Kimi K3.1 Pro joins Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon at $2 in and $10 out. Its cache hits cost double, and it spends 2.2x Sol's tokens per task, so the matching price tag hides a 3x spread.
Three months ago Moonshot tripled its prices with Kimi K3. On Friday it took a third back. K3.1 Pro lists at $2 in and $10 out, which is now the most crowded price point in the market.
Photo by Raymond Petrik on Unsplash
Two things the $2/$10 tag hides
- Kimi charges double for a cache hit but nothing extra for a cache write. Cache under about 83% of your input and it beats Sonnet 5.5 and Sol on the token bill; cache more and it loses.
- It writes 2.2x Sol's output tokens on Artificial Analysis, so $2.11 a task against $0.72 at identical rates. Against Sonnet 5.5 at max it's one point behind for under 40% of the cost.
Four cards at $2 and $10, and the ones above them
| Model | Input | Cache write (5m) | Cache hit | Output |
|---|---|---|---|---|
| Kimi K3.1 Pro | $2.00 | $2.00 | $0.20 | $10.00 |
| Claude Sonnet 5.5 | $2.00 | $2.50 | $0.10 | $10.00 |
| GPT-6.1 Sol (to 272K) | $2.00 | $2.50 | $0.10 | $10.00 |
| Gemini 4 Argon (to Nov 30) | $2.00 | none | $0.10 | $10.00 |
| Kimi K3 | $3.00 | $3.00 | $0.30 | $15.00 |
| Kimi K3.1 Pro Highspeed | $4.00 | $4.00 | $0.20 | $20.00 |
| Gemini 4 Argon (from Dec 1) | $4.00 | none | $0.20 | $20.00 |
| Claude Opus 5.5 | $4.00 | $5.00 | $0.20 | $20.00 |
Per million tokens, read October 24, 2026. Kimi's own price table rendered empty when we checked, so the Kimi rows come from OpenRouter, Vercel AI Gateway and Artificial Analysis, which agree with each other. Argon uses implicit caching only, so there is no write fee. Sol bills $4/$15 on requests over 272K input tokens.
The model ID is kimi-k3.1-pro. Moonshot describes it as a post-trained K3, so it keeps the same 1,048,576-token context and flat billing across it. Max output defaults to 131,072 and can be raised to the full window. Caching is automatic with a 5-minute TTL, and you can ask for an hour.
K3 hasn't been retired. Moonshot's docs call it the previous-generation flagship and still list it at $3 and $15, so anyone pinned to kimi-k3 keeps paying 50% more until they change one string. We covered K3's launch price in July.
The cache line decides the tie
Input and output are identical across the $2 group, so the only lines that move a bill are caching. We priced an agent month: 1 billion input tokens, 85% served from cache, and 50 million output tokens. Uncached input is billed as a 5-minute cache write wherever the provider charges for one.
- Gemini 4 Argon (to Nov 30)$885
- Claude Sonnet 5.5$960
- GPT-6.1 Sol$960
- Kimi K3.1 Pro$970
- Kimi K3$1,455
- Kimi K3.1 Pro Highspeed$1,770
- Gemini 4 Argon (from Dec 1)$1,770
- Claude Opus 5.5$1,920
Kimi K3.1 Pro: 150M uncached at $2.00 ($300), 850M cache hits at $0.20 ($170), 50M output at $10.00 ($500). Sonnet 5.5 and Sol: 150M at $2.50 ($375), 850M at $0.10 ($85), output $500. Prompts assumed under Sol's 272K threshold.
At 85% hits, Kimi lands $10 above Sonnet and Sol. That's noise. What matters is which side of 83.3% you sit on. Kimi saves $0.50 per million on every write and loses $0.10 per million on every hit, so the two cancel at five hits per write. Drop to 70% hits, which is normal for a chat app with short sessions, and Kimi comes out $80 cheaper. Push to 95%, typical of a long-running coding agent, and it costs $70 more.
Gemini 4 Argon is the cheapest of the four until November 30, because it never charges for writes. On December 1 every Argon line doubles and it lands on Opus 5.5's card instead. We wrote up that switch on Tuesday.
