Claude Sonnet 4.6 is Anthropic's latest mid-range model — released in February 2026. It sits between the budget-friendly Haiku line and the premium Opus tier, making it the default choice for most production workloads: coding, chat, document analysis, and tool use.
But Anthropic's pricing isn't just "input + output." There's a layered caching system with two TTL tiers, a Batch API discount, and a data residency surcharge that can all stack on top of each other. This guide breaks down every component so you know exactly what you're paying — and how to pay less.
Last updated: April 27, 2026
Disclaimer: Prices in this article are accurate as of the publication date. Anthropic may adjust pricing at any time. Always verify on the official Anthropic pricing page before making production decisions.
The foundation of Claude Sonnet 4.6 pricing is straightforward:
| Token Type | Price per 1M Tokens |
|---|---|
| Input (base) | $3.00 |
| Output | $15.00 |
That's a 5:1 output-to-input ratio. For every dollar you spend on input tokens, you'd spend five dollars on the same number of output tokens. This ratio matters — if your workload is output-heavy (code generation, long-form writing), output costs will dominate your bill.
| Workload | Tokens | Cost |
|---|---|---|
| 1 short chat turn (500 in / 200 out) | 700 total | $0.0045 |
| 1 code review (2K in / 1K out) | 3K total | $0.021 |
| 1 document summary (10K in / 2K out) | 12K total | $0.06 |
| 1 hour of chatbot traffic (500K in / 200K out) | 700K total | $4.50 |
| 1 day of heavy API usage (5M in / 2M out) | 7M total | $45.00 |
Prompt caching is where Anthropic's pricing gets interesting — and where the real savings live.
When you send a request with cache_control enabled, Anthropic stores the computed state of your prompt prefix. On subsequent requests that start with the same bytes (same system prompt, same few-shot examples, same preamble), those tokens are served from cache instead of being reprocessed.
There are two cache duration tiers:
| Cache Operation | Price per 1M Tokens | Multiplier vs Base Input | Duration |
|---|---|---|---|
| 5-minute cache write | $3.75 | 1.25x | 5 minutes |
| 1-hour cache write | $6.00 | 2.0x | 1 hour |
| Cache hit (read) | $0.30 | 0.1x | — |
5-minute cache (1.25x write):
1-hour cache (2.0x write):
| Scenario | Recommended Cache | Why |
|---|---|---|
| Real-time chatbot (many requests/minute) | 5-minute | High request frequency, cache stays warm |
| Batch processing (bursts every few minutes) | 5-minute | Requests cluster within 5-min windows |
| Long-running agent sessions | 1-hour | Requests spread over 10-60 minutes |
| Scheduled jobs (hourly reports) | 1-hour | Predictable hourly pattern |
| One-off requests | No cache | No reuse opportunity |
Automatic caching (recommended for most cases):
Explicit cache breakpoints (fine-grained control):
Every response includes cache metrics in the usage object:
cache_creation_input_tokens: tokens written to cache (charged at 1.25x or 2x)cache_read_input_tokens: tokens read from cache (charged at 0.1x)input_tokens: tokens processed normally (charged at base rate)Your actual input cost for a request is:
Anthropic offers a Batch API for asynchronous processing. You submit requests in bulk, and results are returned within 24 hours. The tradeoff: no real-time responses, but you get a 50% discount on all token types.
| Token Type | Standard | Batch API |
|---|---|---|
| Input | $3.00/M | $1.50/M |
| Output | $15.00/M | $7.50/M |
| 5-min cache write | $3.75/M | $1.875/M |
| 1-hour cache write | $6.00/M | $3.00/M |
| Cache hit | $0.30/M | $0.15/M |
The Batch discount stacks with caching. If you're running a nightly batch job with a consistent system prompt, you get both the 50% batch discount AND the 0.1x cache read discount on repeated prefixes. That's $0.15/M for cached input tokens in batch mode — 95% cheaper than standard base input.
