Cheapest LLM API — Inference Provider Comparison

This is a comparison, not a black box. Every provider is priced for the same reference open model so the numbers line up: pick Llama 3.3 70B and you see what Together, DeepInfra, Groq and the rest each charge to run exactly that model. Two figures drive the decision — price per million tokens and output throughput in tokens per second — and the tool computes both from published data, with no network call and no opinion baked in.

1. Price per million tokens

Every host on this list quotes in US dollars per 1,000,000 tokens, billed separately for input (prompt) and output (completion) tokens. Your estimated monthly cost is a plain weighted sum:

cost = (input ÷ 1,000,000) × input $/1M + (output ÷ 1,000,000) × output $/1M

Output tokens are usually the same price or dearer than input tokens, which is why a chat workload heavy on generation costs more than a classification workload that reads a lot and writes little.

2. Output throughput and generation time

Speed is the second axis. Output throughput — how many tokens per second a host generates — comes from Artificial Analysis' independent cross-provider measurements, not the vendors' own marketing. From it the tool derives a “time to generate 1M tokens” readout, seconds = 1,000,000 ÷ tokens-per-second, so the price↔speed trade-off is visible: a specialised host like Cerebras or Groq is many times faster than a cheap GPU-backed marketplace, and that gap decides whether the extra cost is worth it for a latency-sensitive chat UX.

3. Ranking and the rupee figure

Providers that do not host the selected model are dropped from the ranking rather than counted as free. The cheapest available host is highlighted and the fastest is badged; you can re-sort by cheapest, fastest, or provider name. Every dollar total is converted to rupees at an editable USD→LKR rate, prefilled from the Central Bank of Sri Lanka indicative figure (305 LKR/USD at the last-verified date) — useful when you earn in dollars but budget locally.

4. Cross-check

Providers quote per-1M tokens, but many SDK cost fields and invoices report per-1K. As an internal check the same monthly cost is recomputed from per-1,000-token unit prices (price ÷ 1000) × (tokens ÷ 1000); the per-1M and per-1K methods agree to the cent — the same idea as checking a tax figure against the regulator's own formula. All prices carry a last-verified date because serverless rates change often; this is a directly-comparable snapshot, not a live feed.

• Together AI — model pricing
• Fireworks AI — serverless pricing
• DeepInfra — per-token pricing
• Groq — pricing & tokens/sec
• Cerebras — inference pricing
• OpenRouter — models & pricing
• Novita AI — pricing
• Hyperbolic — pricing
• Artificial Analysis — independent throughput & latency measurements
• Central Bank of Sri Lanka — indicative USD/LKR exchange rate

Every price and throughput figure was last re-confirmed against the sources above on 2026-07-03. Serverless inference pricing and measured speeds change frequently — these are illustrative reference figures for budgeting and comparison, not live quotes. Confirm the current rate on the provider before you commit.