Kimi K2 open model pricing undercuts GPT-5.5 and Claude by up to 12x
A Chinese AI lab is changing the price of smart software. Moonshot AI made a new tool called Kimi K2. It is much cheaper than its big rivals. By one count, it costs up to 12 times less than a top Claude model. And it is still good at writing computer code. Here is what that means in simple words.
First, a few easy terms. An “AI model” is the software brain that answers questions or writes code. An “open model” means the maker shares the model’s files. So other people can download it and run it on their own computers. “Tokens” are tiny pieces of text the model reads and writes. One word is often one or two tokens. “Price per token” is what you pay for each chunk of text the model handles. A “benchmark” is a test that scores how good a model is at a job, like coding.
What Moonshot AI launched
Moonshot AI released Kimi K2 on June 13, 2026. The newest version is called K2.7 Code. It is an “open-weights” model. That means the model files are free to download. You can get them from Hugging Face. Hugging Face is a popular website where people share AI models.
It uses a design called Mixture-of-Experts. Here is the easy version. The model has many small “expert” parts. But only a few wake up for each job. So it is big overall, yet cheaper to run.
“Parameters” are the internal numbers a model learns from. The full model holds 1 trillion parameters. But only 32 billion of them are active for each token. It picks 8 experts out of 384 each time.
It can read and understand text, images, and video. The “context window” is how much text it can hold in mind at once. Kimi K2’s window is 256,000 tokens. That is about a long book’s worth of words in one go.
The price gap explained
AI prices are shown per million tokens. This is often written as MTok. Kimi K2 charges $0.95 to read a million tokens. It charges $4.00 to write a million. Cached input is text the model has seen before. That costs just $0.19 per million.
The “up to 12x” claim comes from the output price. A reported Claude model charges about $50 per million output tokens. Kimi K2 charges only $4. That is about 12 times cheaper. GPT-5.5 is reported at $5 to read and $30 to write per million. So Kimi K2 is far cheaper than that too.
Benchmarks & specs
The table below shows reported figures only. One note: Moonshot’s coding scores come from its own tests. They are not checked by independent outside reviewers yet.
| Spec / benchmark | Kimi K2 (K2.7 Code) | GPT-5.5 | Claude (Opus 4.8 / Fable 5) |
|---|---|---|---|
| Input price (per million tokens) | $0.95 | $5 | $5 (Opus 4.8) |
| Output price (per million tokens) | $4.00 | $30 | $25 (Opus 4.8); $50 (Fable 5) |
| Context window | 256,000 tokens | Not reported here | Not reported here |
| Modalities | Text, images, video | Not reported here | Not reported here |
| Model size | 1T total, 32B active (MoE) | Not reported here | Not reported here |
| Program Bench | 53.6 | 69.1 | Not reported here |
| Kimi Code Bench v2 | 62.0 | 69.0 | Not reported here |
| MCPMark Verified | 81.1 | 92.9 | 76.4 (Opus 4.8) |
What it means: Kimi K2 is behind GPT-5.5 on standard coding tests. But it beats Claude Opus 4.8 on one agent-style task (MCPMark Verified). And it costs a small fraction of the price. So you get “good enough” coding for much less money.
Key facts
| Item | Detail |
|---|---|
| Maker | Moonshot AI |
| Model | Kimi K2 (K2.7 Code version) |
| Release date | June 13, 2026 |
| Type | Open-weights (free to download) |
| Where to get it | Hugging Face |
| License | Modified MIT, with a clause for very large users |
| Input / output price | $0.95 / $4.00 per million tokens |
| Context window | 256,000 tokens |
About the license
Kimi K2 uses a modified MIT license. A “license” is the rule set that says how you may use the software. MIT is a common, very open rule set. It lets you use the software freely. Moonshot added one extra rule. Companies with more than 100 million monthly active users face extra terms. So do companies with over $20 million in monthly revenue. Small teams and most startups will not hit those limits.
Why it matters (especially for India / founders)
For Indian founders and students, cost is often the biggest blocker to using AI. A 12x lower price changes the math. You can build, test, and ship more without a huge bill.
The model is open, so you can also run it on your own servers. That keeps your data in your control. It can also cut costs more as you grow. For a bootstrapped team (one that runs on its own money, with no big outside funding), that freedom is valuable.
The trade-off is simple and honest. The top closed models still lead on the hardest coding tests. So think about your needs. For everyday coding help, cheaper may be smart. For the toughest jobs, the pricier models may still win.
FAQ
What does “open model” mean for Kimi K2?
It means the model files (the weights) are shared in public. Anyone can download Kimi K2 from Hugging Face. Then they can run it on their own machines.
How is Kimi K2 cheaper than GPT-5.5 and Claude?
It charges $0.95 to read and $4.00 to write per million tokens. A reported Claude model charges about $50 to write per million. So Kimi K2 is up to 12 times cheaper on output.
Is Kimi K2 better at coding than its rivals?
Not on every test. It is behind GPT-5.5 on standard coding tests. But it beats Claude Opus 4.8 on one agent-style task. Its main edge is price, not top scores.
Takeaway
Kimi K2 makes strong AI much cheaper to use. It is open, fast, and good at coding for a fraction of the cost. The big closed models still lead the hardest tests. But for many founders and builders, “cheap and good enough” is exactly the deal they need.
Source: The Decoder — Moonshot’s open model Kimi K2.7 Code undercuts GPT-5.5 and Claude.
