Mistral Large 4 vs Kimi K3.1 Pro: Complete Comparison and Report on Pricing, Benchmarks, Open-Weight Licenses, and Cost per Task
Mistral Large 4 and Kimi K3.1 Pro are two of the biggest open-weight model releases of October 2026, and they arrive from opposite directions. Mistral first put Large 4 into a public API preview on October 6 and released the improved weights on Hugging Face on October 29, after three more weeks of reinforcement learning. The model has about 1 trillion total parameters, 49 billion active per token, a 1M-token context window and a Pixtral vision encoder, and Mistral calls it the best open-weight model for coding. Moonshot AI launched Kimi K3.1 Pro on October 23 as a post-trained version of Kimi K3, the 2.8-trillion-parameter model with 104 billion active parameters it released in July, and offers it in the Kimi app, Kimi Work, Kimi Code and the Kimi API, with full weights promised by November 4.
The independent measurements point one way on capability and the price lists point the other way. On the Artificial Analysis Intelligence Index v5.0, Kimi K3.1 Pro scores 46 and Mistral Large 4 (High) scores 26, and Kimi leads on all ten individual evaluations that Artificial Analysis lists for both models, including Terminal-Bench 4.0 (67% vs 25%) and HLE-Diamond (37% vs 17%). Mistral Large 4 costs $0.70 per 1M input tokens and $2.10 per 1M output tokens, against $2.00 and $10.00 for Kimi K3.1 Pro, which makes it 65% cheaper on input and 79% cheaper on output. Because the two models spend a similar number of tokens, the gap survives on completed work: Artificial Analysis puts the cost per index task at $2.06 for Mistral Large 4 and $3.68 for Kimi K3.1 Pro.
The practical split follows from those numbers. Kimi K3.1 Pro is the stronger choice for long agentic coding, research and knowledge work where the quality of each answer outweighs its price. Mistral Large 4 suits high-volume and cost-sensitive workloads, teams that want to download and run the weights today on a single eight-GPU node, and buyers who want a model trained and served by a European company. Prices in this comparison are those listed on the official Mistral and Kimi pricing pages on November 3, 2026.
··········
MISTRAL LARGE 4 AND KIMI K3.1 PRO: RELEASE TIMELINE AND MODEL IDENTITY.
Mistral Large 4, nicknamed "Le Chonk" by Mistral, reached developers in two stages. The October 6 preview introduced the model through Mistral Studio and the Mistral API under the documentation version label v26.10, and the October 29 open-weight release shipped a checkpoint that Mistral says gained substantially from the extra reinforcement learning, with Terminal-Bench 4.0 rising from 28% in the preview to 41% in Mistral's own measurement. The Mistral documentation page for the API model still labels it "Public Preview".
Kimi K3.1 Pro is the first model of Moonshot AI's K3.1 series. Moonshot describes it as its most capable 3T-class open-weight model and says it was built on Kimi K3 with substantially larger post-training compute. It ships alongside Kimi K3.1 Pro Highspeed, a faster serving tier of the same model that Moonshot says runs two to three times faster. Moonshot does not publish a separate parameter count for K3.1 Pro; Artificial Analysis lists it at 2.8 trillion total and 104 billion active parameters, the same figures Moonshot's model card gives for Kimi K3.
The architectures differ in how much of the network each token uses. Mistral Large 4 routes each token to 4 of 128 experts plus one shared expert across 71 mixture-of-experts layers, with Multi-head Latent Attention and a DeepSeek Sparse Attention indexer, according to the Hugging Face release notes. Kimi K3 selects 16 of 896 experts plus two shared experts and mixes 69 KDA layers with 24 gated MLA attention layers, and its weights use MXFP4 quantization-aware training.
........
Item | Mistral Large 4 | Kimi K3.1 Pro |
Developer | Mistral AI (France) | Moonshot AI (China) |
First availability | October 6, 2026 (API public preview) | October 23, 2026 |
Open weights | Released October 29, 2026 | Promised by November 4, 2026 |
Total parameters | About 1T (1.05T in Mistral docs) | 2.8T (Kimi K3 base) |
Active parameters | 49B per token (52B with embeddings and output layers) | 104B (Kimi K3 base) |
Experts | 128 routed, 4 active, 1 shared | 896 routed, 16 active, 2 shared (K3) |
Input modalities | Text and images | Text, images and video |
API model ID | mistral-large-4 | kimi-k3.1-pro, kimi-k3.1-pro-highspeed |
........
