Kimi K3.1 Pro is Moonshot AI's flagship reasoning model, released as the first model in its new K3.1 series and positioned as a sparse Mixture-of-Experts system with a long context window. It ranks 4th on the Artificial Analysis Intelligence Index, scoring 55 points, which makes it the highest-ranked model expected to ship with open weights. It is described as a 3T-class model with a context window reported in the million-token range, priced at $2 per million input tokens and $10 per million output tokens on its standard API tier.
K3.1 Pro is a post-trained version of Kimi K3, built with larger post-training compute and a full-stack optimization of its reinforcement learning framework, long-context inference, and sandbox infrastructure. Moonshot positions it as its most capable open-weight model to date, retaining text and image input with text output while focusing on agentic coding, long-horizon execution, and tool use. Independent testing describes an MoE layout where only a routed subset of parameters runs per forward pass, which underpins its cost and speed profile.
The launch ranking places it behind only Claude Opus 5.5, Claude Sonnet 5.5, and Grok Wonder Full on the Intelligence Index, narrowing the gap between leading proprietary and open-weight models once weights are public. Moonshot has said full model weights for K3.1 Pro are expected to follow, with a faster companion variant, K3.1 Pro Highspeed, launched alongside the standard release.
The headline numbers matter most once you understand what sits underneath them architecturally.
Kimi K3.1 Pro delivers its capacity without paying full compute cost on every token by activating only a routed subset of its experts per token, keeping active compute far below the model's total parameter count. It uses a sparse Mixture-of-Experts layout built for long-context reasoning, where the full model holds a large pool of total parameters but only a routed subset runs for each token.
In a dense model, every parameter participates in every forward pass. In Kimi K3's sparse MoE design, a large bank of experts is stored in memory and a router selects a small active subset per token, so FLOPs per token scale with the active slice, not the total. Moonshot's published spec positions the model as a large total-parameter system with a far smaller activated slice per token, which is how the system keeps inference work close to a much smaller dense model while retaining greater overall capacity.
Practically that splits two costs. Memory and storage still track total parameters because all expert weights must be available across shards, while compute, latency, and token price scale with active parameters plus routing overhead. The quality argument for this tradeoff is specialization: different experts can focus on code, reasoning, or multimodal patterns, and the router composes them per token rather than forcing one shared weight matrix to handle everything.
The structure also changes how you run it. The architecture notes describe 93 layers mixing Kimi Delta Attention with Gated MLA and Attention Residuals to preserve signal across depth, a design choice aimed at making very long sequences stable during decoding without collapsing attention cost.
Kimi K3.1 Pro carries a context window reported in the million-token range. That is not a marketing limit for chunked retrieval. It fits in one prompt, which unlocks three workloads that break under 128K or 200K windows:
For developers, the implication is more KV-cache to manage but fewer tricks to work around truncation. Long-context handling becomes a systems problem (caching, eviction, and resume) rather than a prompt compression problem.
Total parameters is how many weights the model stores; active parameters is how many actually run per token in an MoE. Kimi K3's exact split has not been independently confirmed, but the gap between the two figures is what keeps inference cost close to a much smaller dense model.
Kimi K3.1 Pro does not beat Claude Opus 5.5 on the Artificial Analysis Intelligence Index. It scores 55 points for 4th place as of October 23, 2026, three points behind Claude Opus 5.5 at 58 points.
The common mistake is reading that ranking as a wide capability gap. On this Index, a 3-point spread at the top reflects specific eval tradeoffs rather than overall dominance.
With the architecture established, the real question is how it measures up against the models it's competing with for frontier status. Artificial Analysis confirmed the new order: Claude Opus 5.5 (Max with fallback, 58) holds #1, followed by Claude Sonnet 5.5 (Max with fallback, 56) and Grok Wonder Full (Max), with Kimi K3.1 Pro next at 55, an 11-point gain over Kimi K3 at 44.
