Every frontier AI model’s per-token price, current and historical

Headline tier is API list price. Every figure on this page is the public per-million-token rate the provider documents on its own pricing page. Negotiated enterprise pricing, committed-spend discounts, and special programs (OpenAI startup credits, Anthropic enterprise volume, Google Vertex AI region rates) are out of scope — this is a list-price reference.

Cached input pricing is a third axis that has grown more important since OpenAI introduced automatic prompt caching in October 2024 and Anthropic shipped explicit prompt caching in mid-2024. For workloads with a long stable prefix (agents replaying the same system prompt; RAG against a fixed corpus), the cached input rate dominates the effective per-call cost. The page surfaces it as a column rather than the headline number because non-cached workloads still see the headline rate.

Batch endpoints typically halve both input and output rates (OpenAI Batch, Anthropic Message Batches, Google Batch API) in exchange for asynchronous turnaround up to 24 hours. The row note flags batch availability per model; the headline number is the synchronous-API rate.

Open-weights pricing is a hosting-provider proxy. Meta’s Llama models and Muse Glimmer have no “official” per-token price — the weights are free to download — so those rows record Together AI’s published rate as a reference and carry “(Together)” in the name. DeepSeek, Mistral and Alibaba also publish open weights but sell their own APIs, and those rows quote the vendor’s own rate. Cheaper rates exist on self-hosted inference (vLLM, sglang) or on cost-optimized providers (Fireworks, DeepInfra); more expensive ones exist on the hyperscalers (AWS Bedrock, Azure AI Foundry).

Reasoning tokens count as output. The o-series, Anthropic’s extended-thinking models, and Gemini’s 2.5 Flash thinking budget all bill reasoning tokens at the output rate. A “cheap-looking” reasoning model can be expensive in practice because each call generates more reasoning than visible output. The page records the per-token rate; the effective per-call cost is workload-dependent.

Tokenizer differences cross-provider. $1 / M tokens does not buy the same amount of text from OpenAI’s o200k tokenizer as it does from Anthropic’s claude-3 tokenizer or Google’s SentencePiece. It does not even buy the same amount within one provider: Anthropic’s pricing page states that Claude 4.7 and later, plus the Mythos line, use a newer tokenizer producing approximately 30% more tokens for the same text, while Claude Sonnet 4.6 and earlier use the previous one — so a per-token rate that looks flat across those two groups is a real increase in what a page of text costs. Compare numbers in tokens, not in characters or pages, and treat the table as a same-provider apples-to-apples rather than a strict cross-provider one.

Historical pricing is sourced from Wayback Machine and launch posts. Every effective-date entry links to a primary source — provider blog post, archived pricing page, or release announcement. Where the Wayback snapshot is the only viable cite (the provider has rewritten the page since), the URL points there directly.

Qwen prices are the International (Singapore) list, not the Chinese-mainland one. Alibaba publishes two USD price tables for the same Qwen model — an International deployment scope and a Chinese-mainland one — and they differ by roughly 3–5×. Every current Qwen rate here is the International figure, because the rest of this table is US-hosted APIs and that is the comparable basis. Some pre-2026 Qwen points predate that basis and are quoted on the mainland table instead; those are kept as historical record and their Δ cell reads “rebased” rather than a percentage, because the gap between the two tables is a change of reference, not a price the provider moved.