AI Tool Scout › AI Models Review

AI Models · Open-Weights LLMs · Updated October 2026

Kimi K3.1 Pro Review (2026): $2/$10, 2.8T, Weights Due Nov 4

Item: Kimi K3.1 Pro
Rating: 4.0
Author: AI Tool Scout

Moonshot made the flagship a third cheaper and eleven points smarter, then promised the checkpoint in eleven days. One of those is a strategy. The other is a calendar entry.

By AI Tool Scout · Published October 25, 2026 · 18 min read

⚡ TL;DR — Is $2 / $10 for a frontier-adjacent model a bargain, and should you move now?

Yes to the rate card, not yet to the weights. Moonshot released Kimi K3.1 Pro on Friday, October 23, 2026 — the first model of a new K3.1 series, and a post-trained K3 rather than a new architecture: same 2.8T total / 104B active MoE, same 1,048,576-token context, same native multimodal input, with substantially larger post-training compute spent on long-horizon agency, 3D work and finance. The API price is $2.00 input / $0.20 cached / $10.00 output per 1M tokens, 33% below K3 on all three rates and exactly half of Claude Opus 5.5, with cache writes at $2.00 (5-minute TTL) or $4.00 (1-hour). Kimi K3.1 Pro Highspeed serves the same model 2–3× faster at $4.00 / $20.00 — but cache reads stay at $0.20, so cached agent loops pay far less than double. On the independent Artificial Analysis Intelligence Index v4.3.2 it scores 55, fourth of all models, eleven points above K3's 44, behind only Claude Opus 5.5 (58), Claude Sonnet 5.5 and Meta Muse Wonder Full — the open-versus-closed gap narrows from 12 points to 3. Moonshot's own table: Terminal-Bench 4.0 65.7% (GPT-6 Astra 58.2%, Claude Opus 5.5 66.4%), Kimi Code Bench 3.0 Long-horizon 61.2% (K3 55.95%), BenchCAD 92.7% with tools / 72.8% without. All of it is vendor-run, and the row absent from the announcement thread is Terminal-Bench Science 0.1 at 38.1% against Astra's 68.1%. Full weights are promised by November 4, 2026 — licence unpublished, K3's identical body shipped as 1.56 TB. Costed honestly, a 30-step agent run is about $1.80, or $0.94 at an 80% cache-hit rate, against $0.13 for MiniMax M3.1 Flash or DeepSeek V4.1-Flash off-peak — you are buying frontier capability, not savings. Verdict: 4.0/5.

⚠️ Read this before you compare K3.1 Pro to anything: it is a post-trained K3, not a new model. That single sentence decides most of this review. The parameter count did not move, the context did not move, and the multimodal encoder did not move — what moved is the reinforcement learning, and therefore the price and the agentic benchmark rows. It also means the checkpoint, when it arrives, will be the same enormous object as Kimi K3, which is why the open-weights promise is worth less than it sounds. If you want the July launch — architecture, licence terms, the 1.56 TB arithmetic, the self-hosting break-even — that is the Kimi K3 review. This page is only about what changed on October 23, what it costs, and what the numbers are worth.

What Is Kimi K3.1 Pro?

Kimi K3.1 Pro (kimi-k3.1-pro) is Moonshot AI's October 23, 2026 flagship, and the first release in a new K3.1 series. Moonshot's own description is unusually plain: a post-trained version of Kimi K3, with substantially larger post-training compute. The company calls it "our most capable 3T-class open-weight model to date" — a marketing frame over an official 2.8 trillion total parameters with 104 billion active per token, the same sparse mixture-of-experts body that shipped in July. Nothing about the architecture, the 1,048,576-token context window, or the multimodal front end has changed. If you want the technical detail of the body — the 93 layers split 69 Kimi Delta Attention to 24 Gated MLA, the 896 routed experts of which 16 fire per token, the native MXFP4 format — it is all in the K3 review and none of it was re-announced.

So what is post-training, and why does it buy eleven index points? Pre-training is where a model learns language, code and world structure from raw data; it is also where nearly all of the compute bill and all of the parameter count live. Post-training is everything after: supervised fine-tuning on curated demonstrations, then reinforcement learning against verifiable rewards — a compiler that either builds the project or does not, a test suite that passes or fails, a physics simulator whose output is measurable. Because the reward is checkable, you can run the loop for a very long time on the same base model and get real gains in exactly the behaviours the reward measures. Moonshot spent that budget on long-horizon execution: multi-step tool use, self-diagnosis when an approach stalls, and domain data for coding, 3D modelling and financial research. The result is a model with the same brain and better habits, priced lower than its predecessor because serving cost did not change while the competitive position improved.

The context that makes the launch legible is what else happened the same day and after it. Moonshot shipped Kimi K3.1 Pro Highspeed alongside the standard model — same capabilities, two to three times the speed — across Kimi.ai, the iOS, Android and HarmonyOS apps, Kimi Work desktop for Windows and Apple Silicon (v3.1.0 or later), Kimi Code via /model, and the Kimi API. On the platform side it announced that the K2 family closes to new customers on October 31, 2026 and retires entirely by the end of December, which is a migration deadline dressed as housekeeping. And on November 4, 2026 it has promised the full weights. For how this rate card sits against the rest of the market, see our LLM API cost guide.

Key Features: Six Things Worth Knowing

Most of the specification is inherited. These are the parts that are new, and the parts that have a bill or a catch attached.

