Anthropic pricing: API cost per model, user, and month

Anthropic lists 14 priced models in the ByteCosts index, with Claude Haiku 4.5 at about $2.20 per million tokens on ByteCosts' 70% input / 30% output blend. For most AI apps the bill is driven by output tokens, retry rate, and prompt-cache hit rate far more than the headline input price, so Anthropic is cost-effective when your workload is input-heavy (RAG, classification) or can cache a large shared prefix. Compare Anthropic against alternatives on a real workload - seats, requests, and token mix - before committing to a model or a subscription, because a cheaper per-token price can still lose once power users and long outputs are priced in.

Anthropic pricing at a glance

• Cheapest tracked chat model: Claude Haiku 4.5 at $2.20 per 1M tokens (blended)
• Flagship: Claude Fable 5 at $22.00 per 1M tokens (blended)
• Last recorded price event: Claude Fable 5 batch input on Jul 9, 2026
• 14 priced models in the index, updated July 9, 2026

Anthropic model prices

Per-million-token list prices for current Anthropic models, from the ByteCosts pricing index. Output tokens usually cost several times more than input.

Anthropic subscription plans

Flat-rate Anthropic plans billed per month, separate from the per-token API pricing above. List prices; annual or quarterly billing can lower the effective rate. Verify the current price with Anthropic before subscribing.

Recent Anthropic price events

Source-backed price events recorded for Anthropic in the ByteCosts ledger, newest first. Every event below is a first-time listing - the ledger started tracking that rate; the model itself may be older.

When Anthropic is cost-effective

Token price alone does not decide cost. The variables that actually move an AI bill are: how many output tokens each call emits (output is the expensive side), how often calls retry, what fraction of the input prefix can be served from prompt cache, and how heavily your top 1% of users use the product.

Anthropic tends to win when your workload is input-heavy with short outputs, when you can cache a large shared system prompt or document context, or when a smaller Anthropic model clears your quality bar. It tends to lose when outputs are long and uncached, where a cheaper-per-output model compounds the saving across millions of calls.

Hidden costs to watch with Anthropic

List pricing rarely equals your invoice. Budget for these before you commit:

• Output-token blow-up: reasoning and verbose responses multiply the most expensive token class.
• Retries and tool loops: a single failed agent step can re-send the whole context.
• Cache misses: prompt caching only saves money above a break-even reuse count; cold prefixes pay the write premium.
• Power users: a small share of heavy users can dominate spend and erase a plan's margin.
• Egress and gateway fees: LLM gateways and observability add a percentage on top of raw model cost.

Limitations before production billing decisions

Treat ByteCosts calculations as planning estimates, not final billing totals. Real invoices can differ because token mix, retry rate, cache hit rate, rate limits, taxes, gateway fees, regional pricing, and negotiated discounts change the effective cost.

Verify the provider source before production billing decisions, then compare the estimate with your own logs or invoice once production traffic is live.

Lower-blend alternatives to Anthropic

On ByteCosts' 70% input / 30% output chat blend, these providers screen lower than Anthropic. Treat this as a shortlist, not a decision - check the selected model against your real token mix:

• Mistral - $0.040 blended via Ministral 3B (latest), 28 priced models.
• Cohere - $0.071 blended via Command R7B, 8 priced models.
• Alibaba Qwen - $0.095 blended via Qwen Turbo, 43 priced models.

Frequently asked questions

Anthropic pricing. ByteCosts. Updated July 9, 2026. https://bytecosts.com/pricing/anthropic/

Sources

• Anthropic official pricing
• Anthropic website

Machine-readable

• Markdown
• JSON reference
• CSV reference