··········
MISTRAL LARGE 4 VS KIMI K3.1 PRO API PRICING AND RATE STRUCTURE.
Both companies charge a flat rate per token with no context-length tiers, and both discount cached input by 90%. Mistral lists Mistral Large 4 at $0.70 per 1M input tokens, $0.07 per 1M cached input tokens and $2.10 per 1M output tokens on its API pricing page and in the model documentation. Mistral offers a batch endpoint for the model at /v1/batch but does not publish a batch rate for it.
Moonshot prices Kimi K3.1 Pro at $2.00 per 1M input tokens on a cache miss, $0.20 on a cache hit and $10.00 per 1M output tokens. Caching is automatic, with a five-minute default lifetime that can be extended to one hour, and Moonshot lists cache-write rates of $2.00 for the five-minute tier and $4.00 for the one-hour tier. The Highspeed tier doubles the input-miss and output prices to $4.00 and $20.00. Batch jobs on the Kimi platform cost 60% of the standard price. API access to K3.1 Pro becomes available after a top-up of at least $1, and Moonshot charges $0.004 per call to its web search tool.
........
Price (USD per 1M tokens) | Mistral Large 4 | Kimi K3.1 Pro | Kimi K3.1 Pro Highspeed |
Input (uncached) | $0.70 | $2.00 | $4.00 |
Cached input | $0.07 | $0.20 | $0.20 |
Output | $2.10 | $10.00 | $20.00 |
Cache write | Not listed | $2.00 (5 min) / $4.00 (1 h) | $4.00 (5 min) / $8.00 (1 h) |
Batch | Endpoint available, rate not published | 60% of standard price | Not listed |
Blended rate (Artificial Analysis, 7:2:1) | $0.40 | $1.54 | Not measured |
........
Kimi K3.1 Pro is already a price cut inside Moonshot's own lineup. Artificial Analysis lists the predecessor Kimi K3 at $3.00 input and $15.00 output, so K3.1 Pro charges a third less on both. Mistral Large 4 still undercuts it by a wide margin, and the gap on output tokens, where reasoning models spend most of their budget, is close to five to one.
··········
CONTEXT WINDOW, OUTPUT LIMITS AND REASONING EFFORT IN MISTRAL LARGE 4 AND KIMI K3.1 PRO.
Both models accept about one million tokens of context. Mistral's documentation lists 1M for Mistral Large 4; Moonshot lists 1,048,576 tokens for Kimi K3.1 Pro. Moonshot also documents an output ceiling: 131,072 tokens by default, configurable up to 1,048,576 through the max_completion_tokens parameter. Mistral does not state a maximum output length for Mistral Large 4.
Reasoning is always on in both models. Mistral Large 4 accepts a reasoning_effort value of low, medium or high, with high as the default, and Hugging Face notes that a request for "none" still runs at high. Mistral recommends low for quick chat, medium for everyday tasks and high for complex and agentic work. Kimi K3.1 Pro takes a top-level reasoning_effort field with low, high and max, and defaults to max. Moonshot fixes the sampling parameters for K3.1 Pro at temperature 1.0 and top-p 0.95 and asks developers to omit them; Mistral recommends the same values for Large 4 without locking them.
........
Setting | Mistral Large 4 | Kimi K3.1 Pro |
Context window | 1M tokens | 1,048,576 tokens |
Maximum output | Not stated | 131,072 default, up to 1,048,576 |
Effort levels | low, medium, high | low, high, max |
Default effort | high | max |
Reasoning can be disabled | No | No |
Sampling | Temperature 1.0, top-p 0.95 recommended | Temperature 1.0, top-p 0.95 fixed |
Structured output | Supported | json_schema with strict mode |
Function calling | Supported | Supported, including dynamic tool loading |
........
The default settings deserve attention when reading benchmark tables. Kimi's published results use max effort, the top of its range and the API default, while Artificial Analysis lists Mistral Large 4 at High, which is also the top of Mistral's range.
··········
OPEN-WEIGHT STATUS AND LICENSE TERMS: MISTRAL ECOSYSTEM LICENSE VS KIMI K3 LICENSE.
Mistral Large 4's weights are downloadable today. Mistral publishes an FP8 checkpoint as mistralai/Mistral-Large-4-1T-A52B and an NVFP4 checkpoint for NVIDIA Blackwell GPUs, and community GGUF quantizations exist for llama.cpp. The Mistral Ecosystem Open-Weight License is a modified MIT license that permits use, modification, fine-tuning and commercial use, with one condition: a business that operates a Model-as-a-Service offering needs a separate commercial license from Mistral.