Long-horizon self-correction is the actual product, and the human-atlas demo is the proof. Moonshot's most specific internal test built a walking-dynamics simulator on top of human-atlas, an open-source anatomical model with 296 skeletal components including cartilage and 402 muscle components. Across 167 missions and 420 independent reviews, the model analysed results, diagnosed failures and iterated the controller until the simulated human went from collapsing within seconds to walking continuously with alternating strides for 11.2 seconds. The detail that matters is what happened when reward shaping stopped helping arm motion: instead of turning the reward dial, it ran open-loop probes of the simulator's dynamics, identified and corrected 36 errors in shoulder-muscle moment-arm calculations, and then retrained. That is a model debugging its own evaluation function, which is the failure mode that kills most long agent runs. Treat the 11.2 seconds as a vendor number and the 36 errors as the interesting part.
3D and CAD are treated as a first-class workload, not a demo. The model writes and runs scripts for Blender and CAD tools directly, without MCP connectors, and works inside Unity, Unreal Engine and Godot for full-project builds. The published examples are unusually concrete: a celestial palace scene in Unreal Engine 5.8 with a 36-bay colonnade at a 1,750-metre radius, a dragon procedurally modelled in Blender, 36 million source triangles, Nanite and Lumen; and a 19.5-second 1080p60 Blender showcase video of an airport interior built with Geometry Nodes, Cycles, Pillow, NumPy and FFmpeg. On BenchCAD, which scores 3D reconstruction from multi-view renders through CAD code generation, K3.1 Pro records 92.7% with tools and 72.8% without — against 96.2 and 73.0 for Claude Opus 5.5 and 95.9 with tools for GPT-6 Astra. Without tools it is level with the frontier; with tools it is within 3.5 points.
Factual accuracy was a stated target, and the number fell by two thirds. On Kimi Hallu Bench Hard, an internal benchmark, claim-level hallucination dropped from 23.4% on K3 to 8.3% on K3.1 Pro, with the share of responses containing at least one factual error reported at 42.4% of K3's rate. Moonshot's comparison chart puts GPT-6 Astra at 5.5, K3.1 Pro at 8.3 and Claude Opus 5.5 at 15.0 on the with-tools cut. The company is explicit that these are not overall error rates in daily use, and it is right to be — but a two-thirds reduction in a research-facing error metric is the difference between a workflow you supervise and one you cannot.
Financial research got its own benchmark and its own training. OfficeQA Pro v2 at 64.4 is a self-reported state of the art, and the internal deltas are +7.2 points on financial modelling and +11.6 on financial research over K3, with roughly 4.2× K3 on simulated business operation. The example outputs are a linked-statement NVDA DCF/LBO model with traceable formulas, an investment-committee deck, and a sector report with source-level citations.
Thinking is always on, with three effort levels and one accepted temperature. reasoning_effort accepts low, high and max, and thinking cannot be switched off — there is no "instant" mode here, unlike Inception Mercury 3. Moonshot's published benchmark results are at max with temperature 1.0 and top-p 0.95, and third-party API write-ups report that the model only accepts temperature: 1 and that reasoning_effort defaults to max. If that default holds in your client, the first thing to change in your payload is the effort field, because thinking bills at the output rate.
A Highspeed ID that is the same model, not a downgrade. kimi-k3.1-pro-highspeed runs two to three times faster for $4.00 input / $20.00 output, with cache reads unchanged at $0.20. That asymmetry matters more than the headline: a loop living on cached context pays close to the standard rate per task, and a loop re-sending fresh long prompts every step pays double. Moonshot has not published the throughput ceiling or concurrency terms that separate the two IDs.

Benchmarks: What Moonshot Says, and What the Independent Index Says

Every number in the first table is vendor-run. Moonshot does disclose the harnesses — K3.1 Pro is evaluated with Claude Code, not its own Kimi Code — which is a more honest arrangement than the K3 launch managed, where the model ran on Moonshot's harness against rivals on Claude Code and Codex. BenchCAD is scored on 1,000-case subsets randomly sampled from a 17,900-case dataset, and competitor BenchCAD scores are lifted from Anthropic's and OpenAI's own cards. The rows worth reading:

Benchmark (vendor-run, max effort)	Kimi K3.1 Pro	Claude Opus 5.5	GPT-6 Astra	Claude Fable 5.1	Kimi K3
Terminal-Bench 4.0 (Claude Code)	65.7	66.4	58.2	57.9	—
Terminal-Bench Science 0.1	38.1	63.3	68.1	40.0	—
BenchCAD — with tools	92.7	96.2	95.9	92.6	—
BenchCAD — without tools	72.8	73.0	—	60.6	—
Kimi Code Bench 3.0 Long-horizon (internal)	61.2	58.5	66.5	58.8	55.95
Kimi Hallu Bench Hard (claim-level error, lower is better)	8.3	15.0	5.5	23.3	23.4
OfficeQA Pro v2	64.4	—	—	—	—

Three things in that table are more informative than the headline. First, Terminal-Bench Science 0.1 is a rout. At 38.1% the model is 25 points behind Opus 5.5 and 30 behind GPT-6 Astra on scientific terminal work, and this is the one row Moonshot's announcement thread does not surface. If your agent's work looks like HPC, bioinformatics pipelines or anything with a scientific toolchain, the launch material does not support you and you should test it yourself. Second, the internal benchmark is where the ordering flips. On Kimi Code Bench 3.0 Long-horizon, the only benchmark in the table that is Moonshot's own, K3.1 Pro beats Opus 5.5 by 2.7 points and loses to Astra by 5.3. An internal benchmark is exactly where a lab is most likely to be measuring something it trained for; the fact that Astra still wins there is either reassuring about Moonshot's honesty or unflattering about its target, and no one outside the company can tell you which. Third, BenchCAD without tools is the most impressive row on the card. Two-tenths of a point behind Opus 5.5, on a task where the model has no tool loop to lean on, at half the token price, is a genuinely hard result to explain away.

The independent index — and the number that actually settles the argument

Artificial Analysis scores Kimi K3.1 Pro at 55 on Intelligence Index v4.3.2, which is fourth of all models and eleven points above Kimi K3's 44 at max effort. The three ahead of it are all closed: Claude Opus 5.5 at 58, Claude Sonnet 5.5, and Meta Muse Wonder Full. Artificial Analysis's own framing is that K3.1 Pro will be the leading open-weights model once its weights are released, and that the gap between the best open and the best closed model narrows from 12 points to 3. Read those two numbers together, because they are the strongest honest case for this release: in July, the best model you could download was 12 index points behind the frontier; in October, the best model you will be able to download is 3 behind it, and 2.8 trillion of it fits nobody's GPU.

