Pricing

Mortar charges per token. A prompt token read from the prefix cache costs 0.1× the input price; one that has to be processed is written to the cache for later requests at 1.25×, so caching saves money once more than about 22% of prompt tokens are read from it.

These are sample rates, not a quote.

Mortar has not published prices yet. Each rate is derived from on-demand GPU list prices and measured throughput at 65% utilisation.

How you are charged

Managed Cloud
Per token, at the rates below. A prompt token read from the prefix cache costs 0.1× the input price, and every token that has to be processed is written to the cache at 1.25×. Output tokens are billed at the output rate.
Bring Your Own Cloud
Fees are not published yet. Email hello@mortalinference.si for a quote. Your cloud provider bills you for the GPU nodes.
Currency
US dollars. Figures on this page exclude tax.
Billing month
730 hours, the cloud-billing convention, so a monthly figure reconciles with an invoice.

Rates

US dollars per million tokens. Sample rates, not a quote.
Model Input Read from cache Written to cache Output
Qwen 2.5 7B Instruct Qwen/Qwen2.5-7B-Instruct Sample rate $0.0313 $0.0031 $0.0391 $0.2373 Estimate
Llama 3.1 8B Instruct meta-llama/Llama-3.1-8B-Instruct Sample rate $0.0347 $0.0035 $0.0434 $0.2654 Estimate
Mistral 7B Instruct mistralai/Mistral-7B-Instruct-v0.3 Sample rate $0.0297 $0.0030 $0.0371 $0.2278 Estimate
Qwen 2.5 32B Instruct Qwen/Qwen2.5-32B-Instruct Sample rate $0.1529 $0.0153 $0.1911 $1.1046 Estimate
Llama 3.3 70B Instruct meta-llama/Llama-3.3-70B-Instruct Sample rate $0.6253 $0.0625 $0.7816 $4.5671 Estimate
DeepSeek-V3 deepseek-ai/DeepSeek-V3 Sample rateComing soon $1.1057 $0.1106 $1.3821 $7.7809 Estimate

Context windows, model families and availability are in the model catalogue.

Estimate your monthly cost

Enter your own traffic. Nothing is filled in for you: the cache-hit rate has no default, and Mortar is compared with the same model without a prefix cache, or with prices you give for your current provider.

Or start from an illustrative example:

Your workload

Priced at its rate in the table above.
Requests are counted

Commas are fine, as in 2,000,000.

The whole prompt: instructions, history and the new message.

The reply.

The share of input tokens served from the prefix cache. There is no default. Not sure? Run cost lint on your prompt.
Compare with

Shows what the cache alone is worth.

Enter what you pay today. Mortar never fills in another vendor's prices.

Your provider's prices, from your invoice or its price page.

Enter 0 if they do not cache prompts.

0.1 means a cached token costs a tenth of the input price.

What the estimate leaves out

  • Tax.
  • The GPU nodes of a Bring Your Own Cloud deployment, which your cloud provider bills.
  • Network egress.
  • Committed-use discounts, at Mortar or at your provider.
  • Any surcharge your provider adds for writing to its cache, as Anthropic documents. Leaving it out understates what you pay today, so a saving shown here is, if anything, too small.

Spend controls

Mortar does not offer spend alerts or spend caps yet.

Questions

Why does output cost more than input?

A prompt's tokens are processed together in one pass. A reply is generated one token at a time, so each output token takes far more GPU time than an input token.

What counts as a cached token?

A prompt token that Mortar's prefix cache serves instead of processing again, at 0.1× the input price. The cache matches from the first token of a prompt and stops at the first difference, so only a shared prefix counts; every token after it is processed and written to the cache at 1.25×. Write cache-friendly prompts explains how to order one.

How do I measure my cache-hit rate?

Divide cached prompt tokens by all prompt tokens over a period. OpenAI-compatible APIs report the cached count for each response in usage.prompt_tokens_details.cached_tokens. On Mortar, the telemetry stream reports each session's cacheHitRate: see Measure your hit rate.