Moonshot AI released Kimi K3.1 Pro on October 23, 2026, and the headline number is the price: $2 per million input tokens and $10 per million output tokens, a third cheaper than the Kimi K3 it upgrades, with a 1M-token context window and no extra charge for long prompts. It is a post-trained upgrade of K3 rather than a brand-new base model (Source: Codersera).
The launch comes with two kinds of scorecards. Moonshot published its own benchmark results, and independent evaluators measured the model separately within a day or two. They mostly agree, but the details matter. This post separates vendor-reported figures from independent ones, explains the Highspeed tier, and flags the one promise worth verifying: Moonshot said downloadable weights would follow by November 4, 2026.
Quick summary: Kimi K3.1 Pro (Moonshot AI, October 23, 2026) costs $2.00 input, $10.00 output and $0.20 cached input per million tokens, with a 1,048,576-token context window. The Highspeed tier is $4 and $20 for the same model at a claimed 2 to 3 times the speed. K3 was $3 and $15. Vendor-reported: Terminal-Bench 4.0 at 65.7%. Independent (Artificial Analysis, as relayed by Codersera and AICYBR): Intelligence Index 55, Terminal-Bench 4.0 at 67%. Open weights were promised by November 4, 2026, but were not public as of October 26, and we could not confirm a release as of November 3, the day before the deadline.
What launched, and when
LLM Gateway's release timeline lists both Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed on October 23, 2026, tagged as flagship models (Source: LLM Gateway). Codersera attributes the release date to Artificial Analysis and gives the API model IDs as kimi-k3.1-pro and kimi-k3.1-pro-highspeed (Source: Codersera). The model takes text, image and video input and produces text output. Reasoning is always on: thinking cannot be switched off, though a reasoning_effort setting accepts low, high or max, with max as the default (Source: Codersera).
Size is reported as roughly 2.8 trillion total parameters with about 104 billion active per token. That figure comes from Artificial Analysis and from the K3 model card, as relayed by the write-ups. Codersera notes Moonshot has not published a separate architecture table for K3.1 Pro and calls the model '3T-class' (Source: Codersera). Treat the parameter counts as secondhand.
Pricing: Standard vs Highspeed vs the old K3
| Model | Input per 1M | Output per 1M | Cached input per 1M |
|---|---|---|---|
| Kimi K3.1 Pro | $2.00 | $10.00 | $0.20 |
| Kimi K3.1 Pro Highspeed | $4.00 | $20.00 | $0.20 |
| Kimi K3 (predecessor) | $3.00 | $15.00 | $0.30 |
These prices are identical in both write-ups (Source: AICYBR). Standard K3.1 Pro is one-third cheaper than K3 on both input and output; Highspeed doubles the standard rates for what Moonshot says is the same model running 2 to 3 times faster. Codersera also lists cache writes at $2.00 for a five-minute cache and $4.00 for one hour, and a default maximum output of 131,072 tokens (Source: Codersera). Pricing pages change, so confirm on Moonshot's platform before budgeting.
Benchmarks: vendor-reported vs independent
Moonshot's launch post reports Terminal-Bench 4.0 at 65.7%, Kimi Code Bench 3.0 (a long-horizon coding test) at 61.2%, BenchCAD with tools at 92.7%, OfficeQA Pro v2 at 64.4, and a hallucination test called Kimi Hallu Bench Hard at an 8.3% claim-level error rate, where lower is better. These are vendor-reported, run at maximum reasoning effort, and several benchmarks are internal, with the hallucination test using deliberately hard prompts (Source: AICYBR). Codersera adds that Moonshot ran K3.1 Pro itself and took some competitor scores from leaderboards or system cards, so the results are not controlled head-to-head tests. In that comparison Moonshot lists Claude Opus 5.5 at 66.4% on Terminal-Bench 4.0, and 51.7 versus 46.3 for K3.1 Pro against Opus 5.5 on SWE-Marathon v1.1 (Source: Codersera).
Independent numbers come from Artificial Analysis. As of October 24 it gave K3.1 Pro an Intelligence Index of 55, ranking 4th of 215 models, and measured Terminal-Bench 4.0 at 67%, slightly above Moonshot's own 65.7% (Source: AICYBR). Codersera reports the same index of 55 against 58 for Opus 5.5, and a cost per Intelligence Index task of $2.11 versus $5.98 for Opus 5.5 (Source: Codersera). Speed readings vary between write-ups (AICYBR reports about 73.5 output tokens per second), so measure latency yourself.
| Metric | Kimi K3.1 Pro | Claude Opus 5.5 | Type |
|---|---|---|---|
| Terminal-Bench 4.0 | 65.7% | 66.4% | Vendor-reported (Moonshot) |
| SWE-Marathon v1.1 | 51.7 | 46.3 | Vendor-reported (Moonshot) |
| Artificial Analysis Intelligence Index | 55 | 58 | Independent |
| Cost per Intelligence Index task | $2.11 | $5.98 | Independent |
| Terminal-Bench 4.0 (measured) | 67% | Not given | Independent |
The open-weights promise
Weights were not available at launch. Moonshot scheduled downloadable K3.1 Pro weights by November 4, 2026, and as of October 26 Hugging Face had no K3.1 Pro repository and Artificial Analysis listed the weights as not public (Source: Codersera). The final files, license and serving requirements were unconfirmed (Source: AICYBR). Check Moonshot's channels for the current status before you plan around it.
Even when weights ship, running them yourself is a data-center job. AICYBR's estimates, which are theoretical rather than measured, put raw weights near 5.6 TB in BF16, 2.8 TB at 8-bit, 1.4 TB at 4-bit and 700 GB at 2-bit, and say four 80 GB GPUs are not enough even for the 4-bit estimate. The 104B active parameters do not mean a 104B-sized download (Source: AICYBR). For nearly everyone, the practical way to use K3.1 Pro is the API.
- Cheaper than K3 by one-third on input and output; Highspeed costs double the standard tier.
- Flat pricing across the full 1M context, per Codersera, so long documents do not trigger a higher tier.
- Vendor-reported benchmarks are competitive with Opus 5.5 on some coding tests, independent index is a few points lower, and cost per task is reported much lower.
- Weights promised by November 4; verify before relying on self-hosting or local use.
Before switching a workflow to a cheaper model, run 20 of your own real tasks through both models and compare cost per finished task, not price per token. A model with a lower sticker price that needs more retries can cost more.
FAQ
01How much does Kimi K3.1 Pro cost?
$2.00 per million input tokens, $10.00 per million output tokens and $0.20 per million cached input tokens. Highspeed is $4.00 and $20.00.
02What is the context window?
1,048,576 tokens (about 1M), with flat pricing across the window, per Codersera.
03Is Kimi K3.1 Pro better than Claude Opus 5.5?
Mixed. Moonshot reports 65.7% vs 66.4% on Terminal-Bench 4.0 and a lead on SWE-Marathon (vendor-reported). Artificial Analysis gives an Intelligence Index of 55 vs 58 but a much lower cost per task, $2.11 vs $5.98.
04Are the weights open?
Moonshot promised downloadable weights by November 4, 2026. As of October 26 they were not public, and we could not confirm a release as of November 3, so verify the current status.
05What is the Highspeed tier?
The same model at double the price ($4 input, $20 output), which Moonshot says runs 2 to 3 times faster.
Price and benchmark tables only go so far; the real test is your own prompts. LumiChats offers 40+ premium AI models side by side in one subscription, so you can run the same task on a budget model and a flagship and judge the difference yourself.
