AI prices
Half price per token isn’t half the bill
Kimi’s new model costs half as much per token as Claude Opus 5.5. It also writes about twice as many tokens, so at default settings you end up paying about the same.
Kimi K3.1 Pro charges $10 per million tokens it writes. Claude Opus 5.5 charges $20. So the same answer should cost half as much from Kimi.
It mostly doesn’t. We gave both the same everyday prompts from our internal benchmark, three times each. On 67% of them the two bills came out within 15% of each other. And on a math research question, Opus’s answer cost $0.076 and Kimi’s $0.197, 2.6× as much.
The reason is thinking. K3.1 Pro reasons at length before it answers, and those tokens are billed like the answer. Asked to fix a long function so its tests pass, it wrote 13,358 tokens on average, 12,688 of them thinking, and took 5 min 7 s. Opus wrote 2,696 and was done in 24 seconds. Kimi’s fix cost 2.3× as much.
- A math research question2.59× the cost
- Fix a long function so its tests pass2.28× the cost · 5.0× the tokens
- A river-crossing logic puzzle1.13× the cost · 2.3× the tokens
- Plan 3 days in Lisbon on $6001.10× the cost · 2.2× the tokens
- Write a Python function, with tests1.10× the cost · 2.2× the tokens
- Draw a fox, then edit it twice (SVG)1.04× the cost · 2.2× the tokens
- A polite email to the landlord0.98× the cost · 1.9× the tokens
- Draw a world map from memory (SVG)0.98× the cost · 2.0× the tokens
- Explain how a mortgage works0.58× the cost · 1.2× the tokens
Line: same cost. Token counts include thinking. The math question shows only the two costs.
Show the numbers
| Prompt | Opus: cost, tokens | Kimi: cost, tokens |
|---|---|---|
| A math research question | $0.076 | $0.197 |
| Write a Python function, with tests | $0.067, 3,308 | $0.073, 7,261 |
| Fix a long function so its tests pass | $0.059, 2,696 | $0.134, 13,358 |
| A polite email to the landlord | $0.0044, 212 | $0.0044, 412 |
| Plan 3 days in Lisbon on $600 | $0.069, 3,425 | $0.075, 7,513 |
| Explain how a mortgage works | $0.051, 2,540 | $0.030, 2,961 |
| A river-crossing logic puzzle | $0.013, 606 | $0.014, 1,381 |
| Draw a world map from memory (SVG) | $0.160, 8,005 | $0.157, 15,676 |
| Draw a fox, then edit it twice (SVG) | $0.289, 12,162 | $0.301, 26,222 |
Twice the tokens, the same bill
Opus costs 2 times as much per token, so Kimi only comes out cheaper when it writes less than 2 times as much. Across these prompts it typically wrote 2.2× Opus’s tokens, and the typical bill was about level. Kimi was clearly cheaper only once, explaining a mortgage (0.6× the cost), where it wrote little more than Opus.
Long jobs don’t change the picture. They just take longer. Drawing a world map, Kimi wrote twice as much as Opus, cost the same, and took 3 min 47 s to Opus’s 1 min 19 s.
On the big benchmarks, it flips
All of the above is at default settings. Claude also has a “max” effort setting, and there Opus is the wordier model. Running the Artificial Analysis Intelligence Index, which runs Claude at “max”, Opus 5.5 wrote 260 million tokens to Kimi’s 150 million. Running the whole index cost $8,708 on Opus and $3,117 on Kimi: $5.98 a task against $2.11. Even one notch down, at “xhigh”, Opus cost $3.46 a task. Claude Sonnet 5.5, which scores about the same as Kimi, wrote 420 million tokens and cost $7,260 at max.
Another leaderboard that publishes cost per task points the same way. On LiveBench, Kimi cost $0.74 a task to Opus’s $3.65, and Opus wrote 3.3× as many tokens as Kimi.
- Our everyday prompts, default settings (typical)1.10×
- Artificial Analysis index, Opus at “xhigh”0.61×
- Artificial Analysis index, Opus at “max”0.35×
- LiveBench0.20×
Each source runs its own tests, so compare the multiples, not the dollars. LiveBench doesn’t say which effort setting it used.
So if you run Opus at its default settings, as you should most of the time, you’re better off with Opus than with Kimi K3.1 Pro: the bill usually comes out about the same or lower, and the answer typically arrives about three times sooner. Kimi pulls clearly ahead on cost when Opus is turned up to its higher effort settings.
None of this makes K3.1 Pro a bad model. It is basically an open-source Claude Sonnet, and its weights are due out by November 4. But the price list only tells you half the bill. The other half is how much a model writes, and the only way to know is to run your own prompt.
Run your prompt on both and see what each answer really costs.
How we did it
We ran both models through the AI Gateway on 23 and 25 October at their default settings, three answers each per prompt (two for the fox, which is three turns long). Costs are at list prices, input included. Kimi’s two drawings are from our launch-day run on the same prompts. The math research question is from our internal benchmark: “How many 4×4 grids of 0s and 1s have exactly two 1s in every row and every column? Derive the count, don’t just state it.” Only its two costs are shown. The Intelligence Index token counts are from each model’s page on Artificial Analysis, and its costs from the Artificial Analysis API, fetched on 2026-10-25. LiveBench costs per task were read on 2026-10-25.