.

What a month actually costs

Cost tracks agentic steps, not calendar days. One step is a model round trip inside a task — roughly 25,000 input and 1,200 output tokens once the system prompt, conversation and fresh file context are counted. At Kunavo rates, uncached:

Twenty such days on Sonnet 4.6 is about $74 — versus about $186 at Anthropic's list price for the same model. Those are deliberately pessimistic numbers: no caching, one model for everything.

For a reality check against real deployments, Anthropic's own cost documentation reports an average of about $13 per developer per active day and $150–250 per developer per monthacross enterprise deployments, with 90% of users staying under $30 per active day. That is at list price on a mixed model diet, and it is the number to plan a team budget against.

The break-even

Put the two routes on the same axis. At $0.037 per step on Sonnet 4.6:

• $20 (Pro) buys about 540 agentic steps a month — roughly 27 twenty-step tasks.
• $100 (Max, entry tier) buys about 2700 steps — roughly 134 tasks.

So the rule is simple: if you comfortably clear a couple of dozen real tasks a month, every month, the flat fee is good value. If your usage is bursty — some weeks heavy, some weeks nothing — per-token wins, because the quiet weeks are free. Teams almost always want keys regardless, for per-seat attribution and hard spend caps. Run your own numbers in the cost calculator.

Costing a session from its own numbers

Claude Code's /usage prints the four token counts for the current session. That is the honest input to a cost estimate, because it separates cache reads from fresh input:

# What one Claude Code session costs, from its own token counts.
# Run /usage in Claude Code to get the four numbers for a real session.
RATES = {  # USD per 1M tokens on Kunavo: (input, output)
    "claude-haiku-4-5":  (0.40, 2.00),
    "claude-sonnet-4-6": (1.20, 6.00),
    "claude-opus-5":     (2.00, 10.00),
}

def session_cost(model, inp, out, cache_read=0, cache_write=0):
    i, o = RATES[model]
    return (
        inp / 1e6 * i            # fresh input
        + out / 1e6 * o          # output, incl. thinking tokens
        + cache_read / 1e6 * i * 0.10   # cache reads bill at 10% of input
        + cache_write / 1e6 * i * 1.25  # 5-minute cache writes at 125%
    )

# The example session Anthropic's docs print for /usage:
print(session_cost("claude-sonnet-4-6", 1_200, 5_300, 940_000, 50_000))

Note what the split does to the total: in that example the session carries 940,000 cache-read tokens against 1,200 fresh input tokens, and the cache reads bill at a tenth of the input rate.

The four levers that move the number

1. Let prompt caching work. Cached input bills at 10% of the input rate, and Claude Code re-reads the conversation on every step, so this is the largest single effect available (how it works).
2. Clear between unrelated tasks. Context is re-sent every step, so one all-day session is quadratically expensive. A fresh task is a small context again.
3. Match the model to the job. Leaving the largest model as the default is the most common cause of a surprising bill. Sonnet handles most coding work; reserve Opus for architecture and hard multi-step reasoning.
4. Route background calls to Haiku. Claude Code makes its own calls for summaries and titles. ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-haiku-4-5 is one line and it is pure saving.

Running Claude Code on Kunavo

Claude Code reads ANTHROPIC_BASE_URL natively, so it works against any endpoint that serves the Anthropic Messages API — no plugin, no proxy, no patched binary. That is the supported way to keep the same CLI and the same models at a lower per-token rate:

export ANTHROPIC_BASE_URL=https://api.kunavo.com
export ANTHROPIC_AUTH_TOKEN=sk-kn-...
export ANTHROPIC_MODEL=claude-sonnet-4-6
export ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-haiku-4-5

ANTHROPIC_BASE_URL is the origin only — Claude Code appends /v1/messages itself. Use ANTHROPIC_AUTH_TOKEN rather than ANTHROPIC_API_KEY: the two go in different headers, and the latter needs a one-time interactive approval that is never re-prompted once declined. Create the sk-kn- key in the dashboard after signing up and topping up $5 — no monthly fee, and the balance doesn't expire, so a quiet month genuinely costs nothing. Full setup, including what stops working behind a gateway, is in the Claude Code router guide; credential details are in the Claude Code API key guide.

FAQ

Does Claude Code cost extra on top of the model calls?

No. The CLI, the VS Code and JetBrains extensions, hooks, skills, subagents and MCP servers are all free. You pay for inference and nothing else. Even idle, background jobs like conversation summarisation typically stay under $0.04 per session.

Why did my usage climb in a session where I barely typed?

Because every request carries the full conversation, and every tool call is another request carrying that batch of results. A one-line question in a session that has been open all day still pays for the whole history. Cache misses after a long break are the other common cause — the first message back reprocesses everything at full price.

Do agent teams cost more?

Substantially. Each teammate runs its own context window as a separate instance, and Anthropic's documentation puts agent teams at roughly 7× the tokens of a standard session when teammates run in plan mode. Keep teams small and shut teammates down when their work is done.

Is Claude Code free?

Free to install, not free to run — the long version is in is Claude Code free.

How do I install it in the first place?

See install Claude Code for the per-OS commands, first-run login, and the errors people actually hit.

What about Codex CLI — is it cheaper?

Cost is a property of the model and the step count, not the CLI. The comparison that matters is in Claude Code vs Codex CLI.