Kimi K3.1 Pro at a Third Less Cost: Specs, Pricing, Open-Weights Timeline, and Benchmarks
Moonshot AI shipped Kimi K3.1 Pro on October 23 — a heavily post-trained version of Kimi K3 that cuts API pricing by roughly a third, claims near-frontier coding-agent results, and is promised as open weights by November 4.
Key takeaways
- Released: October 23, 2026 on Kimi.com, Kimi Work, Kimi Code, and the Kimi API.
- Architecture: unchanged from Kimi K3 — 2.8T total parameters (sparse MoE), 104B active per token, 1M-token context, text + image input.
- API pricing: $2.00 / $0.20 (cached input) / $10.00 per million tokens (input / cached / output). Kimi K3 was $3.00 / $0.30 / $15.00.
- Highspeed variant: exactly 2× the price for a claimed 2–3× faster serving rate.
- Open weights: due by November 4, 2026; license terms not yet announced.
- Benchmarks: vendor-reported only so far — within about a point of Claude Opus 5.5 on Terminal-Bench 4.0.
Specs at a glance
| Released | October 23, 2026 |
|---|---|
| Model IDs | kimi-k3.1-pro, kimi-k3.1-pro-highspeed |
| Total parameters | 2.8 trillion (sparse Mixture-of-Experts) |
| Active parameters | 104B per token |
| Context window | 1M tokens |
| Input modalities | Text and images |
| Relationship to K3 | Post-trained version; identical architecture |
| Open weights | Due by November 4, 2026 (license TBD) |
| API compatibility | OpenAI-compatible endpoints |
Moonshot attributes the capability gains entirely to post-training: reinforcement learning at much larger scale across hundreds of domains, supported by new work on the RL framework, long-context inference, and sandbox infrastructure. Because the base model is unchanged, any deployment planning you did for K3 carries over directly.
Pricing: the flagship got cheaper
Per-million-token rates on Moonshot's API:
| Model | Input | Cached input | Output |
|---|---|---|---|
| Kimi K3.1 Pro | $2.00 | $0.20 | $10.00 |
| Kimi K3.1 Pro Highspeed | $4.00 | $0.20 | $20.00 |
| Kimi K3 (previous) | $3.00 | $0.30 | $15.00 |
Cache writes are billed separately on K3-series models: $2.00 per million for a five-minute cache and $4.00 per million for a one-hour cache on K3.1 Pro (double that on Highspeed). Cached reads are billed at the cache-hit rate only.
Benchmark claims — read with caveats
All figures below come from Moonshot's own launch materials and harness; independent replication has not happened yet. The rows most relevant to engineering teams:
| Benchmark | K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra | DeepSeek V4.1 Flash | GLM 5.3 |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 | 65.7 | 66.4 | 58.2 | 42.3 | 41.8 |
| DeepSWE v1.1 | 75.5 | 74.2 | 74.1 | 77.0 | 66.9 |
| SWE-Marathon v1.1 | 51.7 | 46.3 | 42.5 | 40.0 | 42.5 |
| HLE-Full (with tools) | 70.0 | 67.7 | 57.2 | 58.8 | 62.5 |
| GPQA Diamond | 94.2 | 93.9 | 96.1 | 92.2 | 91.7 |
| Artificial Analysis Intelligence Index* | 55 | 58 | — | — | 45 |
* The Intelligence Index is the only independently run line in the table. Moonshot reports an 11-point jump there versus Kimi K3 — the clearest signal in the set. Comparison rows mix vendor-reported scores, so treat cross-model deltas as indicative, not definitive.
Two additional vendor claims worth tracking: Moonshot says the claim-level hallucination rate on its internal test suite dropped to 8.3% from 23.4% for K3, and that cyber-related data was identified and removed from mid-training and the RL mixture ahead of the open release — a deliberate safety trade that costs some vulnerability-research capability.
Self-hosting: the hardware math is unchanged
Since K3.1 Pro is a post-trained K3 with identical parameter counts, the deployment footprint is the same: if the checkpoint ships in the same MXFP4 format, expect roughly 1.56 TB of weights on disk and a multi-node cluster with 2 TB+ of aggregate GPU memory plus fast interconnect. Unless you already operate such a cluster, the API remains the practical access route — and the 104B active parameters are precisely what keeps per-token pricing where it is.
What a realistic workload costs
Assuming 10M input and 2M output tokens per day, no caching, 30 days, at list prices:
| Model (illustrative route) | Input / output per 1M | Per day | Per month |
|---|---|---|---|
| DeepSeek V4 Pro | $0.99 / $2.97 | $15.84 | $475.20 |
| GLM 5.3 | $1.40 / $4.40 | $22.80 | $684.00 |
| Kimi K3.1 Pro | $2.00 / $10.00 | $40.00 | $1,200.00 |
| Kimi K3 | $3.00 / $15.00 | $60.00 | $1,800.00 |
| GPT-6 Astra | $10.00 / $50.00 | $200.00 | $6,000.00 |
K3.1 Pro now sits at a fifth of a GPT-6 Astra bill while claiming near-Opus coding-agent performance — yet it still costs well over twice a tier-below open flagship like DeepSeek V4 Pro. The familiar question stands: does the capability difference show up in your output quality? Benchmark it on your own workload before switching.
How to try it today
Call it on Moonshot's API with model ID kimi-k3.1-pro — or kimi-k3.1-pro-highspeed when latency matters more than cost. The endpoint is OpenAI-compatible, so an existing client typically needs only a base-URL and model-string change. When the weights land on November 4, the self-hosting route opens for teams with the hardware — watch for the model card and license terms, which will define what "open weight" actually permits.
- Is Kimi K3.1 Pro open source?
- Moonshot calls it open weight and promises full weights by November 4, 2026. The license is unannounced; the previous K3 shipped under a custom (non-OSI) license, so review the terms before building products on it.
- What is the difference between Pro and Pro Highspeed?
- The same model, served at a claimed 2–3× faster speed for exactly double the price.
- Is it really as good as Claude Opus 5.5?
- On Moonshot's own numbers, it is within about a point on Terminal-Bench 4.0 and ahead on two long-horizon SWE benchmarks; on the independent Intelligence Index it trails by three points. Nobody has replicated the vendor figures yet.
- How much VRAM does self-hosting require?
- The same as Kimi K3 if the checkpoint format is unchanged: about 1.56 TB of weights and a 2 TB+ aggregate GPU-memory cluster.
Bottom line
K3.1 Pro is the same 2.8T-parameter model as K3 with far more post-training, priced a third lower, and carrying unverified frontier-level coding-agent claims. The price cut signals Moonshot is competing on unit economics, not just capability. For most teams the playbook is unchanged: call it through an API, measure it against the open flagships you already run, and revisit when the weights and license actually drop.
Sources: Moonshot AI launch announcement (Oct 23, 2026) and public coverage. All benchmark and hallucination figures are vendor-reported and not independently verified. Prices are list prices at time of writing and subject to change.