Claude's API is pay-as-you-go, billed per million tokens with no subscription. Claude API pricing starts at $1/$5 per MTok for Haiku 4.5; the flagship Claude Fable 5 costs $10 input and $50 output per MTok. This guide lists what every model and operation costs.
The Claude API is pay-as-you-go, billed per million tokens (MTok) with no subscription fee and no minimum spend. The current flagship, Claude Fable 5, costs $10 input and $50 output per MTok. Claude Opus 4.8 is $5/$25 per MTok and is the recommended production model for most teams. The most affordable model, Claude Haiku 4.5, starts at $1 input and $5 output per MTok. The Batch API cuts all token costs 50%, and prompt caching cuts repeated input to as low as 10% of the standard rate.
Claude API billing is usage-based and metered per million tokens. Every call generates an input token count (your prompt, conversation history, and any tool schemas) and an output token count (the model's response), and you pay for both at separate rates. There is no monthly fee, no per-request charge, and no seat licence. Billing is postpaid on monthly invoices with no minimum spend requirement.
Prompt caching writes cost a small premium over standard input rates but pay back on the second cache read, which costs just 10% of the standard input rate. The Batch API processes requests asynchronously within 24 hours at 50% off all token costs, making it the single largest cost lever for non-real-time workloads.
Yes. The Claude API supports spend limits scoped to individual users, seat tiers, RBAC groups, or the whole organisation, with daily, weekly, or monthly periods. You can also query your organisation's rate limits programmatically via GET /v1/organizations/rate_limits to see current RPM and TPM limits per model group, and retrieve specific spend limits via GET /v1/organizations/spend_limits/{id}. Both endpoints require an OAuth token. View and manage limits in the Limits section of Claude Console.
Claude.ai subscriptions (Free, Pro, Team, Enterprise) are billed per seat per month and give access to the Claude.ai interface. The Claude API is for developers building applications and is billed per token used, independently of any Claude.ai plan. You can hold both simultaneously. API billing is postpaid monthly with no minimum spend.
Claude API pricing for the current model lineup spans four tiers: Fable 5 (next-generation intelligence for long-running AI agents), Opus 4.8 (complex agentic coding and enterprise work), Sonnet 5 (high performance for coding and agents at introductory pricing), and Haiku 4.5 (fastest, most cost-efficient). All four carry a full 1M-token context window at standard pricing with no long-context surcharge. The table below shows both 5-minute and 1-hour prompt cache write rates; cache reads cost the same regardless of which write TTL was used.
| Model | Input / MTok | Cache Write 5m / MTok | Cache Write 1h / MTok | Cache Read / MTok | Output / MTok | Context | Max Output |
| Claude Fable 5 | $10.00 | $12.50 | $20.00 | $1.00 | $50.00 | 1M | 128K |
| Claude Opus 4.8 | $5.00 | $6.25 | $10.00 | $0.50 | $25.00 | 1M | 128K |
| Claude Sonnet 5* | $2.00* | $2.50* | $4.00* | $0.20* | $10.00* | 1M | 128K |
| Claude Haiku 4.5 | $1.00 | $1.25 | $2.00 | $0.10 | $5.00 | 200K | 64K |
*Claude Sonnet 5 introductory pricing of $2/$10 per MTok is in effect through August 31, 2026. Standard pricing of $3/$15 per MTok takes effect September 1, 2026. Cache write and read rates adjust proportionally.
Claude Opus 4.7 and later Opus models, Fable 5, Mythos 5, and Sonnet 5 use a newer tokenizer that produces approximately 30% more tokens for the same text compared to Sonnet 4.6 and earlier models. If you are migrating from an older model, your token counts and costs will increase even if the per-token rate is lower. Account for this when budgeting a migration.
Claude Mythos 5 shares Fable 5's capabilities and pricing ($10/$50 per MTok) but is available only to approved customers through Project Glasswing and is not part of general API access.