Agentic work is where the eleven points came from. Artificial Analysis places K3.1 Pro third on GDPval-AA at 1,846 Elo, up 312 Elo from Kimi K3, behind only Claude Opus 5.5 at 1,866 and Muse Wonder Full at 1,857 — so on occupational-task work the model is within 20 Elo of the frontier. AutomationBench-AA rises from 58% to 73%, which is the agentic-SaaS-workflow cut and the one that maps to what tools like OpenAI Dots and Meta Muse actually ask a model to do.
Artificial Analysis and Moonshot disagree on the flagship coding row, and both are defensible. Artificial Analysis reports K3.1 Pro as achieving the highest score of any model on Terminal-Bench 4.0 in its runs, while Moonshot's own table puts it 0.7 behind Claude Opus 5.5 at 65.7 against 66.4. The gap is inside harness noise. The lesson is not that one is lying; it is that an agentic benchmark score belongs to a model plus a scaffold, and you should benchmark your own scaffold before you standardise on any model on the strength of a row.
On cost per unit of capability it is the best value at the top of the board. At $2 and $10, K3.1 Pro is 50% of Claude Opus 5.5's token price for 95% of its index score and, on Moonshot's table, 99% of its Terminal-Bench 4.0 result. That is a genuinely different position from July, when K3 cost roughly $2.00 per index task against $1.82 for Opus 5.5 at high effort while scoring ten points lower — the model was both worse and more expensive per completed task. The direction of that reversal is the story of this release.

Pricing

Every figure below is approximate, in US dollars before tax, checked on October 25, 2026 against Moonshot's API pricing documentation, the Kimi pricing resource page, and five independent model-catalogue trackers that mirror the same rate card. Confirm current pricing on platform.kimi.ai before you build — Moonshot has changed consumer plan availability and catalogue rows more than once this year.

Kimi API — the official rate card

kimi-k3.1-pro: $2.00 per 1M input · $0.20 per 1M cached input · $10.00 per 1M output · 1,048,576-token context, priced flat

kimi-k3.1-pro-highspeed: $4.00 input · $20.00 output · $0.20 cached input (unchanged) — same model, 2–3× the speed

Cache writes are billed by time to live: $2.00 per 1M at 5 minutes and $4.00 at 1 hour on the standard ID; $4.00 and $8.00 on Highspeed. Thinking tokens bill at the output rate. No long-context surcharge, no tier above the window, and no subscription: the API is strictly pay-as-you-go.

What changed against K3 — and what did not

Kimi K3 was $3.00 / $0.30 / $15.00. K3.1 Pro is $2.00 / $0.20 / $10.00. That is a 33% cut on all three legs at a flagship launch, which almost never happens, and it makes the successor cheaper than the model it replaces while scoring eleven index points higher. Against Claude Opus 5.5's $4.00 / $20.00 it is exactly half. Against Gemini 4 Argon — $2.00 / $10.00 promotional through November 30, 2026, then $4.00 / $20.00 — K3.1 Pro matches the promo price permanently, which is a real advantage once the Google discount expires.

What did not change: the 50× spread between input and output that makes agent design matter. Output is 5× input here, not 50×, but the output leg is still where bills are made, and thinking bills at the output rate. Cache reads at 10% of input are the single largest lever Moonshot gives you, and unlike K3's 90% discount era the absolute number is now $0.20, which is competitive against MiniMax's $0.015 only in relative terms — it is still 13× MiniMax and roughly 33× DeepSeek V4.1-Flash's cache rate. If your loop lives on cache hits, that line decides the bill.

Consumer, Kimi Work, Kimi Code and enterprise

K3.1 Pro is live on Kimi.ai, the iOS, Android and HarmonyOS apps, Kimi Work desktop (Windows and Apple Silicon, v3.1.0 or later) and Kimi Code via /model. Kimi's consumer plans carry a free allowance with limits and paid memberships above it — the tiers and allowances change often, so read them in the app rather than from a review. On the enterprise side, Kimi Enterprise administrators can enable K3.1 Pro for their organisation, and access is off by default at launch — a small detail with real consequences, because a rollout that "did not work" may simply be disabled.

The API has no free tier. Every token is billed from the first one, and third-party setup guides report that the API refuses K3.1 Pro calls until the account holds a positive balance, with a minimum top-up of about $1. Consumer membership does not include API credit; they are separate billing systems.

What a real agent run costs

Headline rates are useless without a token count, so here is the model we use across these reviews, sized to the way a coding or browser agent actually behaves. A 30-step agent run: each step sends 20,000 input tokens — system prompt, tool schema, accumulated conversation, tool results — and receives 2,000 output tokens of reasoning plus action. Over 30 steps that is 600,000 input tokens read and 60,000 output tokens written, which fits comfortably inside the 1,048,576-token window.

The arithmetic, written out

Input: 600,000 tokens × $2.00 ÷ 1,000,000 = $1.20
Output: 60,000 tokens × $10.00 ÷ 1,000,000 = $0.60
Total: $1.80 per task.

With 80% of input served from cache: 480,000 × $0.20 ÷ 1,000,000 = $0.096, plus 120,000 × $2.00 ÷ 1,000,000 = $0.24, so input falls to $0.336 and the task to $0.94 — a 48% saving from caching alone.

