Claude API Pricing Explained: Plans, Usage Limits and Hidden Costs in 2026
Claude API pricing is pay-as-you-go per million tokens, starting at $1 input / $5 output per MTok on Haiku 4.5 and ranging up to $10 / $50 on Fable 5.1, with prompt caching (90% off repeated input) and Batch API (50% off) stacking to cut effective spend by up to 95%. Anthropic also sells flat-rate subscriptions (Free, Pro $20/month, Max $100–$200/month, Team seats $25–$125/month) that are billed separately from API usage. If you want to extend Claude-powered agents with tested, secure skills at no cost, SkillzBench offers a free skills directory and free MCP connector while it finishes its pre-launch build.
TL;DR
- Claude API is pay-as-you-go: Haiku 4.5 costs $1 input / $5 output per million tokens, Sonnet 5.5 $2 / $10, Opus 5.5 $4 / $20, and Fable 5.1 $10 / $50.
- Prompt caching cuts repeated input by 90% (cache reads at 0.05x–0.1x base input) and the Batch API takes 50% off all token rates; stacked, they can reduce effective spend by 95%+.
- Claude subscriptions (Free, Pro $20/month, Max $100–$200/month, Team $25–$125/seat/month) are billed separately from API usage — a paid plan does not create a shared pool of API tokens.
- Hidden cost levers include cache writes (1.25x for 5-minute, 2x for 1-hour), Fast Mode (2x–6x standard rates), web search at $10 per 1,000 searches, code execution at $0.05 per container-hour after 50 free hours, and Managed Agents runtime at $0.08 per session-hour.
- Claude 4.6 and later models include the full 1M-token context window at standard pricing — no long-context surcharge (Sonnet 4.5 was the last current-gen model with one).
- Max plans include $100–$200/month in API credits and Team plans up to $500/month in pooled API credits, which can offset API spend.
- SkillzBench is building a free directory of verified agent skills and a free MCP connector so businesses can extend Claude-based agents without adding vendor costs.
Head to head
| Dimension | Claude API | SkillzBench |
|---|---|---|
| Pricing | Pay-as-you-go per token: $0.10/$0.50 (Haiku 5.5, ≤100K prompts) to $10/$50 (Fable 5.1) per MTok; subscriptions $0–$200/mo; Enterprise $20/seat/mo annual + usage | Free — directory, MCP connector, benchmarks and alerts; ad-supported 🏆 |
| Ease of use | Console quickstart with cURL, Python and TypeScript examples; requires developer comfort | Toggle-based loadouts designed for non-technical users 🏆 |
| Features | Four model tiers, 1M context, caching, batch, tool use, Managed Agents, Fast Mode 🏆 | Discover, Verify, Benchmark, Equip pillars with 0–100 trust scores |
| Support | Documentation and quickstart; dedicated support tied to Enterprise 🏆 | Waitlist and newsletter live; product not yet public |
| Integrations | Tool use, server-side web search and code execution, MCP connectors | Free MCP server for Claude Code, Codex, Cursor, Gemini CLI and chat-app agents |
| Setup | Create an API key in the Console and start billing per token immediately 🏆 | Join the waitlist; full setup pending launch |
| Transparency | Token rates published to the cent; subscription usage allowances stated only as unstated-token multiples | Trust scores and Lift published separately; money never touches scores 🏆 |
The contenders
The Claude API is Anthropic's developer surface (the Claude Platform/Console) for building on Claude models, billed per token rather than per seat. As of late 2026 the current lineup spans Claude Haiku 5.5 ($0.10/$0.50 per MTok for prompts up to 100K tokens), Sonnet 5.5 ($2/$10), Opus 5.5 ($4/$20) and Fable 5.1 ($10/$50), with 4.x models still available as legacy options. Models from the 4.6 generation onward include a full 1M-token context window at standard pricing. The API supports prompt caching (90% discount on cache reads), Batch processing (50% off), tool use, server-side web search and code execution, Managed Agents sessions, and Fast Mode on Opus models. Anthropic separately sells consumer/team subscriptions (Free, Pro, Max, Team, Enterprise) that include Claude Code and chat access but do not share a token pool with the API — a distinction that causes most Claude billing confusion. Small businesses typically use the API for programmatic automation and the subscriptions for interactive work.
pricing 0/month
- + Token rates are published to the cent, making API costs calculable in advance
- + Caching plus batch processing can cut effective spend by 95%+ on repetitive workloads
- + 1M context at standard pricing on current models removes long-context surcharges
- + No subscription fee for the API itself — you pay only for tokens and tools used
- + Wide model range lets you route cheap tasks to Haiku and hard tasks to Opus/Fable
- − Multi-turn agents re-send context every turn, so bills compound faster than expected
- − Cache writes cost extra (1.25x–2x input), and Fast Mode runs 2x–6x standard rates
- − Subscription usage allowances are published as multiples of an unstated token base, so plan-vs-API comparisons can't be verified from Anthropic's own pages
- − Setting ANTHROPIC_API_KEY in your shell makes Claude Code bill at API rates, ignoring your subscription
- − US-only inference on 4.6+ models adds 10%
SkillzBench is AIQ Labs' free, independent directory of AI agent skills plus a free MCP tool that connects any agent to the skills you choose. Agent skills — SKILL.md instruction packages for Claude Code, Codex, Cursor, Gemini CLI and chat-app agents — are everywhere, but published research found roughly 26% of them carry at least one vulnerability, and no one tells you which actually help. SkillzBench is building four pillars: Discover (continuously crawling and deduping skills from GitHub), Verify (pinning every version to a commit SHA with validation, license checks, security scans and quality judging), Benchmark (testing every skill in a sandbox with the same model, with and without the skill, and publishing the Lift in percentage points), and Equip (one free MCP server plus per-agent loadouts you toggle, so each agent only sees what you chose). A 0–100 trust score combines security, quality, license, maintenance and community signals, with Lift always shown separately and money never touching scores. Everything is free; revenue comes from blog advertising. The product is pre-launch — the waitlist and newsletter are live.
