LLM API pricing, compared

Mistral AIMinistral 3 (3B)price changed 2026-10

Mistral Large 3 ($0.50 / $1.50) is still sold but superseded by Large 4. OCR and voice models are priced per page or per minute and aren't modeled here.

Standard-tier prices for models also sold by their own labs; many more open models are listed. Cached input is billed lower.

GroqGPT OSS 20B

Prices from Groq's production model list (groq.com/pricing no longer shows a price table). Llama 3.1 8B and Llama 3.3 70B are listed as contact sales.

International (BytePlus) prices for the lowest input-length tier. Flex (off-peak) online inference and batch inference cost about half. Mainland prices (Volcano Engine Ark) are in yuan.

Cache-miss input prices. Cache hits cost a fraction of a cent (Pro $0.0036, Flash $0.0028 per million) and cache writes are free for now. The Batch API halves Pro and Flash prices. Mainland prices are in yuan, and a separate Token Plan subscription covers coding tools.

ClaudeClaude Haiku 5.5price changed 2026-10

Base input/output prices; prompt-cache reads are billed much lower and cache writes higher. US-only inference is 1.1x, and fast mode for Opus 5.5 is 2x. Batch processing is discounted.

International prices for the lowest input-length tier. ERNIE 5.0, which adds image input and a thinking mode, costs $0.84 / $3.38.

OpenAI APIGPT-6 Luna

Standard-tier prices for short context (the page doesn't state the cutoff for GPT-6 models; older models switch at 272K input tokens); long-context requests cost more (Astra $20 / $75, Sol $4 / $15, Luna $0.20 / $0.75 per million). Batch is half price; cached input is billed lower.

Cached input is billed lower (Hy4 preview $0.042, Hy3 $0.033 per million). Hy4 is still a preview release.

International (Singapore) list prices for the lowest input-length tier. Qwen3.7-Plus is marked "limited-time 20% off" on the page; the list price is shown. Batch calls are half price.

Fireworks AIGLM 5.3 Flash

Serverless standard prices for the models also sold by their own labs. Models not listed are priced by size, from $0.10 per million tokens under 4B parameters.

International (Z.ai) prices, flat across input lengths. GLM-4.7-Flash, GLM-4.5-Flash and GLM-4.6V-Flash are free. Mainland prices (bigmodel.cn) are in yuan and tiered by length for some models.

Together AIgpt-oss-120B

Serverless prices for the models also sold by their own labs; many more open models are listed on the page. Cached input is billed lower where shown, and the Batch API is cheaper.

A rate-limited free trial (about 1M tokens a day) and a one-time $5 credit are available. Qwen 3.8 27B is the other shared model, at $0.99 / $1.49.

DeepSeekdeepseek-flash

Peak-hour prices with cache-miss input (peak is 01:00–04:00 and 06:00–10:00 UTC, Monday–Friday). All other hours are half price (v4-pro $0.66 / $1.98, flash $0.15 / $0.60), and cache-hit input is far cheaper, so most workloads pay less than shown.

Cache reads cost $0.06 per million. The small MiniMax-M3.1-Flash-Preview is sold only through subscription plans, not pay-as-you-go. Mainland prices (platform.minimaxi.com) are in yuan.

Gemini APIGemini 3.1 Flash-Lite

Paid-tier Standard text prices. Gemini 3.8 Flash is $0.75 / $3.75 through December 31, 2026 and rises to $1.50 / $7.50 on January 1, 2027. Gemini 3.1 Pro Preview costs $4 / $18 for prompts over 200K tokens. Batch and Flex are 50% off. A free tier exists but may use prompts to improve Google's products.

Grok APIGrok 4.3

Prices for prompts under 200K tokens; at 200K and above input and output prices double. Web and X search tools are billed separately as tool calls.

Cache-miss input prices. Cache hits are far cheaper (K3 $0.30, K2.6 $0.16, K2.7 Code $0.19 per million). A HighSpeed variant of K2.7 Code costs twice as much. Mainland prices (platform.moonshot.cn) are in yuan.

Standard-tier prices for muse-spark-1.3 (also 1.2 and 1.1); cached input is $0.15 per million and there is no long-context premium. The contributor tier is left out because it trades your data for the lower price.