Claude API pricing (May 2026)

Every Claude model and its per-token price

All prices in USD per 1 million tokens. Last reviewed 2026-05-15. Provider pricing pages are authoritative, confirm before contracting.

How Claude pricing actually works

Claude pricing in 2026 prices the flagship aggressively against OpenAI. Opus 4.7 at $5 / $25 per 1M is roughly 1/6 the cost of GPT-5.5 Pro while leading the Arena coding leaderboard at 1567 Elo. Sonnet 4 at $3 / $15 with a 1M-token context window is the production default for most teams. Haiku 3.5 at $0.80 / $4 is the cheap tier for classification, routing, and high-volume extraction.

Claude is the default LLM for serious coding work. Claude Code authors approximately 4% of all public GitHub commits, by a wide margin the largest share of any AI coding tool. Cursor, the most popular AI IDE, uses Claude Opus 4.7 as its default in agent mode. Most Fortune 500 enterprises run Claude via AWS Bedrock or Google Vertex AI for VPC residency and procurement integration.

Anthropic offers 90% prompt cache discount, 50% batch discount, and ZDR by default for paid API. Bedrock and Vertex expose Claude with VPC residency.

Prompt caching: the 90% discount most teams ignore

Anthropic’s prompt cache is the deepest discount in the market: 90% off the cached input rate. You opt in by attaching cache_control on the message blocks you want cached (typically the system prompt, tool definitions, and any large stable prefix like a codebase or RAG corpus). The 5-minute TTL covers most agent loops; the 1-hour TTL (at 25% of input rate) covers longer-running back-office workloads. Cache hit rate above 80% is achievable on most production workloads with deliberate prompt structure.

On a typical coding agent run that re-sends the same 200K-token codebase across 10 turns, prompt caching reduces effective input cost by 80-90%. The cached-input column in the table above is the right number to plug into a production budget; the headline input rate is the "new conversation" rate, not the steady-state rate.

Batch inference, and half-price overnight

Anthropic’s Message Batches API offers 50% off both input and output for batches of up to 100,000 requests, with a 24-hour completion SLA. Combined with prompt caching, a cache-warm batched call can land under 10% of the headline price. The right workloads are nightly evaluation suites, document enrichment, content classification, and synthetic training data generation.

Batch + cache stack. The combined effective rate for a cache-warm, batched call is often 5-10% of the headline price. For workloads like nightly eval suites, large-scale classification, document enrichment, and synthetic data generation, batching is free money.

Four real production cost scenarios

The routing pattern that cuts Claude spend 60-80%

The dominant production pattern is: default to Sonnet 4; promote to Opus 4.7 on a small set of triggers (complexity score, multi-step planning intent, prompt size above 100K, more than 5 tool calls expected); fall back to Haiku for any pure-classification turn. Most fleets run 70% Sonnet / 20% Haiku / 10% Opus by request count, which lands at a similar split by spend because the cheaper tiers dominate by volume.

A typical production fleet settles into a 70/25/5 split. 70% of requests handled by the smallest competent tier, 25% by the mid-tier workhorse, 5% promoted to the flagship. Done well, this cuts model spend 60-80% versus naive single-model use without any measurable quality drop on the bulk of requests.

With an AI gateway in front, the routing rule is one config block: declare a default model, declare promotion triggers, declare a fallback to a second provider for availability. Applications keep using a single OpenAI-compatible endpoint. See Swfte for a managed runtime that bundles the gateway, observability, eval, and per-team cost ceilings.

Enterprise considerations

Claude reaches enterprise via three channels. Anthropic direct API has zero data retention by default for paid contracts. AWS Bedrock exposes Claude with regional residency, IAM-based access, and integration with AWS data services. Google Vertex AI does the same on GCP with Vertex Pipelines / Workbench integration. Most regulated industries (finance, healthcare, government) standardise on Bedrock or Vertex for procurement reasons; pricing matches direct API at scale.

• Prompt caching: Available — use it from day one; the headline rate is misleading without it.
• Batch inference: Available, 50% discount, up to 24h SLA.
• Fine-tuning: Not currently offered.
• On-prem / VPC: Available via Bedrock / Vertex / Azure or direct VPC contract.
• Zero data retention: Available; default on enterprise contracts.

How Claude compares to the rest of the market

Against OpenAI, Claude wins on coding, agent loops, and per-token price at the frontier; Opus 4.7 is roughly 1/6 the price of GPT-5.5 Pro with stronger coding benchmarks. Against Gemini, Claude wins on coding and long-form writing; Gemini wins on raw price and multimodal. Against DeepSeek, Claude has the quality edge on the hardest benchmarks but DeepSeek V4 Pro is ~3x cheaper at similar Arena Elo on bulk workloads. The standard combination is Claude for code/agents, Gemini or DeepSeek for high-volume background work.

For a full side-by-side, see the API pricing index and the AI model leaderboard for quality / speed / value rankings.

Frequently asked questions about Claude API pricing