StackProofStackProof

Model pricing · Anthropic · mid tier · August 2026

Claude Sonnet 5 API pricing, in numbers that matter

List price, real cost per task, rank against the field, and the self-hosting break-even, all computed from our open dataset.

Input, $/1M tokens

$2

Output, $/1M tokens

$10

Cost per task

$0.02400

Cost rank

#7 of 10

The task profile behind every figure: 1,500 input and 500 output tokens per call, 3 calls per task. On that profile Claude Sonnet 5 costs $0.02400 per task and $24,000 per million tasks. Output is billed at 5x the input rate, so despite the profile sending three times more input than output, output makes up 63% of the bill. Capping response length is the highest-return prompt-level optimisation on this model.

Claude Sonnet 5 against the field

#Model$/taskvs Claude Sonnet 5Head-to-head
1Mistral Small$0.001580.07xcompare
2DeepSeek-V4-Flash$0.003960.17xcompare
3Gemini 2.5 Flash$0.005100.21xcompare
4Grok 4.3$0.009380.39xcompare
5GPT-5.4 mini$0.010130.42xcompare
6Claude Haiku 4.5$0.012000.50xcompare
7Claude Sonnet 5$0.02400--
8Gemini 3.1 Pro$0.027001.12xcompare
9Claude Opus 4.8$0.060002.50xcompare
10GPT-5.5$0.067502.81xcompare

Self-hosting break-even

Against a reference $2/hour dedicated instance ($1,460/month fixed, 788,400 tasks/month capacity at full utilisation), the break-even against Claude Sonnet 5 is 60,833 tasks per month. That sits inside the box's capacity, so past that volume self-hosting genuinely wins on cost. Work through your own numbers in the $/task calculator or the break-even guide.

Frequently asked

How much does the Claude Sonnet 5 API cost?
Claude Sonnet 5 lists at $2 per million input tokens and $10 per million output tokens as of August 2026. On a realistic task of 1,500 input and 500 output tokens across 3 calls, that is $0.02400 per task, or $24,000 per million tasks.
Is Claude Sonnet 5 expensive compared to other models?
Claude Sonnet 5 ranks #7 of 10 tracked models on cost per task. It costs 15.2x more than the cheapest tracked model (Mistral Small), while the most expensive (GPT-5.5) costs 2.81x more than it.
When does self-hosting beat the Claude Sonnet 5 API?
Against a $2/hour dedicated GPU ($1,460/month), self-hosting becomes cheaper than Claude Sonnet 5 above 60,833 tasks per month, within that box's physical capacity of 788,400 tasks.

See also: all tracked models · every head-to-head comparison · download the dataset