Model the whole request
Separate uncached input, cache writes, cache reads and generated output. Long prompts can trigger another rate tier.
MULTI-MODEL GUIDE · UPDATED OCTOBER 12, 2026
LLM API Pricing
Comparison
Compare published input and output token rates, then estimate a real workload. Rates are snapshots—not promises—so verify the cited provider before deployment.
01 / PRICE TABLE
Compare LLM prices per million tokens.
This version includes models whose rates are already documented on this site. New providers are added only with a source and verification date.
| Model | Provider / route | Input / MTok | Output / MTok | Guide |
|---|---|---|---|---|
| Claude Haiku 5.5 · ≤100K | Claude Platform | $0.10 | $0.50 | Details |
| Claude Haiku 5.5 · >100K | Claude Platform | $0.50 | $2.50 | Calculator |
| Claude Sonnet 5.5 | Claude Platform | $2.00 | $10.00 | Details |
| Claude Opus 5.5 | Claude Platform | $4.00 | $20.00 | Details |
| Step 5 Preview | OpenRouter | $1.00 | $2.70 | Details |
A lower input rate does not guarantee a lower completed-task cost. Include output length, retries, caching, latency and usable-result rate.
02 / TOKEN COST CALCULATOR
Estimate one workload across models.
All calculations run locally. No usage data is submitted.
Estimated monthly token cost$2.00$0.10 input + $0.50 output per MTok
03 / HOW TO COMPARE
Price per token is only the starting point.
Measure accepted-task cost
Divide spend by usable answers, resolved tickets or accepted patches—not raw request count.
Record route and date
The same model can differ across first-party and aggregator routes.
04 / FAQ
LLM API pricing questions
How do I calculate LLM API cost?
Divide input and output tokens by one million, multiply each by its rate, then add caching, tool, image, regional or provider-specific charges.
What is the cheapest LLM API?
There is no universal answer. Low token rates can lose their advantage when a workload needs more output, retries or escalation.
Why can two sites show different prices?
They may refer to different providers, regions, prompt-length tiers, cache rules or verification dates.
Should I choose a model using benchmarks?
Use benchmarks to shortlist candidates, then test your own prompts and acceptance criteria. Separate independent results from vendor claims.
Need Claude-specific detail?
Compare Haiku, Sonnet and Opus rates and scenarios.