List prices per 1M tokens for the major hosted LLMs in 2026. Use the calculator to see what they would cost for your workload.
Cite asHarpd. "Harpd Model Pricing Dataset." harpd.com/llm-pricing/ Accessed: 2026-08-18. License: CC BY 4.0.
Quick answer
Harpd publishes list prices per 1M tokens for 16 hosted models in 2026 — Claude, GPT, Gemini, DeepSeek, Llama, Mistral, Qwen — each record verified against its provider's own pricing page, with the last-verified date shown on every row. The data is also downloadable as JSON and CSV.
Data source: Harpd Model Pricing DatasetLast updated:
A citable price table needs more than a number — it needs the source and the date. By verifying every price against the provider's own page and exposing the source URL and verification date per row, Harpd makes each figure independently checkable and safe for an AI system to quote.
Data source: Harpd Model Pricing DatasetLast updated: August 18, 2026
| Model | Provider | Tier | $ / M input | $ / M output | Est. 100k-call month | Verified | Source |
|---|---|---|---|---|---|---|---|
| Claude Opus 4.8 | Anthropic | flagship | $5.00 | $25.00 | $5,0005k in / 1k out | 2026-08-18 | Anthropic ↗ |
| Claude Sonnet 5Introductory pricing ($2/$10) runs through 2026-08-31; standard $3/$15 applies from 2026-09-01. | Anthropic | fast | $2.00 | $10.00 | $2,0005k in / 1k out | 2026-08-18 | Anthropic ↗ |
| Claude Sonnet 4.5 | Anthropic | fast | $3.00 | $15.00 | $3,0005k in / 1k out | 2026-08-18 | Anthropic ↗ |
| Claude Haiku 4.5 | Anthropic | mini | $1.00 | $5.00 | $1,0005k in / 1k out | 2026-08-18 | Anthropic ↗ |
| GPT-5 | OpenAI | reasoning | $1.25 | $10.00 | $1,6255k in / 1k out | 2026-08-18 | OpenAI ↗ |
| GPT-5 mini | OpenAI | fast | $0.25 | $2.00 | $3255k in / 1k out | 2026-08-18 | OpenAI ↗ |
| GPT-5 nano | OpenAI | mini | $0.05 | $0.40 | $655k in / 1k out | 2026-08-18 | OpenAI ↗ |
| GPT-4.1 | OpenAI | flagship | $2.00 | $8.00 | $1,8005k in / 1k out | 2026-08-18 | OpenAI ↗ |
| Gemini 2.5 ProPrompts over 200k tokens bill at $2.50 input / $15 output. Context caching ~$0.125/1M for ≤200k. | flagship | $1.25 | $10.00 | $1,6255k in / 1k out | 2026-08-18 | Google ↗ | |
| Gemini 2.5 Flash | mini | $0.30 | $2.50 | $4005k in / 1k out | 2026-08-18 | Google ↗ | |
| Gemini 2.5 Flash-Lite | mini | $0.10 | $0.40 | $905k in / 1k out | 2026-08-18 | Google ↗ | |
| Mistral Large 3 | Mistral | fast | $2.00 | $6.00 | $1,6005k in / 1k out | 2026-08-18 | Mistral ↗ |
| Mistral Small 3.1 | Mistral | mini | $0.10 | $0.30 | $805k in / 1k out | 2026-08-18 | Mistral ↗ |
| DeepSeek V3DeepSeek moved to peak/off-peak pricing on 2026-08-17. Figures shown are approximate off-peak list rates — verify before budgeting. | DeepSeek | open | $0.27 | $1.10 | $2455k in / 1k out | 2026-08-18 | DeepSeek ↗ |
| Llama 3.1 70BResale price via Together AI. | Meta · Together | open | $0.88 | $0.88 | $5285k in / 1k out | 2026-08-18 | Meta · Together ↗ |
| Qwen 2.5 72BResale price via OpenRouter. | Alibaba · OpenRouter | open | $0.40 | $0.40 | $2405k in / 1k out | 2026-08-18 | Alibaba · OpenRouter ↗ |
All prices are USD per 1,000,000 tokens, last verified 2026-08-18 against each provider's official pricing page (Source column). Estimates assume 100,000 calls/month, 5,000 input tokens, 1,000 output tokens. Your actual bill depends on caching, batching, free tiers and any negotiated enterprise rates — verify with each provider. Machine-readable copy: /data/llm-pricing.json.
"Flagship" usually means the strongest model in a provider's line; "mini" usually means the cheapest. The tier label is a navigation aid, not a quality rating — a cheap model can still beat a flagship on your specific task.
Almost every provider charges 2–5× more for output tokens. If your workload is generation-heavy (long agent loops, code synthesis, reports), the output column matters more than the input column.
Enterprise contracts, committed-use discounts, prompt caching and batch APIs can change the math 2–10×. This page is for planning; for actual billing, check your invoice or your provider's pricing dashboard.
A model that costs 1/10 the price but only succeeds on 60% of your tasks is more expensive per successful task. Cost per successful task models this; verify the assumptions on your real workload.
Each price is read from the provider's own published pricing page; the record stores both the source URL and the date it was verified. Prices are list prices in USD per 1M tokens, input and output reported separately. A record is only re-verified when the provider page is re-checked, so the data date moves with the data, not with a rebuild. Full methodology →
Underlying data date: