Kimi K2.6 vs Kimi K3
Kimi K3 is one of Moonshot AI's newest models and Kimi K2.6 an earlier one, so this is also what moving up costs or saves. Kimi K2.6 is the cheaper of the 2: for the same tokens, 3 read for each 1 written, Kimi K3 costs 3.5 times as much. Below are both price lists, what 8 everyday jobs cost a month on each, and how many tokens the same text takes on each, which can change the answer.
- For the same tokens
- Kimi K2.6 is cheaper; Kimi K3 costs 3.5 times as much
- Cheaper to read
- Kimi K2.6, $0.95 against $3
- Cheaper to write
- Kimi K2.6, $4 against $15
- Scores higher
- Kimi K3, 157 against 151
- Larger window
- Kimi K3, 1.05M against 262K tokens
Price per million tokens, side by side
In US dollars, as each maker prints them. The better figure in each row is green.
| Kimi K2.6Moonshot AI · earlier model | Kimi K3Moonshot AI | |
|---|---|---|
| InputTokens the model reads | $0.95 | $3 |
| Cached inputRead again from the cache | $0.16 | $0.30 |
| Cache writePutting text in the cache | Not listed | $3 |
| OutputTokens it writes, thinking included | $4 | $15 |
| Blend 3:13 tokens read for each 1 written | $1.7125 | $6 |
| Batch inputAnswers collected later | Not listed | Not listed |
| Batch output | Not listed | Not listed |
| Context windowTokens it holds at once | 262K | 1.05M |
| Capability scoreEpoch Capabilities Index; GPT-5 is 150 | 151 | 157 |
| Longest answerTokens it writes in 1 reply | Not listed | Not listed |
What 8 everyday jobs cost a month on each
Each job at its usual number of runs a day, for 30 days, at the standard price. The cheaper of the 2 is green; open any job in the calculator to change it.
| Job | Kimi K2.6 | Kimi K3 | Difference |
|---|---|---|---|
| A chat question300 in, 500 out, 100 a day | $6.860.23 cents a run | $25.200.84 cents a run | 3.7 times |
| A hard question, with thinking300 in, 1,400 out, 100 a day | $17.650.59 cents a run | $65.702.2 cents a run | 3.7 times |
| A long report, summarised60,000 in, 2,000 out, 20 a day | $39.006.5 cents a run | $12621 cents a run | 3.2 times |
| A 40-turn agent task2,800,000 in, 40,000 out, 10 a day | $24983 cents a run | $659$2.20 a run | 2.6 times |
| A support bot's reply4,000 in, 300 out, 2,000 a day | $2050.34 cents a run | $6661.1 cents a run | 3.2 times |
| Sort a message into a category250 in, 5 out, 50,000 a day | $3860.026 cents a run | $1,2380.083 cents a run | 3.2 times |
| Read an invoice into fields1,500 in, 250 out, 1,000 a day | $72.750.24 cents a run | $2480.83 cents a run | 3.4 times |
| Review a pull request25,000 in, 4,500 out, 50 a day | $62.634.2 cents a run | $21414 cents a run | 3.4 times |
Your own job on both
Start from a job, then set its size, how often it runs and whether the answers can wait. Both bills follow every change.
A short question and a few paragraphs back.
about 230 words
Instructions or documents the model has read before cost far less the second time.
about 380 words
A month of this job
- Kimi K2.6$6.86
0.23 cents a run
- Kimi K3$25.20
0.84 cents a run
Kimi K2.6 saves $18.35 a month; Kimi K3 costs 3.7 times as much.
This job on every modelWhy the price per token is only the start
Every maker prints 2 main prices: one for the tokens a model reads (input) and one for the tokens it writes (output). On most current models writing costs 3 to 5 times as much as reading, so which model is cheaper can depend on the job. A job that reads a long document and writes a line pays mostly the input price; a chat that writes paragraphs pays mostly the output price. That is why each comparison prices 8 jobs and not one figure.
The same text is not the same number of tokens on every model. Each maker cuts text its own way, and Anthropic says its models from Claude 4.7 on produce about 30 percent more tokens for the same text than earlier ones. A model with the lower price per token can still send the larger bill, so each comparison counts one of our lessons the way both makers count it.
Discounts change the order too. Work that can wait costs half on a batch at the makers that offer one, and text read again from a cache costs a quarter of the input price or less on every current model here. The price table on each comparison shows who offers what.
Thinking is billed as output, and models differ in how much they think and how long they answer. A price list cannot show that. Before you switch, run the same 10 tasks on both models and compare the bills.
Questions
- Which is cheaper, Kimi K2.6 or Kimi K3?
- Kimi K2.6 is cheaper: $0.95 per million tokens read and $4 per million written, against $3 and $15 on Kimi K3. On a mix of 3 tokens read for each 1 written, Kimi K3 costs 3.5 times as much.
- Which is better, Kimi K2.6 or Kimi K3?
- On the Epoch Capabilities Index, an independent score from Epoch AI that combines over 50 benchmarks, Kimi K3 scores 157 and Kimi K2.6 151, above the top of the range of doubt around Kimi K2.6's score (GPT-5 is 150). One score is an average; test both on your own work before you switch.
- What does a support bot cost a month on Kimi K2.6 and Kimi K3?
- At 2,000 replies a day, each reading 4,000 tokens (50 percent of them from cache) and writing 300, it costs $205 a month on Kimi K2.6 and $666 on Kimi K3: Kimi K3 costs 3.2 times as much for this job.
- What does an AI agent task cost on Kimi K2.6 and Kimi K3?
- One 40-turn task that reads 2,800,000 tokens in all, 90 percent of them from cache, and writes 40,000, costs 83 cents on Kimi K2.6 and $2.20 on Kimi K3. At 10 tasks a day that is $249 and $659 a month.
- Which has the larger context window, Kimi K2.6 or Kimi K3?
- Kimi K3, with 1.05M tokens against 262K on Kimi K2.6. The window holds your instructions, the files you attach, the conversation so far and the answer, all at once.
- Does the same text take the same number of tokens on Kimi K2.6 and Kimi K3?
- Yes, both are counted the same way: our lesson "How AI writes", 1,700 words long, is about 2,100 tokens on each.
- Is it cheaper to move from Kimi K2.6 to Kimi K3?
- No, it costs more per token: Kimi K3 charges $3 per million tokens read and $15 written, against $0.95 and $4 on Kimi K2.6, 3.5 times as much for the same tokens. It pays off only if better answers save more than that, for example in fewer retries.
- Can I pay less on Kimi K2.6 and Kimi K3 for work that can wait?
- Not with a batch price: neither maker's price page lists one for these models. A cache still lowers the price of text the model reads again.
Sources