The same 30-step run	Input read	Output written	Cost	vs K3.1 Pro
Kimi K3.1 Pro, no cache	600,000	60,000	~$1.80	—
Kimi K3.1 Pro, 80% cache hits	600,000	60,000	~$0.94	0.52×
Kimi K3.1 Pro Highspeed ($4 / $20)	600,000	60,000	~$3.60	2.0×
Kimi K3 ($3 / $15)	600,000	60,000	~$2.70	1.5×
MiniMax M3.1 Flash ($0.15 / $0.60)	600,000	60,000	~$0.13	0.07×
DeepSeek V4.1-Flash-1017, off-peak / peak	600,000	60,000	~$0.13 / ~$0.25	0.07× / 0.14×
Inception Mercury 3 ($0.20 / $0.70)	600,000	60,000	~$0.16	0.09×
Claude Opus 5.5 ($4 / $20)	600,000	60,000	~$3.60	2.0×
GPT-6 Astra ($10 / $50)	600,000	60,000	~$9.00	5.0×

Our arithmetic at published list prices, with the assumptions stated above. Cache-hit share is an assumption — measure yours from your provider dashboard. Competitor rates approximate, checked October 25, 2026.

Three conclusions. First, K3.1 Pro is the cheapest route to this class of capability, and the most expensive route to any capability. At $1.80 it is exactly half of Claude Opus 5.5 and one fifth of GPT-6 Astra for a model that scores within three index points of the former. At the same time it is 14 times MiniMax M3.1 Flash and DeepSeek V4.1-Flash at off-peak for the same token counts. Both statements are true; which one matters depends on whether your bottleneck is quality or volume. A thousand runs a month is $1,800 on K3.1 Pro against $126 on MiniMax, and that is not a rounding difference you can absorb.

Second, caching is worth more than shopping. At an 80% cache-hit rate the K3.1 Pro task falls from $1.80 to $0.94 — a 48% saving from one architectural decision about how you structure your loop, which is a bigger lever than the difference between most vendors on this page. Moonshot's cache writes are cheap ($2.00 for a 5-minute TTL), so a stable system prompt and tool schema that you deliberately pin is the highest-return engineering you can do on this model.

Third, Highspeed is a latency product, not a quality product, and it is priced against the frontier rather than against itself. At $3.60 per task it lands exactly on Claude Opus 5.5's price for a model three index points behind it. That is defensible only if the two to three times speed is worth more to you than the capability. If you are serving an interactive coding assistant where a user is watching, it probably is. If you are running overnight batch agents, it is a 2× tax for throughput you were not going to notice.

Is It Free?

Partly, and the split matters.

The consumer side is free within limits. K3.1 Pro is available on Kimi.ai and the mobile apps through a free allowance. Moonshot does not publish a token budget for it — allowances are expressed as message counts and feature limits that change — so the honest answer is that you can use the model for free and cannot use it for free at volume. Read the current allowance in the app before you plan around it.
The API has no free tier at all. There is no trial credit, no per-key grant and no free model ID. Billing starts at the first token, and third-party setup guides report a minimum top-up of about $1 before K3.1 Pro calls are accepted. This is a real difference from the 100 million free tokens Inception hands every new key, and from the generous free chat tiers at DeepSeek and Qwen.
The open-weights route is zero-cost in principle and unavailable in practice today. Weights are promised by November 4, 2026. Until a repository exists there is nothing to download, and once it does the object will be roughly K3-sized — 1.56 TB in MXFP4, against community quants from 466 GB to 649 GB — so "free" will mean "free to rent eight accelerators to hold it".

Practical read: evaluate on the free Kimi app or on a $20 API top-up, then budget at $2.00 / $10.00 for anything you keep running. There is no configuration in which this model is both free and in production.

Pros & Cons

✅ Pros

A 33% price cut at a flagship launch, on all three rates. $2.00 / $0.20 / $10.00 against K3's $3.00 / $0.30 / $15.00, and exactly half of Claude Opus 5.5.
Eleven index points for nothing. Artificial Analysis 55 against K3's 44 — fourth of all models, and the highest score any open-weights model has recorded.
The open-versus-closed gap narrows from 12 points to 3. That is the single most consequential number in the release, and it is third-party.
Terminal-Bench 4.0 at 65.7%, ahead of GPT-6 Astra's 58.2 and 0.7 behind Claude Opus 5.5 — for half the token price.
BenchCAD without tools at 72.8% is level with Opus 5.5's 73.0, which makes it a serious 3D and parametric-CAD model rather than a demo.
The human-atlas run is real evidence of self-correction: 167 missions, 420 reviews, 11.2 seconds of continuous gait, and 36 moment-arm calculation errors the model found and fixed on its own.
Claim-level hallucination down from 23.4% to 8.3% on Kimi Hallu Bench Hard, and OfficeQA Pro v2 at 64.4 — the research workflow is now defensible.
A 1M context with no long-context surcharge, plus cache reads at 10% of input and cheap $2.00 cache writes.
Highspeed is the same model, not a distilled substitute, and its cache reads stay at $0.20.
Harness disclosure improved. K3.1 Pro is evaluated on Claude Code rather than Moonshot's own harness, and competitor BenchCAD numbers are sourced from their own cards.
Zero Data Retention available to API customers, and a real distribution footprint: web, three mobile platforms, desktop, CLI and OpenRouter.