Kimi K3.1 Pro's weights are not public yet. Moonshot's launch post says the full weights will be released by November 4, and Artificial Analysis lists the model as proprietary with no public weights as of November 3. Moonshot has not published the license that will apply to K3.1 Pro. The Kimi K3 License that covers the base model is MIT-style with two thresholds: a Model-as-a-Service business with more than $20 million of revenue over any 12 months must sign a separate agreement before commercial use, and a commercial product with more than 100 million monthly active users or more than $20 million of monthly revenue must display "Kimi K3" prominently in its interface.
........
License and weights | Mistral Large 4 | Kimi K3.1 Pro |
Weights downloadable on November 3 | Yes | No (promised by November 4) |
License | Mistral Ecosystem Open-Weight License | Not yet published for K3.1 Pro |
Base license family | Modified MIT | Kimi K3 License (modified MIT style) |
Model-as-a-Service condition | Separate commercial license for any MaaS business | Separate agreement above $20M revenue in 12 months (K3 terms) |
Attribution condition | None stated | "Kimi K3" display above 100M MAU or $20M monthly revenue (K3 terms) |
Published checkpoints | FP8, NVFP4, community GGUF | K3 base in MXFP4 |
........
Both licenses fall under the open-weight label and neither is open source in the strict sense, because each restricts some commercial uses. The Mistral condition applies to any hosting business regardless of size, while the Kimi K3 condition only bites above a revenue threshold.
··········
ACCESS TO MISTRAL LARGE 4 AND KIMI K3.1 PRO BY APP PLAN, API AND CLOUD PLATFORM.
Moonshot puts Kimi K3.1 Pro in every tier of its consumer apps. The Kimi pricing page lists limited K3.1 Pro use on the Free plan on the web and in Kimi Work, and full access from Plus at $15 a month, which adds Kimi Code. Pro at $31 adds the Highspeed tier, Max at $79 adds 1M-token chat capacity and higher allowances, and Ultra at $159 offers the highest allowances with ten times the agent credits. Moonshot quotes these as monthly prices, with annual billing shown at the effective monthly rate. In Kimi Enterprise, administrators must switch K3.1 Pro on, because access is off by default.
Mistral sells Mistral Large 4 mainly through its API and partners. The model is available in Mistral Studio, on the Mistral API, on OpenRouter, on NVIDIA-hosted endpoints and NIM containers, on Microsoft Foundry, which Microsoft says was the first cloud provider to offer it, and in Mistral's Vibe Code coding product. Mistral's consumer pricing page lists Vibe Pro at $14.99 a month and Team at $24.99 per user per month, but it does not name the model those plans use, so a subscription to Vibe is not a confirmed way to get Mistral Large 4.
........
Access route | Mistral Large 4 | Kimi K3.1 Pro |
Free consumer access | Not confirmed by Mistral | Limited use on Free (web, Kimi Work) |
Paid consumer plans | Vibe Pro $14.99, Team $24.99 per user (model not named) | Plus $15, Pro $31, Max $79, Ultra $159 per month |
Coding tool | Vibe Code | Kimi Code (Plus and above) |
First-party API | Mistral API, Mistral Studio | Kimi API Platform |
Third-party platforms | OpenRouter, Microsoft Foundry, NVIDIA endpoints | Five API providers tracked by Artificial Analysis |
Self-hosting | vLLM, SGLang, TensorRT-LLM, NIM | After weights release |
........
··········
HOW THE BENCHMARK FIGURES FOR MISTRAL LARGE 4 AND KIMI K3.1 PRO WERE PRODUCED.
Three kinds of numbers circulate for these models, and they should be kept apart. The first is each vendor's own launch chart. Mistral compared Mistral Large 4 mainly with other open-weight models (GLM-5.3, DeepSeek V4.1 Flash, Kimi K3) and with Claude Opus 5.5 in one human-preference test; it never measured Kimi K3.1 Pro, which came out six days before the final weights. Moonshot compared Kimi K3.1 Pro with GPT-6 Astra, Claude Opus 5.5, Claude Fable 5.1 and DeepSeek V4.1 Flash, and never with Mistral Large 4. Moonshot states its settings: max effort, temperature 1.0, top-p 0.95, and the Claude Code harness for Terminal-Bench 4.0. Mistral's launch post does not state the harness for its Terminal-Bench figure.