Starting with Claude Sonnet 4.5 and newer models (including Sonnet 4.6), Anthropic charges a 1.1x multiplier if you specify US-only inference via the inference_geo parameter.
| Token Type | Global (default) | US-only (1.1x) |
|---|---|---|
| Input | $3.00/M | $3.30/M |
| Output | $15.00/M | $16.50/M |
| Cache write (5min) | $3.75/M | $4.125/M |
| Cache hit | $0.30/M | $0.33/M |
This surcharge stacks with everything else. If you use US-only + Batch + caching, all multipliers apply.
Most users don't need this — global routing is the default and has no surcharge. Only enable inference_geo if you have strict data residency requirements.
Through Crazyrouter, Claude Sonnet 4.6 is available at 55% of official pricing — a 45% discount on base token rates.
| Token Type | Anthropic Official | Crazyrouter (55%) |
|---|---|---|
| Input | $3.00/M | $1.65/M |
| Output | $15.00/M | $8.25/M |
Crazyrouter supports both OpenAI-compatible and native Anthropic API formats, so you can use whichever SDK you prefer.
OpenAI-compatible format:
Anthropic-native format:
curl:
Let's compare costs for three common workloads: Anthropic direct vs Crazyrouter.
| Anthropic Direct | Crazyrouter | |
|---|---|---|
| Input cost | $3.00 | $1.65 |
| Output cost | $7.50 | $4.125 |
| Daily total | $10.50 | $5.775 |
| Monthly (30 days) | $315.00 | $173.25 |
| Monthly savings | — | $141.75 (45%) |
| Anthropic Direct | Crazyrouter | |
|---|---|---|
| Input cost | $1.50 | $0.825 |
| Output cost | $30.00 | $16.50 |
| Daily total | $31.50 | $17.325 |
| Monthly (30 days) | $945.00 | $519.75 |
| Monthly savings | — | $425.25 (45%) |
With caching on Anthropic direct:
With Crazyrouter (no native cache, but 45% off base):
Even with Anthropic's caching at 60% hit rate, Crazyrouter's flat 45% discount still comes out cheaper for this workload. The gap narrows at very high cache hit rates (80%+), where Anthropic's cache reads at $0.30/M become extremely cheap.
At what cache hit rate does going direct to Anthropic beat Crazyrouter?
For a pure input workload (ignoring output for simplicity):
Solving: 3.75 + x × $0.30
At cache hit rates above ~61%, going direct to Anthropic with 5-minute caching is cheaper for input tokens. But remember: output tokens have no caching discount, and Crazyrouter's 45% off applies to output too. For output-heavy workloads, Crazyrouter wins at any cache hit rate.
| Component | Anthropic Official | Crazyrouter |
|---|---|---|
| Base input | $3.00/M | $1.65/M |
| Base output | $15.00/M | $8.25/M |
| 5-min cache write | $3.75/M (1.25x) | — |
| 1-hour cache write | $6.00/M (2.0x) | — |
| Cache hit | $0.30/M (0.1x) | — |
| Batch input | $1.50/M (50% off) | — |
| Batch output | $7.50/M (50% off) | — |
| US-only surcharge | 1.1x all prices | — |
| Supported formats | Anthropic API | OpenAI + Anthropic |
Base pricing is 15/M output. Output is 5x more expensive — optimize for shorter outputs when possible.
Prompt caching saves up to 90% on input tokens. The 5-minute cache pays for itself after just 1 reuse. The 1-hour cache needs 2 reuses.
Batch API cuts everything by 50%. Stack it with caching for up to 95% savings on cached input tokens.
Crazyrouter offers a flat 45% discount on base token rates, with no caching complexity to manage. For output-heavy workloads, this is often the better deal.
The optimal strategy depends on your workload. High cache hit rates + input-heavy = go direct. Output-heavy or unpredictable traffic = Crazyrouter wins.
Sign up at crazyrouter.com to access Claude Sonnet 4.6 at 45% off — along with 300+ other models from OpenAI, Google, xAI, and more, all through one API key.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。