❌ Cons

The weights do not exist. November 4, 2026 is an eleven-day-old promise at publication time, with nothing on Hugging Face on launch day and no licence text published anywhere.
And when they land, nobody reading this can run them. Post-training does not shrink a model: K3's identical body shipped as 1,561 GB and needs eight B300-class accelerators to load.
Every benchmark is vendor-run. BenchCAD samples 1,000 of 17,900 cases, the internal Kimi Code Bench is Moonshot's own, and no independent reproduction of K3.1 Pro exists yet.
Terminal-Bench Science 0.1 at 38.1% against 63.3 for Opus 5.5 and 68.1 for GPT-6 Astra — a 25-point deficit on scientific terminal work, absent from the announcement thread.
14× the Flash tier on a modelled agent run. $1.80 against $0.13 for MiniMax M3.1 Flash or DeepSeek V4.1-Flash off-peak. This is a capability purchase, not a savings purchase.
"3T-class" is marketing. The official figure is 2.8 trillion total parameters, and the phrasing exists to sit next to competitors described as 3T while being 200 billion short.
Thinking cannot be switched off, there is no instant mode, and third-party write-ups report reasoning_effort defaults to max — the most expensive setting — with temperature pinned at 1.
The licence is unpublished, and K3's Kimi K3 License carried a $20M model-as-a-service revenue gate plus a UI attribution rule. Assume nothing until the text appears.
K2 retires on a deadline. New customers lose access October 31, 2026 and the series shuts down by year-end, so migration work is real and unscheduled by you.
Highspeed's limits are undocumented. No throughput ceiling, rate limit or concurrency terms distinguish the two IDs on the public pages.
Jurisdiction and compliance questions remain. A Beijing lab's hosted API means data leaving your jurisdiction, and Kimi Enterprise has K3.1 Pro disabled by default at launch.

5 Best Kimi K3.1 Pro Alternatives

MiniMax M3.1 Flash

Best if your bottleneck is volume rather than the last three index points

~$0.15 input / ~$0.60 output / ~$0.015 cached per 1M flat, with no peak pricing — ~$0.13 on our 30-step run against K3.1 Pro's ~$1.80, a 14× difference.
428B total / ~23B active MoE with a 1,048,576-token context, a native vision encoder, and five thinking levels from Low to Max.
MiniMax claims agent-task cost down 70% and throughput above 200 tokens per second at peak. Reasoning is always on and cannot be disabled, so there is no cheap routing mode.
Weights are promised "later" rather than dated, and there is no free API tier — the free path is a limited preview in MiniMax Code.
The honest trade: you give up roughly 30 index points and the long-horizon self-correction behaviour. Read the MiniMax M3.1 Flash review.

DeepSeek V4.1-Flash-1017

Best if you want the weights in your hands today, under a permissive licence

MIT-licensed weights already published: 763B total parameters across 48 safetensors shards, about 510 GB, with a vLLM memory floor near 614 GB — rentable, and the only route on this page to real self-hosting.
~$0.15 / ~$0.60 per 1M off-peak, doubling at peak — about $0.13 per 30-step run off-peak, and the same 1,048,576-token context.
Be clear about what you are trading: Terminal-Bench 4.0 is 42.3% for DeepSeek against 65.7% for K3.1 Pro. That is a 23-point gap on exactly the long-horizon agentic work this model was post-trained for.
Cache hits at roughly $0.003 per 1M make it the cheapest agent memory on the market — 66× below K3.1 Pro's $0.20.
Off-peak pricing is 50% of peak, so scheduling flexible batch work is a real lever. See the DeepSeek V4.1-Flash review.

Inception Mercury 3

Best if latency dominates and you want a reasoning dial you can turn to zero

~$0.20 / ~$0.70 per 1M at list — roughly $0.16 per 30-step run, a tenth of K3.1 Pro — with a 50% launch discount and 100 million free tokens on every new key.
A diffusion language model, not autoregressive: it denoises a whole block in parallel, which is a genuinely different architecture rather than another MoE.
1,000,000-token context, 131,072-token output, and four reasoning levels including instant — so a router or classifier never pays for a thinking block.
Big caveat: no weights, no parameter count and no licence of any kind, which is a harder position than K3.1 Pro's dated promise.
Speed is the pitch and it survives discounting: ~1,068 tokens/s claimed, ~243 measured on live OpenRouter traffic. Read the Inception Mercury 3 review.

Claude Opus 5.5

Best if three index points and a published licence are worth 2× the token price

58 on Artificial Analysis Intelligence Index v4.3.2 against K3.1 Pro's 55 — three points, at exactly double the price at $4.00 / $20.00.
Terminal-Bench 4.0 66.4% against 65.7, and Terminal-Bench Science 0.1 at 63.3% against 38.1 — the second number is the one that decides hard scientific workloads.
BenchCAD 96.2% with tools and 73.0% without, so it remains the reference for 3D reconstruction, by 3.5 points with tools and 0.2 without.
No licence to read, no jurisdiction question, no checkpoint to wait for, and a documented system card that Moonshot's competitors quote from.
At ~$3.60 per 30-step run it is twice the bill for a 3-point gain — which is either obviously worth it or obviously not, depending on whether your work is science or software. See the Claude review.

Gemini 4 Argon

Best if you want the same headline price with the lowest measured hallucination rate

~$2.00 / ~$10.00 per 1M with $0.10 cached — matching K3.1 Pro today, but only as a promotion running to November 30, 2026, after which it doubles to $4.00 / $20.00.
53 on the index against K3.1 Pro's 55, so it is slightly behind on the composite while being $0.10 cheaper on cache reads.
The lowest hallucination rate measured at its launch — 15% against 51% for GPT-6 Astra — and it took first place on the Vals Index at 68.90%.
Google's own table has it last on FrontierSWE v2 at 55.0 and Terminal-Bench 4.0 at 57.4, both well behind K3.1 Pro's 65.7. Benchmark skepticism applies in both directions.
No free API tier, implicit caching only, and the best model access sits behind the $99.99 Ultra tier. Read the Gemini 4 Argon review.

Comparison Table

Model	Input /1M	Output /1M	Cached /1M	Context	Weights	AA index
Kimi K3.1 Pro	$2.00	$10.00	$0.20	1,048,576	Promised Nov 4, 2026 — licence unpublished	55
Kimi K3.1 Pro Highspeed	$4.00	$20.00	$0.20	1,048,576	Same as above	55
Kimi K3	$3.00	$15.00	$0.30	1,048,576	Yes — 1.56 TB, Kimi K3 License	44
MiniMax M3.1 Flash	$0.15	$0.60	$0.015	1,048,576	Promised, undated	—
DeepSeek V4.1-Flash-1017	$0.15	$0.60	~$0.003	1,048,576	Yes — MIT, ~510 GB	—
Inception Mercury 3	$0.20	$0.70	$0.02	1,000,000	No — none announced	—
Claude Opus 5.5	$4.00	$20.00	$0.20	1M	No — proprietary	58
Gemini 4 Argon	$2.00 → $4.00	$10.00 → $20.00	$0.10	1,048,576	No — proprietary	53

All prices are approximate, in US dollars, checked October 25, 2026, and change often. Confirm on each vendor's own pricing page. DeepSeek rates are off-peak and double during peak hours. Gemini 4 Argon's promotional rate runs to November 30, 2026 and then doubles. A dash in the index column means the model is not scored on Artificial Analysis Intelligence Index v4.3.2 at the time of writing, not that it scored zero.

