LLM API pricing in 2026: what a million tokens really costs

LLM API pricing shifted under everyone’s feet this summer: DeepSeek — the default “cheap tier” of 2025 — repriced upward (and has since trimmed V4.1 Flash back down a little), Google is running promo pricing with a scheduled increase, and new flagship models landed at aggressive prices. Here is every major model’s current cost per million tokens as of 2026-10-12, and what an actual production workload costs on each.

The full price table

What changed in 2026

• DeepSeek repriced — twice. V4 Flash went from $0.14/$0.28 to $0.44/$1.32 per million (input/output) this summer; by late September the default deepseek-flash endpoint had moved to V4.1 Flash at $0.30/$1.20. Off-peak hours are half price and cache hits stay near-free, but the “10× cheaper than everyone” era is over — budget workloads built on 2025 DeepSeek math need re-costing.
• Promo pricing with expiry dates. Gemini 3.8 and 3.6 Flash list at $0.75/$3.75 (3.7 and 3.5 Flash were deprecated on 8 October and now auto-route to newer Flash models) only through 2026 — the posted price doubles to $1.50/$7.50 in January 2027. If you’re budgeting a year ahead, use the 2027 number.
• Cheaper flagships. Claude Sonnet 5 launched at $2/$10 — below the $3/$15 its predecessor charged — and Opus 5.5 ($4/$20) undercuts Opus 5 ($5/$25). OpenAI’s GPT-6 family (Astra $10/$50, Sol $2/$10, Luna $0.10/$0.50) halves the GPT-5.6 Sol and Luna rates. The point releases kept those prices: Claude Sonnet 5.5 (28 Sep) is still $2/$10, and GPT-6.1 Sol (30 Sep) is $2/$10 with cached input cut to $0.10. Raw intelligence per dollar keeps improving even as the floor rises. At the budget end, Claude Haiku 5.5 (7 October) lists at $0.10/$0.50 — a tenth of Haiku 4.5 and level with GPT-6 Luna — but only for prompts up to 100K tokens; a longer prompt bills the whole request at $0.50/$2.50. One caveat: GPT-5.6 Sol’s $4/$20 is a promo that OpenAI says runs at least through 21 November 2026 (it was $5/$30).
• Caching is the real discount. OpenAI’s cached input is ~10× cheaper, DeepSeek’s near-free, Anthropic’s prompt caching similar. For chat workloads where 90% of each request is repeated context, the cache price matters more than the list price.

A real workload, costed

Per-token prices are abstract, so: a support chatbot handling ~100k conversations a month — roughly 100M input tokens (system prompt, context, history) and 10M output tokens. Uncached monthly cost per model:

The spread is the story: the same workload runs from about twenty dollars to fifteen hundred per month depending on the model — a 70× range for the identical token volume. Most production systems land on a two-model split — a budget model (Luna, Haiku, DeepSeek off-peak) for the 80% of easy traffic and a flagship for the hard 20% — plus aggressive prompt caching, which can cut the input bill by an order of magnitude on chat-shaped workloads.

Methodology: prices are from official provider pricing pages (Anthropic, OpenAI, Google, DeepSeek, Mistral, xAI) as of 2026-10-12, date-stamped and refreshed with the rest of our data. Tiered prices use the base tier (≤200k prompt where providers split); cached-input and batch discounts are noted but not included in the workload math. No AI provider pays us anything.