Claude's legacy models remain available for teams with existing integrations. Sonnet 4.6 and Opus 4.7 are the most capable legacy options at $3/$15 and $5/$25 per MTok. Opus 4.1 is deprecated at $15/$75 per MTok. Note that Sonnet 4.6 and earlier use the previous tokenizer, so migrating to Sonnet 5 or Opus 4.7+ will increase your token count by approximately 30% for the same text.
| Model | Input / MTok | Cache Write 5m / MTok | Cache Read / MTok | Output / MTok | Tokenizer | Status |
| Claude Sonnet 4.6 | $3.00 | $3.75 | $0.30 | $15.00 | Previous | Legacy |
| Claude Opus 4.7 | $5.00 | $6.25 | $0.50 | $25.00 | New | Legacy |
| Claude Opus 4.6 | $5.00 | $6.25 | $0.50 | $25.00 | Previous | Legacy |
| Claude Sonnet 4.5 | $3.00 | $3.75 | $0.30 | $15.00 | Previous | Legacy |
| Claude Opus 4.5 | $5.00 | $6.25 | $0.50 | $25.00 | Previous | Legacy |
| Claude Opus 4.1 | $15.00 | $18.75 | $1.50 | $75.00 | Previous | Deprecated (retires Aug 5, 2026) |
Three mechanisms change what you pay per token. The Batch API processes requests asynchronously within 24 hours and cuts all input and output token costs 50% across every model. Prompt caching stores repeated prompt prefixes so subsequent requests pay just 10% of the standard input rate on cache reads. The 5-minute TTL suits most conversational workloads; the 1-hour TTL suits long documents and codebases at a higher write premium. Fast mode, in research preview for Opus 4.8, delivers up to 2.5x faster output at 2x standard pricing and is not available with the Batch API.
| Mechanism | Discount / Premium | Latency | Best for |
| Batch API | 50% off all tokens | Within 24 hours | Bulk extraction, classification, content generation, evals |
| Prompt caching (5-min TTL) | Cache read = 10% of input rate; write = 1.25x input rate | Same as standard | Repeated system prompts, shared context, RAG pipelines |
| Prompt caching (1-hour TTL) | Cache read = 10% of input rate; write = 2x input rate | Same as standard | Long documents, large codebases, extended sessions |
| Fast mode (Opus 4.8) | 2x standard pricing ($10/$50 per MTok) | Up to 2.5x faster | Latency-sensitive production, interactive agentic tasks |
| Data residency (US) | 1.1x on all token rates | Same as standard | Regulatory or compliance requirements |
Yes. Batch API and prompt caching discounts stack. For Claude Opus 4.8, a cached batch read costs $0.25 per MTok versus $5.00 standard input: a 95% reduction. Place stable content (system prompts, shared documents) at the start of every request and route those requests through the Batch API for maximum savings on high-volume workloads.
Tool use requests are priced on the total input tokens sent to the model, including the tool definitions in the tools parameter, plus the output tokens generated. Claude also automatically adds a tool use system prompt when tools are present, which adds a fixed number of input tokens per request depending on the model and tool choice. Understanding this overhead is important for budgeting AI agent and AI coding applications where tool calls are frequent.
The table below shows the system prompt token overhead added per request when tools are present. These tokens are billed as input tokens at the model's standard rate and are in addition to your own prompt and any tool result content.
| Model | tool_choice: auto or none | tool_choice: any or forced |
| Claude Opus 4.8 | 290 tokens | 410 tokens |
| Claude Fable 5 / Sonnet 5 | 354 tokens | 474 tokens |
| Claude Opus 4.7 | 675 tokens | 804 tokens |
| Claude Opus 4.6 / Sonnet 4.6 / Haiku 4.5 | 497 tokens | 589 tokens |
Beyond the system prompt overhead above, tool use also adds tokens from tool names, descriptions, and schemas in the tools parameter, plus tool_use and tool_result content blocks in the conversation. Additional fixed overhead per tool: the bash tool adds 325 tokens per request for Opus 4.8, or 244 tokens for Opus 4.6, Sonnet 4.6, and earlier; the text editor tool adds 700 tokens per request; computer use adds 466–499 system prompt tokens plus 735 input tokens per tool definition for Claude 4.x models. The minimum cacheable prompt length is 512 tokens on Claude Fable 5 and 1,024 tokens on all other models.