Which One Should You Choose?

Agentic coding, high volume — Kimi K3.1 Pro, and it is not close on value. Terminal-Bench 4.0 at 65.7% for $1.80 a run against Claude Opus 5.5's 66.4% for $3.60. Pin your system prompt and tool schema so the cache does the work, set reasoning_effort to high rather than letting it default to max, and measure before you scale.
3D, CAD and parametric modelling — K3.1 Pro is the value pick and the only one priced for iteration. 92.7% with tools and 72.8% without on BenchCAD, against 96.2 and 73.0 for Opus 5.5. When a modelling task takes twenty attempts, paying half per attempt matters more than the 3.5-point tool gap. Blender and CAD scripting without MCP connectors is a real workflow advantage.
Financial research — try it, but verify the SOTA claim. OfficeQA Pro v2 at 64.4 and claim-level hallucination down from 23.4% to 8.3% are the numbers that matter, and both are self-reported. Run it against three of your own research questions with sources you can check before you put it in a deliverable.
Scientific and HPC terminal work — do not standardise on it yet. Terminal-Bench Science 0.1 at 38.1% against 63.3 for Opus 5.5 and 68.1 for GPT-6 Astra is a 25-to-30-point deficit, and it is the one row the launch material omits. Test your own toolchain first.
Budget-sensitive or extremely high volume — go to the Flash tier. MiniMax M3.1 Flash and DeepSeek V4.1-Flash both cost ~$0.13 on the same run. At a million runs a month the difference is $1.67 million. You will feel the quality drop on long-horizon agentic work and you will not feel it on classification, extraction or summarisation.
You need to self-host today — Kimi K3.1 Pro cannot help you. There is no checkpoint. DeepSeek V4.1-Flash under MIT at ~510 GB is the only realistic open route on this page, and it needs roughly 614 GB of VRAM, so "self-host" still means rented accelerators or a serious capital purchase.
You are already on Kimi K3 — migrate, and migrate before December. The successor is cheaper on all three rates and better on every published row. Change the model ID to kimi-k3.1-pro, audit your reasoning_effort payloads, and use the K2 shutdown at year-end as the forcing function to finish the job.

"Open Weights by November 4" — What the Promise Is Actually Worth

Moonshot has committed to releasing the full weights by November 4, 2026. That commitment is worth real money to three groups and roughly nothing to everyone else, and the difference is worth spelling out because "open-weight" is doing a lot of marketing work in this launch.

The licence is not published. K3 shipped under the Kimi K3 License: use, copy, modify, fine-tune and sell free of charge, with a gate requiring a separate agreement above $20 million of model-as-a-service revenue measured on aggregate group revenue, and a UI attribution rule above 100M monthly users or $20M monthly revenue. That is permissive for almost everyone and not open source in the OSI sense. Whether K3.1 Pro repeats it, tightens it, or ships under something more permissive is unknown, and a lab that has spent this much post-training compute on a frontier model has every incentive to tighten. Do not plan a commercial product on an unpublished licence.
The size is inherited, and it is the whole story. Post-training does not shrink a model. K3's checkpoint is 1,561,018,243,668 bytes across 96 MXFP4 shards; K3.1 Pro's will be the same order of magnitude because it is the same architecture. The most aggressive community quantisations of K3 run 466 GB to 649 GB, and at those bit-widths you are running a different model from the one on the benchmark table. The realistic hardware floor is a node of eight B300 or MI355X class accelerators, and the cheapest legitimate path is renting one — which, on K3's published economics, only beat the API above roughly two billion output tokens a month. At $2/$10 the crossover is worse, not better.
So who is it for? Three groups, and they are all legitimate. Researchers get to verify the benchmark claims and the safety story independently, which is the entire point of releasing weights and is why the November 4 date matters more than the licence. Enterprises with hard data-residency requirements get a path that a hosted API cannot offer, at a price only large organisations pay. Model builders get a 2.8T post-trained checkpoint to distil from, which is how the 466 GB community quants exist at all. If you are an individual developer, the honest reading is that the open-weights release changes nothing about your options: you will call the API, and your real decision is whether $1.80 a task buys enough over $0.13 a task.
One safety footnote worth knowing. Moonshot says it removed cyber-related training data from K3.1 Pro ahead of the open-weights release. That is a claim you cannot check today and, once the checkpoint exists, one that outside researchers can partially test. Hold it as a statement of intent rather than a fact.

Privacy & Data

Zero Data Retention is available on the API. Moonshot documents ZDR for API customers, which means prompts and completions are not stored for training or review. Confirm it is enabled on your account rather than assuming it — it is a setting, not a default you can infer from a marketing page.
The consumer apps are a different deal. Kimi.ai and the mobile apps are consumer products with consumer terms; a free allowance is not a data agreement in your favour. Read the policy before you paste anything you would not publish, and do not assume the API's ZDR extends to the chat surface.
Jurisdiction is the unresolved issue. Moonshot is a Beijing lab, so a hosted API means inference and any retained metadata crossing into Chinese jurisdiction. For regulated industries in the EU, UK or US public sector, that is typically a blocking question regardless of benchmark scores, and it is the single strongest argument for the open-weights path — which is also the path that requires eight accelerators.
Enterprise access is off by default. Kimi Enterprise administrators can enable K3.1 Pro, but a fresh organisation will not have it. If a rollout appears broken, check the admin toggle before you file a bug.
Output rights are your problem to check. No vendor on this page warrants that model output is free of third-party claims, and the 3D examples in Moonshot's launch material use external asset libraries under their own licences. If you are shipping generated assets commercially, track provenance yourself.

