AI Model Pricing

On input, GLM-4.7-FlashX at $0.070 per 1M tokens. On output, GLM-4.7-FlashX at $0.400 per 1M. Which one is cheapest for you depends on the shape of your traffic: a 10,000-token prompt with a 1,000-token answer costs $0.00110 on GLM-4.7-FlashX and $0.032 on GPT-5.6 Terra. Set the two token fields above to your own volumes and the table re-sorts around them.

Weight input and output by how you actually use the model. Providers charge three to five times more for what the model writes than for what you send — GPT-5.6 Terra is $2.00 in and $12.00 out — so a summariser that reads a lot and writes a little ranks the models completely differently from a code generator that does the opposite. Enter both volumes above rather than comparing input rates alone, and check the notes: several providers re-price the entire request once a prompt crosses an input threshold.

When consecutive requests share the same prefix — a system prompt, a document, a long few-shot block — providers can serve that prefix from cache and charge less for it. The cached column is that cache-hit read rate, usually 10% to 20% of the standard input rate: Claude Sonnet 5 reads cached input at $0.200 against $2.00 standard. Writing to the cache is not always free, and a dash means the provider publishes no cached rate at all.

Every figure was read from the provider's own pricing page on August 22, 2026, and each provider card links back to the page it came from. Nothing here is scraped live, so treat it as a snapshot: providers change rates without notice, and the ones with promotional or off-peak pricing (MiniMax, DeepSeek) move most often.

Because the provider does not publish that figure, and guessing would be worse than leaving it blank. Perplexity publishes no context window for the Sonar models, Z.ai publishes no per-variant context length for GLM-4.5-X, and Meta’s hosted LLaMA row has no cached rate. A dash always means "not published", never "zero".

11 of the 56 rows — the OpenAI models, because OpenAI publishes its tokenizer and the calculator loads the real BPE vocabulary (o200k_base) in your browser. Everything else is marked "est." and counted with a characters-per-token approximation, which lands within roughly 10–15% on English prose but drifts on code, JSON and non-Latin scripts. The prices are first-party for every row; only the token counts differ in confidence.

No — the table shows token rates only, and a few providers bill on top of them. Sonar adds $5 to $12 per 1,000 requests for search depending on context size, Sonar Deep Research adds citation-token, reasoning-token and per-search charges, and hosted open-weight models carry whatever the host charges rather than a first-party rate. The notes column flags every row where the token rate is not the whole bill.

Take 200,000 input tokens — a full novel, or a few hundred pages of a contract. That single prompt costs $0.400 on GPT-5.6 Terra, $0.400 on Claude Sonnet 5, $0.600 on Kimi K3 and $0.020 on Qwen3.5 Flash. Watch the tier notes at that size: on Grok 4.6 a prompt of 200K tokens bills the whole request at double the listed rates.