Three platform features carry additional charges on top of token costs: Managed Agents, web search, and code execution. Managed Agents add $0.08 per session-hour for active runtime, billed to the millisecond; idle time is not charged. Web search costs $10 per 1K searches; input and output tokens required to process the results are billed at the model's standard rates separately. Code execution gives every organisation 50 free hours per day; additional usage is $0.05 per hour per container, with a 5-minute session minimum. Web fetch has no additional charge beyond standard token costs.
| Feature | Rate | Notes |
| Managed Agents | $0.08 per session-hour (active runtime only) | Billed to the millisecond. Idle time not charged. Standard token rates also apply. |
| Web search | $10.00 per 1K searches | Input and output tokens for results billed at model rates. Errors not billed. |
| Code execution | 50 free hours/day per org; $0.05/hour per container after that | Free when used alongside web search or web fetch. 5-min session minimum otherwise. |
| Web fetch | Free | Only standard token costs for fetched content apply. Use max_content_tokens to limit token consumption. |
| Computer use | Standard token rates only | Adds 466-499 system prompt tokens + 735 input tokens per tool definition for Claude 4.x. Screenshots billed as vision input tokens. |
The same 1,000 chat turns cost $17.50 on Claude Opus 4.8 and $3.55 on Claude Haiku 4.5, an 80% difference that makes model selection the single biggest cost lever available before even using the Batch API. Apply the Batch API and those figures halve again. The table below shows worked examples across models and operations so you can estimate your actual bill before you build.
| Scenario | Math | Cost |
| Opus 4.8 chat turn (1,500 in / 400 out) | (1,500 x $5 + 400 x $25) / 1M | ~$0.0175 |
| 1,000 such Opus 4.8 turns | 1,000 x $0.0175 | ~$17.50 |
| Same 1,000 turns via Batch API | $17.50 x 0.50 | ~$8.75 |
| Haiku 4.5 chat turn (1,500 in / 400 out) | (1,500 x $1 + 400 x $5) / 1M | ~$0.00355 |
| 1,000 Haiku 4.5 turns | 1,000 x $0.00355 | ~$3.55 |
| Sonnet 5 document summary (100K in / 1K out) | (100K x $2 + 1K x $10) / 1M | ~$0.21 |
| Same summary via Batch API | $0.21 x 0.50 | ~$0.105 |
| 1-hour Managed Agent session, Opus 4.8 (50K in / 15K out) | (50K x $5 + 15K x $25) / 1M + $0.08 | ~$0.705 |
| 10,000 support tickets, Haiku 4.5 (~3,700 tokens/ticket) | ~37M tokens x blended $1/$5 rate | ~$37.00 |
| 1,000 web searches | 1 x $10.00 | $10.00 |
Use these exact model ID strings in your API requests. IDs with a date suffix are pinned snapshots; dateless IDs are also pinned to a specific release, not evergreen pointers.
| Model | Claude API ID | Description |
| Claude Fable 5 | claude-fable-5 |
Next-generation intelligence for long-running agents |
| Claude Opus 4.8 | claude-opus-4-8 |
Complex agentic coding and enterprise work |
| Claude Sonnet 5 | claude-sonnet-5 |
Best combination of speed and intelligence |
| Claude Haiku 4.5 | claude-haiku-4-5-20251001 |
Fastest, most cost-efficient model |
Making your first call takes three steps. New users receive a small amount of free credits to test the API.
input_tokens, output_tokens, cache_creation_input_tokens, and cache_read_input_tokens. Log these per request to track actual cost.Start with Claude Haiku 4.5 for testing and simple tasks at $1/$5 per MTok, the most affordable entry point. Move to Claude Sonnet 5 for production workloads at $2/$10 per MTok (introductory through August 31, 2026). Upgrade to Claude Opus 4.8 at $5/$25 per MTok for complex, long-running agentic tasks. If you need the highest capability available, Claude Fable 5 at $10/$50 per MTok is designed for next-generation long-running agents.
Prompt caching stores repeated prompt prefixes (minimum 512 tokens for Claude Fable 5; 1,024 tokens for all other models) so subsequent requests retrieve them at a cache read rate of 10% of the standard input price. For Claude Opus 4.8, a cache read costs $0.50 per MTok versus $5.00 standard input: a 90% saving. Cache writes cost 1.25x the input rate for the 5-minute TTL or 2x for the 1-hour TTL.
The Batch API processes requests asynchronously within 24 hours at a flat 50% discount on all input and output tokens across every model. Use it for document summarisation, data extraction, content generation at scale, and model evaluations. Batch API and prompt caching discounts stack.
