Blended across the normalized routing mix.
Free LLM inference cost calculator
Compare inference architectures before you benchmark them.
Project the cost, latency, and quality of three model configurations, including routing, retries, provider prompt caching, and full-response caching. Then export the assumptions for a representative local test.
04 / Project the frontier
Model the experiment before paying to run it.
Replace the illustrative inputs with a measured workload profile and provider prices. The planner identifies configurations that appear dominated and makes the assumptions portable.
- Three configurations
- Two cache types
- Browser-only calculations
Projection workspace
Illustrative defaults. Replace prices and performance assumptions with your own evidence.
Shared workload
Traffic + routing mix
Configuration A
Tokens per model call
Provider price per 1M tokens
Full-response cache
A hit serves a stored answer and avoids the model call entirely.
Configuration B
Tokens per model call
Provider price per 1M tokens
Full-response cache
A hit serves a stored answer and avoids the model call entirely.
Configuration C
Tokens per model call
Provider price per 1M tokens
Full-response cache
A hit serves a stored answer and avoids the model call entirely.
Projection assumptions
Requests are routed once and traffic shares are normalized to 100%. A response-cache hit avoids the model path. Extra retry attempts apply only on cache misses: 25% means 1.25 model attempts per miss and 150% means 2.5. Retry latency is modeled as sequential, full-path time. Input includes prompt-cached input; output must include every billable output category, including provider reasoning tokens when they are charged as output. Costs are USD provider usage only and exclude implementation, infrastructure, and labor.
Projected blended portfolio
A planning baseline—not a measured benchmark.
$11,724projected provider cost / month
Derived from entered path, cache, and retry assumptions.
User-supplied score, not evaluated by this page.
After response-cache hits and retry assumptions.
Positive means lower cost versus the same mix with no caching.
Projected configuration map
Cost × latency, with quality shown by point size
Lower-left is less expensive and faster. A ring marks a non-dominated configuration.
| Configuration | Status | Normalized traffic | Cost / 1K | Monthly cost | Average latency | Quality |
|---|---|---|---|---|---|---|
| Frontier model | Frontier | 40% | $23.40 | $9,360 | 988 ms | 92 / 100 |
| Efficient model | Frontier | 45% | $4.5873 | $2,064 | 509 ms | 86 / 100 |
| Specialized endpoint | Frontier | 15% | $1.9958 | $299 | 297 ms | 81 / 100 |
Projection boundary
Prices are not verified, quality is user-supplied, and latency is modeled as a weighted average. This page does not test rate limits, tail latency, failures, cache eligibility, stale responses, or workload-specific quality. Validate the frontier with a representative local benchmark.
Method version: 2026-09-05.
What the projection separates
Cost, speed, and quality belong in the same decision.
A lower token price is not automatically a better architecture. Routing, cache eligibility, retries, response length, latency, and workload quality can change which configuration is economically defensible.
Configuration frontier
Compare cost per thousand requests, projected average latency, and a user-supplied quality score without collapsing them into one arbitrary rating.
Two kinds of caching
Provider prompt caching discounts eligible input tokens. A full-response cache avoids the model call. The planner models them separately.
Portable experiment
Download the entered workload and configurations as JSON so the assumptions can become a repeatable local benchmark instead of a screenshot.
What still requires a real benchmark
This page cannot measure workload quality, tail latency, provider reliability, cache eligibility, retrieval relevance, or production failure behavior. Those require representative prompts, real endpoints, and an evaluation rubric run locally with credentials controlled by the user. The public planner helps decide which configurations are worth that test—it does not replace the test.