Claude AI costs, the four ways you actually pay

Three published surfaces and one stacking trick. Free trial credits on a new API account, $17 to $20 per month for Pro, $100 or $200 per month for Max, and a per-token API from $1 input ($5 output) on Haiku 4.5 to $5 input ($25 output) on Opus 4.7 per million tokens. The trick is that the Pro and Max plans expose an OAuth scope (user:inference) that external builders can consume from, so a chat turn that would otherwise be metered against an API key can be billed against your flat monthly subscription instead.

The official table is at https://platform.claude.com/docs/en/about-claude/pricing. Verified on April 29, 2026: Opus 4.7, 4.6, and 4.5 are $5 input and $25 output per million tokens; Opus 4.1 and 4 are $15 / $75; Sonnet 4.6, 4.5, and 4 are $3 / $15; Haiku 4.5 is $1 / $5; Haiku 3.5 is $0.80 / $4. Cache writes are 1.25x base for 5-minute cache or 2x base for 1-hour cache. Cache reads are 0.1x base. Batch API is a 50% discount on input and output. The Pro and Max plans live at https://claude.com/pricing.

It does not save you money in absolute terms. It moves you from a metered-per-token cost model to a flat monthly cost model. If you spend a heavy day prompting Sonnet through a builder, the metered API would charge you proportional to the tokens. The OAuth path consumes from the Pro or Max usage allowance you already pay $20 or $200 a month for. For users who already pay Anthropic, that means an app build session adds zero new line items. The constraint is the rate-limit window on the plan, not a credit balance.

src/app/api/chat/route.ts line 187 reads `const apiKey = oauthToken ? "" : sharedApiKey ?? "";`. If the user has a connected Claude OAuth session, oauthToken is the JSON blob for the user's credentials and apiKey is set to the empty string. The ACP subprocess that drives Claude inside the E2B sandbox is spawned with that empty string in ANTHROPIC_API_KEY and the OAuth blob in CLAUDE_CODE_OAUTH_TOKEN, plus the anthropic-beta: oauth-2025-04-20 header. Inference for that turn is billed against the user's plan, not against mk0r's API key.

Anonymous use is free. The default model on a fresh session is pinned to Haiku at src/app/api/chat/model/route.ts line 5 (FREE_MODEL = 'haiku'); inference goes through mk0r's shared ANTHROPIC_API_KEY and you pay $0. The mk0r Pro tier is $19 per month with a 21-day free trial (src/app/checkout/page.tsx, captured as 'price: 19, trial_days: 21' in the PostHog checkout_started event). That tier unlocks Sonnet and Opus inside the model picker. Separately, you can connect your own Claude.com Pro or Max plan via OAuth, and mk0r will route inference through your subscription instead. The two settings stack: an mk0r Pro user with a connected Claude plan pays $19 to mk0r and consumes Claude tokens from their existing Anthropic plan.

src/components/header.tsx defines MODEL_LABELS as: default → 'Fast', haiku → 'Scary', sonnet[1m] → 'Fast+', opus[1m] → 'Smart'. The 'haiku' label is intentionally dramatic because it is the only model an anonymous or free-tier user can run, and the landing copy advertises sub-30-second builds against it. The 1m suffix on Sonnet and Opus selects the 1M-token long-context variant. Long context is included at the standard per-token rate on Claude Sonnet 4.6, Opus 4.6, and Opus 4.7, so the 'Smart' label does not carry a context premium.

It changes it a lot. Cache writes cost 1.25x base for the 5-minute cache or 2x base for the 1-hour cache. Cache reads cost 0.1x base. For a builder that re-sends a large system prompt and a long conversation history on every chat turn, the reusable bytes hit the cache-read rate after the first write. On Sonnet 4.6 that is $0.30 per million input tokens for cached content versus $3.00 for fresh input, a 10x reduction. The Anthropic worked example in their pricing doc shows a one-hour Opus 4.7 session dropping from $0.705 to $0.525 when 40,000 of 50,000 input tokens are cache reads.

Batch is a 50% discount on input and output for asynchronous workloads. It is not relevant to a real-time builder because the builder needs the response to stream back into the iframe immediately. Fast mode is the opposite trade: 6x standard rates on Opus 4.6 ($30 input, $150 output per million) for higher throughput, currently in beta. Builders that target a chat-driven preview loop (mk0r included) do not enable either, because batch breaks streaming and Fast mode burns cost faster than the user gains by waiting one fewer second.

A few. Web search is $10 per 1,000 searches on top of token costs. Code execution is free if you bundle web_search or web_fetch in the same request, otherwise $0.05 per container hour after a free monthly allowance of 1,550 hours. Tool use adds 245 to 700 tokens per tool definition. Computer use adds 466 to 499 tokens to the system prompt and 735 tokens per tool definition on Claude 4.x. Data residency on US-only inference adds a 1.1x multiplier on Opus 4.6, 4.7, and newer. None of these apply to the OAuth path because they bill against the API account directly, not the consumer plan.

Because the friction of putting a credit card on Anthropic, generating an API key, and remembering to rotate it kills the maker loop on a Saturday afternoon. The shared key is not free for mk0r, it is metered against our Anthropic bill at the rates above. We absorb that cost on Haiku because the unit economics work out on the conversion rate to mk0r Pro and to the connected-Claude-plan path. Sonnet and Opus on the shared key would not, which is why anonymous and free-tier accounts are pinned to Haiku in src/app/api/chat/model/route.ts.