LLM API pricing comparison
What 17 current large language models cost through their APIs, in US dollars per 1M tokens: fresh input, cached input and output side by side. Put in your own token counts to see what the same workload costs on each model, cheapest first.
$0.10 to $20 per 1M input tokens
GPT-6 Luna and Claude Haiku 5.5 are the cheapest here at $0.10 in and $0.50 out. GPT-6 Cyber is the most expensive at $20 in and $100 out. Gemini 4 Argon's $2 / $10 is an introductory rate that doubles on December 1, 2026.
Cost calculator
Enter the tokens for one job, or for a month. Every model is priced at the rates in the table below.
| Model | Estimated cost |
|---|
List prices only. Long-context tiers, batch discounts, reasoning tokens billed as output, and regional uplifts are not included.
Price per 1M tokens
Standard API list prices in USD, most expensive first. Cached input is the price of a prompt-cache hit. Model names link to a full guide with its own calculator where one exists.
| Model | Maker | Input | Cached in | Output | Note | Fingerprint |
|---|---|---|---|---|---|---|
| GPT-6 Cyber | OpenAI | $20.00 | $2.00 | $100.00 | Restricted access | — |
| GPT-6 Astra | OpenAI | $10.00 | $1.00 | $50.00 | Above 272K input: $20 / $75 | Verify |
| Claude Fable 5.2 | Anthropic | $8.00 | $0.20 | $40.00 | Batch API half price | Verify |
| Muse Wonder | Meta | $4.00 | $0.20 | $20.00 | Wonder Full: subscription only | Verify |
| Gemini 4 Argon | $2.00 | $0.10 | $10.00 | Intro price; $4 / $20 from Dec 1, 2026 | Verify | |
| Gemini 4 Carbon | $2.00 | $0.10 | $10.00 | Gated access; standard $4 / $20 | Verify | |
| Kimi K3.1 Pro | Moonshot AI | $2.00 | $0.20 | $10.00 | Highspeed variant $4 / $20 | Verify |
| GPT-6 Sol | OpenAI | $2.00 | $0.20 | $10.00 | Above 272K input: $4 / $15 | Verify |
| Tencent Hy4 | Tencent | $0.834 | $0.042 | $2.501 | 50% off on TokenHub and OpenRouter | — |
| Mistral Large 4 | Mistral | $0.70 | $0.07 | $2.10 | Open weights | — |
| LongCat 2.5 | Meituan | $0.30 | $0.006 | $1.20 | Limited-time discount | — |
| Inkling 2 Small | Thinking Machines Lab | $0.30 | $0.03 | $1.00 | On Tinker | — |
| Mercury 3 | Inception | $0.20 | $0.02 | $0.70 | 50% off at launch | — |
| MiniMax M3.1 Flash | MiniMax | $0.15 | $0.015 | $0.60 | — | Verify |
| DeepSeek V4.1-Flash-1017 | DeepSeek | $0.15 | $0.003 | $0.60 | Off-peak; input $0.30 at peak | Verify |
| GPT-6 Luna | OpenAI | $0.10 | $0.01 | $0.50 | Above 272K input: $0.20 / $0.75 | Verify |
| Claude Haiku 5.5 | Anthropic | $0.10 | $0.01 | $0.50 | Prompts over 100K: $0.50 / $2.50 | Verify |
Not in the table: Reflection Beam has no published per-token price, and Muse Spark Open is open weights you run on your own hardware. OpenAI's Decisions API ($0.10 per 1M input, free output) and TypeSafe's Jev ($0.042 per 1M input, free output) are decision models, not chat models.
Is your provider really serving that model?
A cheap reseller price is only a saving if you get the model you pay for. Rows marked Verify are in the reference library of the Model Fingerprint Lab: send the same random-number prompt to your API three times, paste the replies, and it shows which of 68 reference models the pattern matches. Only the numbers leave your browser, and nothing is stored.
Reading the prices
- Cost per task beats cost per token. A model that needs three times the tokens to finish a job costs three times as much per job. Artificial Analysis puts Kimi K3.1 Pro at about $1.98 per index task and MiniMax M3.1 Flash at $0.22 per agent task.
- Caching changes the order. Cache hits cost 90 to 98 percent less than fresh input on most of these models. A long agent loop with a stable prompt prefix can cost a fraction of the headline rate; try the cached-agent preset above.
- Watch the context line. GPT-6 models switch the whole request to higher rates above 272K input tokens, and Claude Haiku 5.5 above 100K. The calculator uses the short-context rate.
- Discounts end. Gemini 4 Argon goes from $2 / $10 to $4 / $20 on December 1, 2026. Mercury 3, LongCat 2.5 and Tencent Hy4 are on discounts with no published end date.
Frequently asked questions
What is the cheapest LLM API right now?
Among the models on this page, GPT-6 Luna and Claude Haiku 5.5 tie at $0.10 per 1M input tokens and $0.50 per 1M output. MiniMax M3.1 Flash and DeepSeek V4.1-Flash-1017 (off-peak) follow at $0.15 input and $0.60 output. The cheapest model for you depends on how many tokens it needs to finish the task, so compare cost per task, not just per token.
What is the most expensive?
GPT-6 Cyber, at $20 per 1M input tokens and $100 per 1M output, with restricted access. Among generally available models, GPT-6 Astra ($10 / $50) and Claude Fable 5.2 ($8 / $40) are the most expensive.
Which prices on this page are temporary?
Gemini 4 Argon's $2 / $10 is an introductory price that doubles to $4 / $20 on December 1, 2026. Mercury 3 is 50% off at launch, LongCat 2.5 is on a limited-time discount, and Tencent Hy4 is 50% off on TokenHub and OpenRouter. The table shows the price you pay today and flags each discount.
How do I know my API provider is really serving the model it bills for?
Use the Model Fingerprint Lab. You send the same random-number prompt to the API three times and paste the replies; it compares the pattern against 68 reference models. Rows marked in the table are in its reference library.
Why are Reflection Beam and Muse Spark Open not in the table?
Reflection AI has not published a per-token price for Beam. Muse Spark Open is an open-weight model you run yourself, so its cost is hardware, not tokens.
Sources. Each linked model's price comes from its guide on this site, which lists the vendor announcement and coverage it was checked against. For the rest: Gemini 4 Argon: Reuters; Gemini 4 Carbon: Reuters; Kimi K3.1 Pro: Artificial Analysis; Muse Wonder: Seoul Economic Daily / Reuters; Reflection Beam: Reflection AI; Muse Spark Open: Hugging Face model card.
Prices change often. This table was last updated on November 11, 2026. Nothing here is affiliated with any model maker.