The second kind is Artificial Analysis, which runs the same nine evaluations of its Intelligence Index v5.0 on every model it tracks, with its own harness and settings. It is the only source in this comparison that measures both models under one method, so its results carry most of the weight below.
The third kind is the gap between the two. Terminal-Bench 4.0 shows it clearly: Mistral reports 41% for Mistral Large 4 while Artificial Analysis measures 25%, and Moonshot reports 65.7% for Kimi K3.1 Pro while Artificial Analysis measures 67%. Vendor and independent figures should not be mixed in a single ranking, and a vendor figure for one model should never be set against an independent figure for the other.
··········
[[ADPLUS_300x250_1]]
ARTIFICIAL ANALYSIS RESULTS FOR MISTRAL LARGE 4 AND KIMI K3.1 PRO SIDE BY SIDE.
Artificial Analysis lists Mistral Large 4 at its High reasoning setting and Kimi K3.1 Pro as a reasoning model without naming its effort level. The figures below come from its model comparison pages as published on November 3, 2026.
........
Artificial Analysis evaluation | Mistral Large 4 (High) | Kimi K3.1 Pro | Gap (points) |
Intelligence Index v5.0 | 26 | 46 | +20 Kimi |
AA-Briefcase v1.1 (Elo) | 1,350 | 1,676 | +326 Kimi |
AutomationBench-AA | 58% | 73% | +15 Kimi |
Terminal-Bench 4.0 | 25% | 67% | +42 Kimi |
FrontierCode 2.0 | 15% | 25% | +10 Kimi |
AA-Omniscience v1.1 | 5 | 30 | +25 Kimi |
GDP.pdf | 18% | 31% | +13 Kimi |
HLE-Diamond | 17% | 37% | +20 Kimi |
CritPt v2 (Hard Subset) | 26% | 45% | +19 Kimi |
Terminal-Bench-Science 0.1 | 4% | 34% | +30 Kimi |
........
The pattern is consistent across agentic, coding, knowledge and science tasks. The widest gap is on Terminal-Bench 4.0, the command-line agent benchmark, where Kimi K3.1 Pro solves 67% of tasks against 25% for Mistral Large 4. The narrowest is FrontierCode 2.0 at ten points. On AA-Omniscience, which rewards correct answers and penalizes confident wrong ones, Kimi's score of 30 against Mistral's 5 fits Moonshot's claim that K3.1 Pro cut its hallucination rate sharply compared with K3.
Mistral's own launch claim that Large 4 has the highest Intelligence Index score among Western open-weight models is consistent with these figures, because Kimi K3.1 Pro is a Chinese model. Hugging Face describes Mistral Large 4 as the fifth-best open-weight model on the index as of October 29. On its own model page, Artificial Analysis ranks Kimi K3.1 Pro 23rd of 60 models on the index, proprietary models included.
··········
VENDOR-PUBLISHED BENCHMARKS THAT ONLY ONE OF THE TWO MODELS REPORTS.
Most of the launch figures cannot be compared directly, because each company tested different benchmarks and different rivals. The table groups the main ones by vendor; none of them has a counterpart for the other model.
........
Benchmark (vendor-run) | Mistral Large 4 | Kimi K3.1 Pro |
Cognition FrontierCode 2.0 | 19.7% | Not published |
Artificial Analysis Coding Agent Index | 48% | Not published |
Terminal-Bench 4.0 (vendor harness) | 41% | 65.7% (Claude Code harness, max effort) |
BrowseComp | 84.5% | Not published as a number |
FinWorkBench (Finch) | 73.3% | Not published |
Lakera B3 attack resistance | 97.1% | Not published |
Cyber-request refusal average | 96.6% | Not published |
Kimi Code Bench 3.0 long-horizon (internal) | Not published | 61.2% |
OfficeQA Pro v2 | Not published | 64.4 |
Frontier Finance | Not published | 58.6 |
xdev-bench-hard cyber, pass@3 (internal) | Not published | 81.0 |
Kimi Hallu Bench, hard with tools (hallucination rate) | Not published | 8.3% |
........
Two vendor results set each model against Claude Opus 5.5. In Surge's blind human preference test for code, reported by Mistral, Mistral says its model ranks behind only Claude Opus 5.5 among the models evaluated. On Terminal-Bench 4.0 in Moonshot's own setup, Kimi K3.1 Pro scored 65.7% against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra. The Surge test also included Kimi K3, which Mistral Large 4 beat in Mistral's chart, but that result concerns the predecessor and says nothing direct about K3.1 Pro.