Per task, the $2 group splits three ways
A rate card tells you the price of a token. It doesn't tell you how many a model spends to finish a job. Artificial Analysis measures that directly:
| Model | Index | $ per task | Output tokens | Tok/s |
|---|---|---|---|---|
| Claude Opus 5.5 (max) | 58 | $5.98 | 260M | 96.1 |
| Claude Sonnet 5.5 (max) | 56 | $5.46 | 420M | 132.0 |
| Kimi K3.1 Pro | 55 | $2.11 | 150M | 77.1 |
| Gemini 4 Argon (high) | 53 | $1.99 | 110M | 76.4 |
| GPT-6.1 Sol (max) | 52 | $0.72 | 67M | 55.2 |
Artificial Analysis model pages, read October 24, 2026. Output tokens are the total used across the whole index run. Argon's cost uses its pre-December price.
Same price per token, nearly three times the price per task. Sol finishes the suite on 67M output tokens. K3.1 Pro uses 150M, and AA flags it as very verbose. Reasoning is always on and defaults to max effort, so that's what you get unless you turn it down. Moonshot offers low and high as well. Nobody has published scores at those settings yet.
Read the other way, the table is good news for Kimi. It scores a point under Sonnet 5.5 at max for 39% of the cost per task, and uses about a third of Sonnet's output tokens to do it. If you're running Sonnet at its top setting for quality, K3.1 Pro is the obvious thing to test against it.
It's also not fast. 77.1 tokens a second puts it level with Argon, behind Opus at 96.1 and well behind Sonnet at 132.
Highspeed costs what Opus costs
kimi-k3.1-pro-highspeed is the same model served faster, at $4 and $20. Moonshot claims 2-3x the speed and gives no tokens-per-second number. If the claim holds, that's somewhere around 150 to 230 tokens a second against Opus 5.5's measured 96.
It's the same headline price as Opus 5.5, and Opus scores three points higher. Highspeed keeps the $0.20 cache hit and charges $4 for writes against Opus's $5, so on our agent month it comes in $150 under Opus. We wouldn't pay for it until someone measures the speed. Paying double for an unmeasured claim is a bet, not a budget.
Things to know before you switch
The benchmark Moonshot leads with is its own pick. Terminal-Bench 4.0 puts K3.1 Pro at 65.7% against Opus 5.5's 66.4%, according to the launch post. Sonnet 5.5 isn't in that chart. Anthropic reports 70.6 for it at max. Moonshot's quickstart page still has "[xx.x%]" where the Terminal-Bench number should be, so the docs were clearly finished in a hurry.
Sampling is fixed at temperature 1.0 and top_p 0.95, and the docs say to leave both out of requests. If your client sends its own values by default, check that first.
Weights are promised by November 4, with no licence published. When K3's weights landed in July, no host undercut Moonshot at first, because the checkpoint needs eight B300s to boot. Three months on, OpenRouter lists K3 from Sail Research at $1.20 and $5.50, Decart at $1.65 and $8.25 and Chutes at $1.85 and $9.50, all under Moonshot's $3 and $15. If K3.1 Pro follows the same path, $2 and $10 is a ceiling, not a floor. For now Moonshot is its only provider on OpenRouter.
Moonshot's own hallucination test has K3.1 Pro at 8.3 against K3's 23.4 and Opus 5.5's 15.0, lower being better. It's an in-house benchmark, and AA hasn't posted an independent hallucination figure yet.
Where K3.1 Pro fits
K3 users should switch today. It's a third cheaper on every line and AA ranks it higher.
Anyone on GPT-6.1 Sol to keep costs down should stay put. Sol is three points behind and costs a third as much per task.
If you're paying for Sonnet 5.5 at max, or thinking about Opus, K3.1 Pro is worth a real trial on your own prompts. Check your cache hit rate first. Above 83% the token bill tilts back to Sonnet, though the per-task gap is far larger than that. Our calculator takes your own token mix.
Sources
- Kimi K3.1 Pro announcement and Kimi platform docs - benchmarks, context, caching rules, sampling, weights date, K3 status
- OpenRouter and Simon Willison - K3.1 Pro and Highspeed rates, provider list, Highspeed speed claim
- Artificial Analysis - Intelligence Index, cost per task, output tokens and speed for every model in the second table
- Anthropic pricing, OpenAI GPT-6.1 Sol and Gemini API pricing - competitor rate cards