Frequently Asked Questions

Is Kimi K3.1 Pro free to use?

The chat product has a free tier and the API does not. On Kimi.ai, the iOS and Android apps, and the HarmonyOS app, K3.1 Pro is available through a free allowance that Moonshot limits by message count and by feature rather than publishing a token budget, so the honest statement is that you can use the model for free and you cannot use it for free at volume; check the current allowance in the app before you plan around it. The Kimi API has no free tier at all. Every token is billed from the first one at $2.00 per million input, $0.20 per million cached input and $10.00 per million output, and third-party setup guides report that the API rejects K3.1 Pro calls until the account holds a positive balance, with a minimum top-up of about one dollar. There is no open-weights route that would let you run it at zero marginal cost today, because the checkpoint is not published: Moonshot promises the full weights by November 4, 2026, and until a repository appears on Hugging Face there is nothing to download, quantise or serve locally. Compare that with DeepSeek V4.1-Flash, whose MIT-licensed weights are already out, or MiniMax M3.1 Flash, where weights are at least promised. If you need a genuinely free path to a frontier-adjacent model today, the free tiers at Kimi, DeepSeek and Qwen are the realistic options.

How much does the Kimi K3.1 Pro API cost?

On Moonshot's own platform, kimi-k3.1-pro costs $2.00 per million cache-miss input tokens, $0.20 per million cache-hit input tokens and $10.00 per million output tokens, with the full 1,048,576-token context priced flat, so there is no long-context surcharge anywhere on the card. Cache writes are billed separately by time to live: $2.00 per million for a five-minute TTL and $4.00 per million for a one-hour TTL. That is a 33 percent cut against Kimi K3's $3.00, $0.30 and $15.00 on all three legs, which is unusual at a flagship launch and is the clearest signal that Moonshot is competing on price rather than capacity this quarter. The Highspeed model ID, kimi-k3.1-pro-highspeed, doubles the regular input, output and cache-write rates to $4.00, $20.00 and $4.00 or $8.00, but leaves cache reads at $0.20, so a cache-heavy agent loop pays much less than double. Worked example: a 30-step agent run that sends 20,000 input tokens and receives 2,000 output tokens per step reads 600,000 input tokens and writes 60,000 output tokens, which bills at 600,000 times $2.00 divided by one million, about $1.20 of input, plus 60,000 times $10.00 divided by one million, $0.60 of output, for a total of about $1.80 per task. At an 80 percent cache-hit rate the same run costs about $0.94. All figures are approximate, in US dollars before tax, and change often, so confirm on the Kimi API pricing page before you build.

Can I download and run Kimi K3.1 Pro at home?

Not yet, and not on home hardware even after the promised release. As of October 25, 2026 there is no checkpoint: Moonshot says the full model weights will be released by November 4, 2026, and on launch day nothing had appeared under the lab's Hugging Face organisation. When it does arrive, the size follows from the architecture, because K3.1 Pro is a post-trained K3 rather than a new model. Kimi K3 shipped as 96 safetensors shards totalling 1,561,018,243,668 bytes in native MXFP4, roughly 1.56 terabytes, and the community quantisations that exist for K3 run from about 649 GB at Unsloth's dynamic one-bit setting down to about 466 GB at the smallest, none of which fits a consumer machine and all of which are a different model from the one Moonshot benchmarked. K3's published minimum serving configuration is a single node of eight B300 or eight MI355X class accelerators, and the consumer ceiling is a 512 GB Mac Studio, which is under a third of the weight floor. The practical answer for almost everyone is to rent the model through the API or through a host such as OpenRouter, and to treat the open-weights release as a research and data-residency event rather than a local-inference one. If local inference is genuinely your requirement, the models that fit are DeepSeek V4.1-Flash under MIT at about 510 GB of weights and a 614 GB VRAM floor, or a smaller open family such as Qwen.

Is Kimi K3.1 Pro actually better than Claude Opus 5.5?

It is close on agentic coding and behind on the composite, and the price changes what close is worth. On Artificial Analysis Intelligence Index v4.3.2, Kimi K3.1 Pro scores 55 and Claude Opus 5.5 scores 58, so Opus leads by three points while costing exactly twice as much per token at $4.00 and $20.00 against $2.00 and $10.00. On Moonshot's own Terminal-Bench 4.0 table the gap is 0.7 points, 65.7 percent for K3.1 Pro against 66.4 for Opus 5.5, and on BenchCAD without tools it is 0.2 points, 72.8 against 73.0, with tools widening it to 3.5 points at 92.7 against 96.2. Moonshot's internal Kimi Code Bench 3.0 Long-horizon actually puts K3.1 Pro ahead of Opus 5.5 at 61.2 against 58.5, though behind GPT-6 Astra at 66.5. Against that, Artificial Analysis reports K3.1 Pro as the highest-scoring model on Terminal-Bench 4.0 in its own runs, which contradicts Moonshot's table and is a reminder that agentic scores belong to a model plus a harness. Where the comparison stops being close is scientific terminal work: Terminal-Bench Science 0.1 is 38.1 percent for K3.1 Pro against 63.3 for Opus 5.5 and 68.1 for GPT-6 Astra. The fair summary is that K3.1 Pro buys roughly 95 percent of Opus 5.5's agentic coding capability for 50 percent of the token price, that it is the better purchase for high-volume coding agents, and that Opus 5.5 remains the right answer when the task is hard science, when a 3-point index gap matters more than the bill, or when you need a vendor with a published licence and no jurisdiction question.