Claude Sonnet 5 is on introductory pricing of $2 input and $10 output per MTok through August 31, 2026. Standard pricing of $3 input and $15 output per MTok takes effect from September 1, 2026. Cache write and read rates adjust proportionally.
Claude Sonnet 5 and Opus 4.7 and later use a newer tokenizer that produces approximately 30% more tokens for the same text than Sonnet 4.6 and earlier models. Your actual token consumption and bill will increase by around 30% when migrating, even if the per-token rate is similar or lower. Run token count tests on your typical prompts before committing to a model upgrade.
New users receive a small amount of free API credits to test the API. There is no ongoing free tier for the Claude API. For extended trials or enterprise evaluation, contact Anthropic sales.
Claude.ai subscriptions are billed per seat per month and give access to the Claude.ai interface. The Claude API is billed per token consumed by your application, independently of any Claude.ai plan. You can hold both simultaneously. API billing is postpaid monthly with no minimum spend.
Specifying US-only inference through the inference_geo parameter applies a 1.1x multiplier to all token pricing categories (input, output, cache writes, and cache reads) for Claude Opus 4.6, Sonnet 4.6, and later models. Global routing (the default) uses standard pricing. Earlier models do not support the inference_geo parameter.
Yes. Claude models are available on Amazon Bedrock and Google Cloud Vertex AI at those platforms' own rates. Claude Platform on AWS and Claude in Microsoft Foundry use a Claude Consumption Unit (CCU) billing model where 100 CCUs represent $1.00 of usage at standard per-model rates. Pricing on these platforms may differ from direct Anthropic API pricing.
Yes. Claude Fable 5 includes safety classifiers that can decline certain requests. When a request is refused before any output is generated, you are not billed for the input tokens. The API returns stop_reason: "refusal" as a successful HTTP 200 response (not an error), with a stop_details.category field indicating which classifier fired. If the classifier fires mid-stream, the input tokens and already-streamed output are billed. Claude Mythos 5 does not include these classifiers. If you use Claude Fable 5, plan to handle the refusal stop reason in your application and consider using the server-side fallback parameter to automatically retry on another model.
The Claude API suits teams building AI agents, AI chatbots, AI writing tools, and AI coding assistants that need a capable, safety-focused model with a 1M-token context window across the entire current lineup. Haiku 4.5 is the right starting point for cost-sensitive or high-volume tasks; Sonnet 5 offers the best current value at introductory pricing; Opus 4.8 handles the hardest agentic and enterprise workloads; and Fable 5 is designed for next-generation long-running agents.
In short: pick the cheapest model that meets your quality bar, apply prompt caching for any repeated system prompt or document, and move non-real-time workloads to the Batch API before launch. If you are migrating from Sonnet 4.6 or earlier, account for the 30% token count increase from the new tokenizer. Claude API pricing changes, so verify current rates on Claude's website before committing a production budget.
Still evaluating which model to build on? Browse verified user reviews for Claude and other large language model providers on G2 to compare real-world performance before committing to a production stack.
The pricing and product details above are sourced from Anthropic's official pricing documentation and were accurate as of publication. Rates are subject to change. Please verify current pricing on Anthropic's website before making a purchase.
Krithika Sathyamoorthy is an SEO Content Specialist at G2. She tests and evaluates different types of software solutions, bringing experience across content, marketing, and analysis, with a growing focus on strategy and first-hand product insights that help buyers make better software decisions. Outside of work, you'll usually find her planning her next trip or hunting for good coffee.
OpenAI's API is pay-as-you-go, billed per million tokens with no subscription. OpenAI API...
by Krithika Sathyamoorthy
xAI's Grok API is pay-as-you-go and billed per million tokens, with Grok 4.3 at $1.25 input...
by Sudipto Paul
HeyGen's API is prepaid and pay-as-you-go, billed per second of output starting at $5 with no...
by Sudipto Paul
OpenAI's API is pay-as-you-go, billed per million tokens with no subscription. OpenAI API...
by Krithika Sathyamoorthy
xAI's Grok API is pay-as-you-go and billed per million tokens, with Grok 4.3 at $1.25 input...
by Sudipto Paul