Anthropic’s current Claude API ladder runs from Haiku 4.5 for utility calls through Sonnet 5 for mainstream production, Opus 5 for premium work, and Fable 5 for the hardest frontier workloads.
The chart, calculator, and table above pull current rates from pricing.json at build time. Use the full Anthropic pricing page for every tracked model and the AI token calculator to enter your own input, cached-input, and output volumes.
Calculate before you migrate: model the same workload across Haiku, Sonnet, Opus, and Fable, then compare cost per accepted result rather than rate per token alone.
Which Claude Model Should You Use?
| Workload | First model to test | Escalate when |
|---|---|---|
| Classification, extraction, and routing | Claude Haiku 4.5 | Errors reach users or create expensive downstream work |
| Customer support, RAG, and document workflows | Claude Sonnet 5 | Complex judgment or long agent loops fail evaluation |
| Production coding and tool-using agents | Claude Sonnet 5 | Difficult repairs or planning need stronger reasoning |
| Premium coding, research, and final review | Claude Opus 5 | Fable materially improves accepted-task cost |
| Frontier evaluations and hardest tasks | Claude Fable 5 | Keep Opus as fallback when the quality gain disappears |
Haiku is the cheapest place to start for narrow jobs with short, verifiable outputs. Sonnet is the practical default when users see the answer or tool reliability matters. Opus belongs on calls where a quality miss costs more than the model premium.
Fable should be an evaluation-backed escalation, not a default selected by model name. Its value depends on reducing retries, human correction, or task failure enough to offset the higher live rate shown above.
How Claude API Billing Works
Claude API usage is metered separately for input, cached input, and output. Long system prompts, retrieved documents, repository context, tool results, and generated reasoning all contribute to the bill in different ways.
Prompt caching is most useful when a large prefix stays stable across calls. Put system instructions, tool schemas, coding rules, or reference documents in a deterministic order; changing timestamps, IDs, or tool order can prevent reuse. Confirm savings from the provider-reported usage fields instead of assuming repeated text qualified.
Output deserves its own budget. Verbose answers, oversized JSON, repeated tool summaries, and retry loops can dominate accepted-task cost. Set output limits, request only needed fields, and trim tool results before passing them back to the model.
Anthropic also documents discounted batch processing for work that does not need an immediate response. Evaluations, nightly enrichment, document backlogs, and offline classification are natural candidates. Verify the current eligibility and billing rules in Anthropic’s official documentation before production.
Hidden Costs and Limits
Retries hide inside the token total. A cheaper model can lose if it needs several attempts, causes extra tool calls, or requires manual repair. Log first-pass success, fallbacks, latency, and human correction alongside token usage.
Large context can be operationally expensive. A large context window makes repository and document workflows possible, but sending everything on every turn is rarely efficient. Retrieve relevant passages, summarize settled history, and cache stable prefixes.
Subscriptions are not API credits. Claude Free, Pro, Max, and Team are human-facing products with their own usage policies. Apps and automations should be budgeted against API usage unless Anthropic explicitly documents another arrangement. Compare current plans in the Claude subscription tracker.
Provider route matters. Direct Anthropic, Amazon Bedrock, and Google Cloud Vertex AI can differ in model availability, regions, billing, and enterprise controls. Confirm the actual route rather than assuming one universal Claude price or feature set.
A Cost-Aware Claude Routing Strategy
- Send simple classification, extraction, and transformation work to Haiku.
- Use Sonnet for most user-visible answers, coding, RAG, and tool use.
- Escalate uncertain or failed requests to Opus.
- Add Fable only where a fixed evaluation set proves a better completed-task result.
- Cache stable context and keep volatile values outside the reusable prefix.
- Batch delay-tolerant work and cap generated output.
- Review cost per accepted task weekly, including retries and human repair.
Run the same production-shaped cases through at least two tiers before choosing a default. Define acceptance first—passing tests, valid schema, grounded citations, or reviewer score—so a quality gain can be compared with its real cost.
For cross-provider routing, compare the OpenAI pricing page, Google AI pricing page, and our Claude vs Gemini pricing guide. The cached-token guide explains how stable prefixes change repeated-context economics.
Open-model fallback: Check Novita’s current model catalog when you want an OpenAI-compatible route for comparing hosted open models against Claude on the same evaluation set.
Affiliate disclosure: AI Pricing Guru may earn a commission from the sponsored Novita link at no extra cost to you. It does not affect the live pricing data or recommendations.
FAQ
What is the cheapest current Claude API model?
Claude Haiku 4.5 is the lowest-cost active Anthropic model in our live tracker. It is the first model to test for routing, classification, extraction, and short structured outputs.
Is Claude Sonnet 5 or Opus 5 better for production?
Start with Sonnet for most production traffic and escalate difficult cases to Opus. Choose Opus as the default only when your evaluation shows fewer failures or less human correction justify its higher live rate.
Does Claude Pro include API usage?
Claude Pro is a user subscription, not a general production API allowance. Budget software, backend automation, and customer-facing features with the API rate table unless Anthropic explicitly states otherwise.
How can I reduce Claude API costs?
Route simple work to Haiku, cache stable context, cap output, batch offline jobs, and reserve premium models for failed or difficult cases. Measure cost per accepted result so retries and manual repair are included.
Sources: Anthropic’s official Claude pricing documentation, prompt-caching guide, batch-processing guide, and AI Pricing Guru’s live pricing API.