That score comes from the v4.3.2 Index methodology, which blends 10 evaluations including GDPval-AA, Terminal-Bench 4.0, AutomationBench-AA, and Humanity's Last Exam. To keep the comparison honest, the table below uses only values observed directly from Artificial Analysis, limited to the models it scored at launch:
| Model | Artificial Analysis Index Score/Rank | Context Window | Parameter Architecture | Pricing (per 1M tokens, in/out) | Release / Availability |
|---|---|---|---|---|---|
| Kimi K3.1 Pro | 55 / #4 | Long-context (reported in the million-token range) | Sparse MoE, 3T-class | $2.00 / $10.00 | Live on Moonshot's platform; open weights expected to follow |
| Claude Opus 5.5 (Max, Default Fallback) | 58 / #1 | — | — | Output pricing varies — check Anthropic's official page | Closed, API-available |
| Claude Sonnet 5.5 (Max, Default Fallback) | 56 / #2 | — | — | — | Closed, API-available |
| Grok Wonder Full (Max) | Scored alongside the above on the same Index | — | — | — | Closed, API-available |
The table above reflects only the figures Artificial Analysis and Moonshot have published; several comparison cells are intentionally left blank rather than estimated.
In practice, the gap to #1 is narrowest where Moonshot focused. On GDPval-AA, Kimi K3.1 Pro reaches 1,846 Elo, just 20 Elo behind Claude Opus 5.5 at 1,866 Elo and ahead of its earlier generation by 312 Elo. On Terminal-Bench 4.0, Kimi K3.1 Pro scores 67%, up from 13% for Kimi K3 and above Claude Sonnet 5.5 at 64%, which Artificial Analysis notes as the highest Terminal-Bench result in the current set.
The remaining three points reflect consistent, smaller deltas on knowledge and factuality evals like Humanity's Last Exam and CritPt, where gains were moderate, rather than any collapse in reasoning. For procurement, that means the choice between #1 and #4 is less about general intelligence rank and more about deployment specifics: Kimi K3.1 Pro offers a long context window under a sparse MoE architecture, priced at $2 per 1M input tokens, $10 per 1M output tokens, and $0.20 per 1M cache-hit tokens.
Ranking aside, Moonshot shipped two versions of this model on day one, and the difference matters for anyone choosing between them.
Kimi K3.1 Pro Highspeed costs $4.00 per million input tokens and $20.00 per million output tokens, double the standard Kimi K3.1 Pro rate published October 23, 2026, and delivers two to three times faster output speed with the same underlying model.
The distinction is infrastructure, not capability. Moonshot AI lists both variants as kimi-k3.1-pro and kimi-k3.1-pro-highspeed on its own platform, and describes Highspeed as the same model run on a faster serving tier, with roughly twice the quota consumption in Kimi Code, priced for speed over standard. Publisher OrcaRouter reports the same finding from Moonshot's documentation, noting identical benchmark behavior with no model change. There is no evidence in the release notes of a distilled or degraded checkpoint for Highspeed; the tradeoff is price per token and quota burn for latency, not accuracy.
On pricing, Moonshot published a flat rate card across the full window with no long-context surcharge:
That puts the new standard variant cheaper than the model it replaces. A 1M-input / 200K-output request costs less on K3.1 Pro than it did on K3 before caching, and the cache-hit price itself fell versus the predecessor, which preserves the discount for agentic loops that re-read large prompts. The pricing also positions K3.1 Pro below some closed frontier outputs at launch; third-party roundups noted Claude Opus 5.5's output pricing sits above K3.1 Pro's $10 rate, though Anthropic's own page should be checked for the current figure.
For developers budgeting agentic or high-volume workloads, the practical effect is twofold. Standard Pro lowers the ceiling on per-task cost for long-horizon agents that accumulate tool outputs and document context, thanks to the flat window and cheaper cache hits. Highspeed doubles token cost but is aimed at interactive coding and rapid iteration where wait time dominates, with Moonshot stating the model always thinks and defaults reasoning effort to max. Changing reasoning effort mid-conversation invalidates prefix-cache hits, which can quietly raise effective cost if you tune prompts inside a long session.
The second milestone that matters is the open-weights release. Moonshot has indicated full model weights are coming, which would make K3.1 Pro open-weight, not just API-only. Once delivered, self-hosting removes the per-token meter for teams with accelerator capacity, enables private fine-tuning on internal code or finance corpora, and shifts competitive comparison toward other open-weight frontier releases such as DeepSeek. Self-hosting at this scale will still require significant GPU memory and serving infrastructure, so managed API and self-hosted routes will likely coexist: API for elasticity and zero data-retention options, weights for control and custom post-training.