pricing 0/month
- + Completely free — directory, MCP connector, benchmarks and alerts
- + Directly addresses the ~26% skill-vulnerability finding with SHA pinning and security scans
- + Lift benchmarking shows whether a skill actually helps before you install it
- + Per-agent loadouts limit what each agent can see, reducing risk
- + Independent: ad revenue never influences scores or listings
- − Pre-launch: the product is still being built, with only the waitlist and newsletter live
- − Focuses on agent skills rather than being a general-purpose AI platform
- − Benchmark coverage will take time to build across the skill ecosystem
Verdict
Winner: Claude API
For small businesses building on Claude, the API's pay-as-you-go model with caching and batch discounts is the most controllable way to pay, while subscriptions suit interactive chat and Claude Code use. There is no single cheaper winner — the right choice depends on whether your usage is programmatic (API) or interactive (subscription).
| Option | Overall | Breakdown |
|---|---|---|
| Claude API | 8.5 | : 8: 8: 9: 7: 8: 9 |
| SkillzBench | 7.5 | : 10: 7: 8: 6: 8: 7 |
Claude API token pricing by model
The discounts that change your real bill: caching and batch
Hidden costs: tools, Fast Mode, agents and the subscription/API split
Where SkillzBench fits for Claude-powered businesses
Which should you pick?
- :
- :
- :
- :
- :
FAQ
How much does the Claude API cost per token in 2026?
Current rates per million tokens are $0.10 input / $0.50 output on Haiku 5.5 (for prompts up to 100,000 tokens; $0.50/$2.50 above that), $2/$10 on Sonnet 5.5, $4/$20 on Opus 5.5, and $10/$50 on Fable 5.1. Legacy 4.x models remain available, including Opus 4.8 at $5/$25, Sonnet 4.6 at $3/$15 and Haiku 4.5 at $1/$5. Prompt caching discounts repeated input by 90% on cache reads, and the Batch API takes 50% off all token rates.
Is a Claude subscription the same as API access?
No. Claude Free, Pro, Max, Team and Enterprise are subscription products for chat and Claude Code, billed at flat rates ($0, $20/month, $100–$200/month, $25–$125/seat/month, and $20/seat/month annual for Enterprise self-serve plus usage). API usage is metered separately through a Claude Console developer account and billed per token. A paid subscription does not create a shared pool of standard API tokens, and the two surfaces bill independently even if you pay for both.
What hidden costs should small businesses watch for with Claude?
Beyond base token rates, watch for: cache writes at 1.25x (5-minute) or 2x (1-hour) input; Fast Mode at 2x–6x standard rates depending on the Opus model; server-side web search at $10 per 1,000 searches; code execution at $0.05 per container-hour after 50 free hours per org per day; Managed Agents runtime at $0.08 per session-hour; and a 10% surcharge for US-only inference on 4.6+ models. Also, if ANTHROPIC_API_KEY is set in your shell, Claude Code bills at API rates and ignores your subscription.
Can caching and batch processing really cut Claude API costs by 95%?
Yes, on suitable workloads. Prompt caching discounts cache reads to 0.05x–0.1x base input (a 90% saving on repeated context), and the Batch API halves input and output for asynchronous jobs. Anthropic states the batch discount stacks with cache pricing — for example, a batched Opus 5.5 cache hit costs $0.10/M with batched output at $10/M. Stacked on repetitive, cacheable workloads, that approaches a 95%+ reduction. Fast Mode does not work with Batch, so the discount applies to standard-speed asynchronous processing.
Do Claude models charge extra for long context windows?
Not on current models. Claude 4.6-generation and later models — including Opus 4.8, Opus 4.7, Opus 4.6 and Sonnet 4.6, and the newer 5.x releases — include the full 1M-token context window at standard pricing with no surcharge. Sonnet 4.5 was the last current-generation model with a long-context premium, charging 2x input and 1.5x output above 200K tokens. Haiku 5.5 handles the length question differently, pricing by prompt length ($0.10/$0.50 up to 100K tokens, $0.50/$2.50 above).
How can I safely extend Claude-based agents with skills?
Agent skills — SKILL.md instruction packages for Claude Code, Codex, Cursor, Gemini CLI and chat-app agents — are widely shared, but published research found roughly 26% carry at least one vulnerability. SkillzBench is building a free answer: it crawls and dedupes skills from GitHub, pins every version to a commit SHA, runs validation, license checks and security scans, and benchmarks each skill in a sandbox to publish its Lift in percentage points. A free MCP server with per-agent loadouts lets you control exactly what each agent sees. Everything is free; the platform is pre-launch with a live waitlist.