What is Kimi K3.1 Pro Highspeed, and does it cost more?

Highspeed is the same model served faster, not a smaller or distilled one, and it does cost more. Moonshot describes kimi-k3.1-pro-highspeed as delivering two to three times the speed of the standard model with the same capabilities. On the rate card it is $4.00 per million input tokens and $20.00 per million output tokens against $2.00 and $10.00 for the standard model, with cache writes also doubled at $4.00 for a five-minute TTL and $8.00 for one hour. The detail that matters is that cache reads stay at $0.20 per million on both models, so the premium is not a clean doubling: an agent loop that runs mostly on cached context pays close to the standard rate, while a loop that re-sends fresh long prompts every step pays the full double. On our modelled 30-step run that reads 600,000 input tokens and writes 60,000 output tokens, the standard model costs about $1.80 and Highspeed about $3.60, which lands exactly on Claude Opus 5.5's price for a model that is faster but three index points behind. Moonshot has not published the throughput ceiling, rate limits or concurrency terms that distinguish the two IDs on its public pages, so if latency is the reason you are considering Highspeed, benchmark it against the standard ID on your own workload before you commit.

Should I switch from Kimi K3 to Kimi K3.1 Pro?

Yes if you are calling the API, because the successor is cheaper on every rate and better on every published benchmark. The model ID changes from kimi-k3 to kimi-k3.1-pro, the request shape is unchanged, and the price falls from $3.00 to $2.00 input, $0.30 to $0.20 cached and $15.00 to $10.00 output, so a workload that cost $2.70 on K3 costs $1.80 on K3.1 Pro for the same 600,000 input and 60,000 output tokens. Quality moves too: Terminal-Bench 4.0 at 65.7 against K3's result, Kimi Code Bench 3.0 Long-horizon at 61.2 against 55.95, and claim-level hallucination falling from 23.4 to 8.3 percent on Kimi Hallu Bench Hard. There is no scenario in which staying on K3 for cost reasons makes sense. Two migration warnings apply. First, thinking is always on and cannot be switched off, and third-party API write-ups report that reasoning_effort defaults to max, the most expensive setting, so audit your call payloads before you assume the new price holds. Second, the wider K2 family is being retired: Moonshot's pricing documentation says new customers lose access to K2 models from October 31, 2026 and the series shuts down by the end of December, so any workload still pinned to K2 has roughly two months to move, and K3.1 Pro is the obvious destination for most of it.

AI ModelsKimi K3.1 ProMoonshot AIOpen WeightsMixture-of-Experts1M ContextAgentic CodingBenchCADPricing2026

Free tools we built

More Free Tools

📒 Bookkeeping Reviews

Honest accounting software reviews: Wave, QuickBooks, Xero & free options.

🎮 Gaming Gear Reviews

Esports hardware decoded: gaming mice, keyboards, headsets & budget picks.

📄 PDF Merge Online

Combine multiple PDF files into one — free, no upload, files stay on your device.

📷 ID Photo Studio

Passport & visa photos with AI background changer, compressor & converter.

🧮 HK Salary Tax Calculator

Estimate Hong Kong salaries tax for 2026/27 in seconds — see what you owe before you file your tax return.

🎁 HK Benefits Quick Check

Government & benefits eligibility checker — OAA, CSSA, housing & more.

📅 HK Calendar 香港月曆

Public holidays, lunar dates & solar terms for 2020-2036 — bilingual, free, no sign-up.

★ Our own product 🏠 Home Expense Bookkeeping

Family bookkeeping: income, spending, net worth and monthly bills. Data stays on your device. HK$98 one-time.

Explore AI Tool Scout: Home About Kimi K3 Review Qwen Review DeepSeek Review DeepSeek V4.1-Flash Review Gemini 4 Argon Review Gemini Review Meta Muse Wonder Review Meta Muse Review Meta Muse Spark Open Review OpenAI Dots Review GPT-6 Astra Review ChatGPT Review Claude Review Grok Review GitHub Copilot Review Cursor Review Perplexity Review Notion AI Review Otter.ai Review LLM API Costs Ideogram Review Midjourney Review Leonardo AI Review Kling AI Review Hailuo AI Review Runway Review Sora Review HeyGen Review ElevenLabs Review Suno Review Unboring AI Review Canva Review Grammarly Review Wave Accounting Review MiniMax M3.1 Flash Review Inception Mercury 3 Review Mistral Voxtral 3 Review All Free Tools GPT-6.1 Terra Review Gemini 4 Carbon Review Mistral Large 4 Review Claude Opus 5.5 Review Meituan LongCat-2.5 Review

AI Tool Scout — independent reviews of AI tools, updated regularly. See more free tools in our directory. Pricing, model specifications, licence terms and benchmark figures in this review were checked on October 25, 2026 against Moonshot AI's Kimi K3.1 Pro technical blog at kimi.ai/blog/kimi-k3-1-pro, the Kimi API pricing documentation at platform.kimi.ai, the Kimi pricing resource page, the OpenRouter model listing, five independent model-catalogue trackers that mirror the same rate card, Artificial Analysis's launch analysis and model page for Intelligence Index v4.3.2, the earlier Kimi K3 model card and technical report for the shared architecture and checkpoint size, and Moonshot's K2 retirement notice. Many of the benchmark results quoted here are vendor-reported on vendor-selected harnesses and have not been independently reproduced; where Moonshot's own table and Artificial Analysis disagree, both figures are stated. The full K3.1 Pro weights were not published as of October 25, 2026, so every claim about the downloadable checkpoint is an inference from the K3 release of the same architecture, and the licence terms are unknown. API prices, consumer allowance, promotion windows, open-weights availability and catalogue rows change frequently — confirm current figures on the vendor's official pages before you buy, subscribe or deploy. Companies and products are the property of their respective owners; this review is not affiliated with Moonshot AI, Anthropic, OpenAI, Google, MiniMax, DeepSeek, Inception or Meta.