Kimi K3.1 Pro is live across Moonshot AI's full product surface as of October 23, 2026, from Kimi.com and mobile apps to Kimi Work desktop, Kimi Code, and the developer API, built to tackle complex coding projects, financial research, and advanced agentic work.
Specs and pricing settle the 'what'; the remaining question is where this model actually shows up and who should care.
Practical availability is broad and launched in parallel. Moonshot lists K3.1 Pro as available today across Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, with full model weights expected in a near-term follow-up release. If you are deciding where to try it, the channels map to different workflows:
Moonshot's stated target use cases are consistent across its launch note and product page: covering the full range of coding work from quick fixes to full-project builds, enabling end-to-end delivery of financial workflows, and running agentic systems that search, execute code, and verify outputs. For teams already maintaining large codebases, research tasks that require cross-document verification, or internal agents that need to keep context across hours of tool calls, those are the surfaces worth testing first.
Who should evaluate now: teams already budgeting for frontier models and willing to run on a hosted stack, developers building coding agents or 3D asset pipelines where long context and tool use are the bottleneck, and financial research teams that need auditable, source-grounded outputs in the same session. If your stack depends on self-hosting, custom fine-tuning, strict data residency, or on-prem inference today, wait for the open-weights drop and watch how licensing, safety filtering, and infrastructure requirements are defined.
For everyone else, the decision is timing. Trial K3.1 Pro now on Kimi Code and the API against your current flagship for your hardest agentic task, measure failure modes, not just headline scores, and hold the build-vs-wait call until you see whether the open-weights release arrives on schedule and what it changes for deployment control.
Kimi K3.1 Pro makes most sense today for long-horizon coding and agentic research teams working on Moonshot's hosted stack. The biggest open question is whether the promised open-weights release holds to its timeline.
HarperFlow publishes highly structured articles featuring FAQs, data tables, and direct-answer blocks that meet the rigorous citation standards required by AI search engines. By continuously auditing and improving your content through AI answer analytics, HarperFlow helps your site build long-term authority and visibility that outlasts ad-dependent strategies.
Yes, it is reported as 2.8T total parameters with 104B active per token per Artificial Analysis. You must store and shard the full 2.8T across hardware, but compute per token tracks the 104B active slice, so API cost behaves like a much smaller dense model.
Pick kimi-k3.1-pro-highspeed when latency matters more than token unit cost. Moonshot says Highspeed delivers two to three times the output speed with identical weights, but pricing doubles to $4 input and $20 output and quota burn in Kimi Code is roughly twice standard.
No, Moonshot published a flat rate card across the full window. The standard rate of $2 input, $10 output and $0.20 cache hit applies up to 1M tokens, so long-document and large-codebase prompts scale linearly rather than with a surcharge.
In Kimi Code CLI and editor integration, use the /model command and choose kimi-k3.1-pro or kimi-k3.1-pro-highspeed. Moonshot lists it as available today on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Today the API meters tokens and offers Zero Data Retention for customers. Full model weights expected by Nov 4, 2026 remove the per-token meter and allow private fine-tuning and on-prem control, but hosting a 2.8T MoE will still require substantial accelerator memory and serving infrastructure.
The Index blends 10 evals, and K3.1 Pro is only 20 Elo behind Claude Opus 5.5 on GDPval-AA at 1,846 vs 1,866, while leading on Terminal-Bench 4.0 at 67%. The remaining points reflect smaller deltas on knowledge evals like Humanity's Last Exam, not a broad reasoning failure.
Cache-hit tokens cost $0.20 per 1M, with cache writes at $2.00 at five-minute TTL and $4.00 at one-hour TTL for standard Pro. Changing reasoning effort mid-conversation invalidates prefix-cache hits, which can raise effective cost for long agentic sessions that re-read large prompts.
Yes, Zero Data Retention is available for API customers using kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Enterprise access through Kimi Business is off by default at launch and must be enabled by admins.
