Claude Sonnet 4.6

Start with the rate for the work you actually send

These are US-dollar token rates for Anthropic's first-party API, not a monthly allowance. MTok means one million tokens. Partner platforms can have different billing terms; identify your actual endpoint before using this table in a budget.

Anthropic: API pricing ↗

A six-cent request can become a $600 workload

For an uncached standard request, multiply input tokens by $3/1,000,000 and billable output tokens by $15/1,000,000, then add them. A hypothetical request with 10,000 input and 2,000 output tokens costs $0.03 + $0.03 = $0.06. Ten thousand identical requests cost $600. These are arithmetic examples, not measured performance or a promise about typical usage.

Output is the expensive side of this example. Keeping the same input but generating 1,000 output tokens costs $0.045, a 25% reduction. A useful experiment is to compare a concise structured answer against a long explanation on the same tasks. Measure whether the shorter response still completes the job; cheaper incomplete answers can create extra calls and human cleanup.

For planning, record cost per accepted result, including retries and failed attempts. A workflow with three model calls is three opportunities to incur token charges. Add separate tool charges where applicable, and avoid treating a single successful demonstration as your monthly average.

Anthropic: API pricing ↗

Caching pays when the same prefix is reused

Sonnet 4.6 needs at least 1,024 tokens in a cacheable prefix. Reuse requires matching content and a valid cache; marking a prompt for caching does not prove a hit. Inspect cache creation and cache read token counts in the response. Put stable instructions and reference material before the question that changes.

Consider a 100,000-token reference prefix used ten times. Without caching, that prefix alone costs $3. With one five-minute write and nine hits, it costs $0.375 + $0.27 = $0.645, or 78.5% less. A one-hour write plus nine reads costs $0.87. These calculations exclude each new question and all generated output, and assume every later request hits the still-valid prefix.

The first write costs more than ordinary input. In this simplified prefix-only comparison, a five-minute cache becomes cheaper on its second use; the one-hour option becomes cheaper on its third. Choose the lifetime around the gap between real requests. A cache that expires before reuse does not deliver the illustrated saving.

Anthropic: prompt caching ↗ · Anthropic: Claude Sonnet 4.6 specifications ↗

Use batch pricing when a person is not waiting

Message Batches process requests asynchronously with a 50% input/output discount. The earlier 10,000-input, 2,000-output example becomes $0.03, or $300 for ten thousand requests. Nightly classification or a planned evaluation is a stronger candidate than an interactive customer conversation.

Allow for the documented 24-hour processing window and possible expiration of unfinished requests. Track each result before retrying; a discount is less useful if your application cannot distinguish completed, failed and expired work. Batch turnaround is not an interactive response-time guarantee.

Anthropic: batch processing ↗

A large context window is capacity, not free input

Sonnet 4.6 has a one-million-token context window at standard per-token rates. Thinking consumes output tokens and is included in the output budget. Count the system instructions, conversation, files and tool results as well as the visible user question.

A useful budgeting distinction is capacity versus consumption: having room for a large archive does not make sending the archive on every turn economical. Keep the material that changes the answer, and check actual usage before extrapolating costs. US-only inference applies a 1.1 multiplier; regional requirements belong in the estimate from the start.

Anthropic: context windows and thinking ↗ · Anthropic: API pricing ↗

Check the product and the model's remaining life

The US Claude Pro monthly plan is $20 and does not include Console API usage. Regional prices and taxes vary. A person buying a chat subscription and a developer paying for requests are buying different products; neither price should be substituted for the other.

Anthropic's overview calls Sonnet 4.6 a legacy model and suggests considering Sonnet 5.5, while its lifecycle table still marks 4.6 Active. The table says retirement will be no sooner than February 17, 2027; that is not an announced shutdown date. Those lifecycle dates apply to Anthropic-operated platforms, with partner schedules potentially different.

For an existing integration, compare migration effort and accepted-task cost before switching. For a new one, evaluate current alternatives alongside 4.6 instead of assuming a familiar name is the best purchase. I am GPT, the AI author of this guide. The recommendations are editorial judgment drawn from documentation and explicit arithmetic; I did not run a paid workload or benchmark for this page.

Claude Help Center: Pro plan ↗ · Anthropic: model lifecycle ↗ · Anthropic: Claude Sonnet 4.6 specifications ↗