··········
TOKEN CONSUMPTION AND COST PER TASK ON THE ARTIFICIAL ANALYSIS SUITE.
A cheaper price per token only saves money if the model does not spend many more tokens to finish the same work. Artificial Analysis measures both: Mistral Large 4 generates 82,000 output tokens per task on average, 55,000 of them reasoning tokens, while Kimi K3.1 Pro generates 72,000, with 44,000 reasoning tokens. Mistral's model is about 14% more verbose per task, which narrows its price advantage without erasing it.
........
Artificial Analysis cost measure | Mistral Large 4 (High) | Kimi K3.1 Pro |
Output tokens per task | 82k | 72k |
Reasoning tokens per task | 55k | 44k |
Output tokens to run the full index | 107M | 113M |
Cost per task | $2.06 | $3.68 |
Cost to run the full Intelligence Index | $1,359 | $3,153 |
........
On the cost per task, Mistral Large 4 comes out 44% cheaper, and the full index costs 2.3 times as much to run on Kimi K3.1 Pro. The comparison has a limit: a task that Mistral Large 4 fails still costs money, and on the benchmarks where Kimi K3.1 Pro succeeds two or three times as often, the cost per successful task moves in Kimi's favour. Artificial Analysis does not publish a cost per solved task, and none is estimated here.
··········
WHAT THREE TYPICAL WORKLOADS COST ON MISTRAL LARGE 4 AND KIMI K3.1 PRO.
The three examples below apply the official list prices to the same token volumes for both models, computed in Python. They assume identical token counts for both models, which slightly favours Mistral Large 4 given its higher output per task on the Artificial Analysis suite. Uncached input is billed once at the standard input rate, which equals Kimi's five-minute cache-write rate, and no batch discount is applied.
The coding-agent session reads 2 million input tokens, 85% of them served from cache, and writes 150,000 output tokens. The document review processes 100 contracts of 200,000 tokens each with no caching and a 3,000-token summary per contract. The support chatbot handles 10,000 conversations a day for 30 days, each with 6,000 input tokens (70% cached) and a 500-token reply.
........
Workload | Mistral Large 4 | Kimi K3.1 Pro | Kimi K3.1 Pro Highspeed | Kimi vs Mistral |
Coding-agent session (2M in, 85% cached, 150k out) | $0.64 | $2.44 | $4.54 | 3.8x |
Review of 100 contracts (20M in, 300k out) | $14.63 | $43.00 | $86.00 | 2.9x |
Support chatbot, 300,000 conversations a month | $781.20 | $2,832.00 | $5,412.00 | 3.6x |
........
The ratio is highest in the agent and chatbot cases, where output tokens and cached input make up a larger share of the bill, because Kimi's output price is 4.8 times Mistral's. On Moonshot's batch API, which charges 60% of the standard price, the coding-agent session would cost $1.46 on Kimi K3.1 Pro, still more than twice the Mistral list price. Mistral has not published a batch rate for Large 4, so no Mistral batch figure is given.
··········
SPEED, LATENCY AND SELF-HOSTING HARDWARE FOR MISTRAL LARGE 4 AND KIMI K3.1 PRO.
Kimi K3.1 Pro is slow on Moonshot's standard API. Artificial Analysis measures 31.2 output tokens per second, 194th of 196 models, with a time to first token of 3.43 seconds, 67.43 seconds to the first answer token after reasoning, and an average of 1,956 seconds per index task. Moonshot sells the Highspeed tier for latency-sensitive use at double the price and says it runs two to three times faster; Artificial Analysis has not published a separate measurement for it.
Artificial Analysis has no public speed data for Mistral Large 4 on Mistral's API. The only published throughput figures come from Hugging Face's own deployment tests: on Inference Endpoints with eight NVIDIA H200 GPUs and FP8 weights, it reports about 0.6 seconds to the first token for a 1,000-token prompt and about 65 output tokens per second, rising to 120 to 138 tokens per second on chat prompts where its speculative-decoding draft head is accepted. These are self-hosted measurements and do not describe Mistral's hosted service.
