1. Direct-provider pricing
30 listings are refreshed daily from the official pricing documentation published by OpenAI, Anthropic, Google, xAI, DeepSeek, and Mistral. These rows are labeled Direct, link to the first-party page, and were last validated 2026-09-09 07:31:17 UTC.
2. Official inference-provider pricing
10 listings are refreshed from pricing sources published directly by Groq, Together AI, and Cerebras. These rows are labeled Official endpoint: the price comes from that inference provider, not from the model creator and not from OpenRouter. The current official-endpoint snapshot was validated 2026-09-09 07:31:18 UTC.
3. OpenRouter endpoint pricing
426 listings are synchronized daily from OpenRouter's public models API. OpenRouter reports dollars per token; LLMPrice multiplies those values by 1,000,000 for display. Invalid or negative router records are excluded. The current snapshot was captured 2026-09-09 07:31:17 UTC.
4. Workload calculation
Billable requests include any retry/extra-call percentage entered in Advanced billing options. Reasoning tokens entered there are treated as output tokens for pricing and output/context-limit checks. When an endpoint does not publish a cache-read or cache-write rate, the calculator uses the normal input rate for that portion. This avoids inventing a discount.
5. Model-specific billing rules
The first-party catalog includes published batch discounts, long-context thresholds, cache-write rates, and DeepSeek peak/off-peak prices where applicable. OpenRouter records use the model-level rates supplied by its API and do not inherit first-party batch rules.
6. Capabilities and limits
OpenRouter context limits, modalities, and supported parameters come from the same API snapshot. A capability means the endpoint advertises the related input modality or parameter; it is not an independent quality score.
7. What is not included
Estimates exclude taxes, rate limits, minimum commitments, tool calls, web-search charges, media-token conversion, cache storage, regional premiums, retry overhead beyond the percentage you enter, and negotiated discounts unless a model record explicitly includes them. Tokenizers also differ, so identical text can produce different token counts across models.
8. Why there are no copied benchmark scores
LLMPrice does not copy another site's proprietary rankings. Quality benchmarks need their own licensed or first-party source, model-version matching, and evaluation date. Price and capability comparison are kept separate until that data can be published responsibly.
9. Automatic update safety
All three source groups run every day. A run publishes only after all tracked direct rows, router records, billing relationships, generated pages, links, and calculator tests pass. Missing rows, malformed prices, suspicious large changes, or a changed source-page layout stop the run, leaving the last known-good catalog live.