Claude API Pricing Guide 2026

1. What does Anthropic charge for the Claude API?

As of mid-2026, Anthropic publishes three model tiers — Opus, Sonnet, and Haiku — each at three latest-generation versions. Prices are per 1 million tokens, billed separately for input (your prompt + context) and output (the model's response):

Three things most pricing guides skip:

• Output is 5× input across every tier. A chatty agent that produces long responses pays 5-10× what a query-and-extract agent pays for the same input.
• Prompt cache write costs more, not less. Writing to cache = 1.25× input price. Reading is where the savings live (0.1× input).
• Thinking variants are gone. Anthropic deprecated the "-thinking-xhigh" SKUs in 2026. All current models reason in the standard variant; budget-cap behavior is configured per-request via tools.

See the full model catalog for spec sheets and per-model code samples. Direct links to each: Claude Opus 4.7, Claude Sonnet 4.6, Claude Haiku 4.5.

2. Why does the Claude API bill spike at scale?

The most-shared developer post pattern of 2026 follows a script: someone runs Claude Code or an OpenClaw agent for a few hours, checks the bill, and finds $80-$200 in API spend in a single day. The Anthropic $200 Max subscription gives them the same usage for $200/month — a 20-30× difference.

The math behind the cliff has three causes:

1. Context bloat.Modern coding agents stuff large portions of the codebase into the prompt — easily 30-60K tokens of input per request at $5/1M means $0.15-$0.30 per request before any output is generated. At 1,000 requests/day that's $150-$300 of input alone.
2. Output amplification.Output costs 5× input. An agent that returns a diff or a multi-step plan generates 3-10K output tokens per request. On Opus that's another $0.075-$0.25 per request, on top of the input cost.
3. Retry loops. When the model gets confused, self-recovery prompts retry. Each retry duplicates the input cost. A misconfigured agent has been documented draining $30 in 8 minutes via a single tight retry loop.

Worth reading on this specifically: Cut your $24K Claude bill to $5K walks through the actual audit a developer published when their monthly Anthropic bill hit $24,096.47 on a $200 plan.

3. What does each Claude model cost per call?

Abstract per-token rates are hard to reason about. Here's what three concrete workload shapes cost at official Anthropic rates:

The lesson: model choice dominates the bill. A workload that needs Sonnet but runs on Opus pays roughly 2× more for the same output quality on most tasks. Haiku-suitable tasks running on Sonnet pay roughly 3× more. The first SEO move on cost is always "am I using the right model for this query?" — before you change vendor.

Deep dive on this: Claude API unit economics for SaaS breaks down voice, chatbot, and code-gen unit economics with the math behind each.

4. How much does prompt caching actually save?

Anthropic's prompt caching changes the math more than any other single mechanic. Understanding it is the second-biggest cost lever after model selection.

Worked example. You're running an agent that has a 20K-token system prompt + tool definitions, and you query it 50 times in an hour:

Three caveats most teams miss:

• Cache hits are not guaranteed. The model has to see an identical prefix. Even one differing character at the start invalidates the cache.
• The 5-min cache resets faster than you think on bursty workloads — agents that pause for tool calls regularly miss the window.
• Some proxies (including legitimate ones) don't forward cache metadata 100% of the time. Verify by checking response headers and benchmarking with cache off vs on.

5. Why do non-US developers pay 25-50% more for Claude?

The list price is the same globally. What you actually pay is not. Developers outside the US and EU face two surcharges that compound on every charge:

The structural payment-rail problem makes it worse. In countries where prepaid cards, Maestro, V-Pay, and UPI virtual cards are common, Stripe's rules silently refuse the transaction — and the error message says "Payment failed" regardless of the actual cause. Retrying with different cards doesn't help.

Region-specific deep dives: Claude API for Indian developers, Pay with crypto (USDT/BTC/ETH), and Claude API without a credit card.

6. Subscription, API, or proxy — which one wins?

There's no single "best" way to pay for Claude. It depends on what you're doing:

Comparison breakdown: full comparison set covers OpenRouter, Anthropic direct, CometAPI, and others head-to-head.

7. How does a Claude proxy actually save 70-80%?

Healthy skepticism is warranted here. A legitimate volume-discount proxy is not magic — it's pooled demand. Three things have to be true for the math to work:

1. Bulk credit purchase. The proxy buys upstream API credit in large blocks and resells it at retail-minus-X. The discount is real because the upstream gives volume customers genuine discount tiers — the proxy is aggregating small-customer demand into volume-tier demand.
2. Crypto-only payment removes Stripe overhead. Processing fees + chargeback risk + fraud reserves run to 3-5% of revenue for card-based SaaS. Crypto-only payment eliminates those, and that margin becomes additional customer-side discount.
3. Lean ops.No sales team, no enterprise tier, no customer-success org. A proxy that runs on $19 lifetime Pro doesn't have the burn rate that needs Acquisition, Marketing, and Sales overhead — so the per-customer overhead is structurally lower.

What to verify before signing up with any proxy:

• Non-quantized model guarantee. Ask in writing. Some open-router-style services quantize. A real proxy returns near-identical output to direct upstream — benchmark before committing.
• Per-token pricing posted publicly per model. No "contact sales for enterprise pricing". If you can't model the savings before signup, the savings claim is unverifiable.
• Balance doesn't expire. Some competitors started enforcing 12-month expiry on unused credits in 2025. That's a cash grab. Confirm balance permanence in writing.
• Public uptime page. Real numbers from a third-party monitor (Instatus, StatusPage, BetterUptime), not a self-reported marketing line.
• Per-request balance audit log. When a retry loop spikes the bill, you need to find the leaking agent in 30 seconds, not 3 days.

ClaudeAPI.cheap publishes all five — see the Anthropic API alternative page for the full feature comparison.

8. How hard is migrating to a Claude proxy?

Drop-in for the official Anthropic Python and Node SDKs. The entire migration is two env vars:

# Anthropic SDK — only the base URL changes
export ANTHROPIC_API_KEY="sk-cc-your-key-here"
export ANTHROPIC_BASE_URL="https://claudeapi.cheap/api/proxy"

# Your existing code stays the same:
# client = Anthropic()
# client.messages.create(model="claude-sonnet-4-6", ...)

# Roll back: unset ANTHROPIC_BASE_URL.
# That's it. Production is back on direct Anthropic.

Works the same way for OpenAI-format clients (point at /v1/chat/completions) and every Claude-compatible tool: Claude Code, Cline, Aider, OpenClaw, Cursor, OpenCode. See per-tool setup at /tools.

For step-by-step screenshots and a checklist: The Anthropic base-URL trick.

What does ClaudeAPI.cheap charge per Claude model?

Pro = $19 lifetime (one-time, crypto), 80% off list. Basic plan is free forever at 50% off — see full plan comparison.

9. FAQ