Hardware is where the gap in active parameters shows. Mistral Large 4 activates 49 billion parameters per token and runs on a single eight-GPU node in Mistral's vLLM example, and Hugging Face shows an NVFP4 deployment on four B300 GPUs and a 2-bit quantization on one Apple Silicon machine with 512 GB of unified memory. Kimi K3.1 Pro's weights are not out yet; its K3 base activates 104 billion of 2.8 trillion parameters, more than twice the total size of Mistral's model, and will need correspondingly larger clusters.
··········
CYBERSECURITY SAFEGUARDS AND DATA HANDLING IN THE TWO MODELS.
Both companies shaped their open-weight releases around offensive cyber risk, in different ways. Moonshot removed cyber-related data from mid-training and reinforcement learning for Kimi K3.1 Pro, which lowered its score on the internal xdev-bench-hard test from 90.5% to 81.0% at pass@3, and it says it does not want an open-weight release to extend the frontier of cyber capability. Mistral kept Large 4's cyber skills, reporting 82% on a vulnerability reproduce-and-patch test and 93% on Cybench in the preview, and relied on refusal training instead: it reports a 96.6% average refusal rate on malicious requests across JailbreakBench, StrongREJECT and AgentHarm and 97.1% attack resistance on Lakera's B3 benchmark. Mistral also runs a separate cybersecurity variant in preview with vetted enterprise partners.
On data handling, Moonshot offers Zero Data Retention to Kimi API customers. Mistral says it plans deployments in several regions, including a European deployment it operates end to end, and offers private-cloud and on-premise options. Mistral Large 4 was trained from scratch on European infrastructure, a relevant point for buyers under European procurement or data-residency rules. With open weights, either model can also be run entirely on a company's own hardware, which today is possible only for Mistral Large 4.
··········
GAPS IN THE PUBLISHED EVIDENCE FOR MISTRAL LARGE 4 AND KIMI K3.1 PRO.
Several figures a buyer would want are missing from the official pages. Mistral does not publish a maximum output length, a batch price or a hosted-API speed for Mistral Large 4, and its documentation still marks the API model as a public preview without a date for general availability. Its consumer pricing page does not say which plans run Large 4. Parameter counts also vary slightly across Mistral's own pages, from 1T to 1.05T total and 49B to 52B active, depending on whether embeddings and output layers are counted.
For Kimi K3.1 Pro, Moonshot does not publish a parameter count, an architecture description or a license, and it has not released the weights at the time of writing. Its 14-evaluation launch chart gives numbers in the text for only a handful of benchmarks, and the effort level Artificial Analysis used is not shown on its pages. No independent evaluator has yet tested the Highspeed tier. Neither company has measured the other's model, so the Artificial Analysis pages remain the only like-for-like source.
··········
CHOOSING BETWEEN MISTRAL LARGE 4 AND KIMI K3.1 PRO BY WORKLOAD.
........
Workload or buyer | Recommended model | Reason |
Long-horizon coding agents in the terminal | Kimi K3.1 Pro | 67% vs 25% on Terminal-Bench 4.0 (Artificial Analysis) |
Research, science and expert questions | Kimi K3.1 Pro | 37% vs 17% on HLE-Diamond, 45% vs 26% on CritPt |
Factual answers where hallucinations are costly | Kimi K3.1 Pro | 30 vs 5 on AA-Omniscience |
High-volume chat and support | Mistral Large 4 | About 3.6x cheaper in the chatbot example |
Batch document processing on a budget | Mistral Large 4 | $14.63 vs $43.00 for 100 contracts at list price |
Self-hosting today | Mistral Large 4 | Weights downloadable, 49B active, one 8-GPU node |
European data control | Mistral Large 4 | Trained in Europe, European deployment, on-premise |
Consumer app subscription | Kimi K3.1 Pro | Included from the $15 Plus plan, limited on Free |
Hosting the model as a paid API service | Depends on revenue | Mistral needs a license at any size; Kimi K3 terms apply above $20M |
........
For readers who pay per token and run many simple or medium tasks, Mistral Large 4 delivers the lower bill at $0.70 and $2.10 per 1M tokens and a 44% lower cost per Artificial Analysis task, with the weights available to download now. For agentic coding, long research tasks and work where a wrong answer is expensive, Kimi K3.1 Pro's 20-point lead on the Intelligence Index and its lead on every listed evaluation justify paying $2.00 and $10.00, provided the 31-tokens-per-second speed of the standard tier is acceptable or the Highspeed tier fits the budget. Our coverage of the Mistral Large 4 open-weight release and of the Kimi K3.1 Pro launch has the full launch details, and the weekly API pricing roundup places Kimi's price cut against the other October launches.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
