Kimi K3.1 Pro costs Sonnet money. Open K3 sells 62% cheaper

Short answer: Don't switch to Kimi K3.1 Pro to save money yet. Moonshot charges $2 in / $10 out per million tokens, exactly what Anthropic charges for Claude Sonnet 5.5, and Sonnet scores higher on the Artificial Analysis Intelligence Index (56.0 vs 55.1). The discount only shows up once the weights are public: Kimi K3, the open model before it, is sold by 21 providers on OpenRouter today, and the cheapest one charges 62% less than Moonshot itself. Moonshot says the K3.1 weights land by Nov 4.

TL;DR: I priced every model on OpenRouter (469 listed, 215 with an intelligence score) and every provider that serves them. At every intelligence tier where open-weight models exist, the cheapest open model costs 3x to 11x less than the median closed one. Above an index of 50 there is no open model yet, and K3.1 Pro walks in at closed-model prices with one seller. If you want a cheap model today, GLM-5.3 scores 44.8 for $0.41 a million blended tokens, about a tenth of K3.1 Pro's $4.00. Checked Oct 24, 2026, 7 AM ET. Every row is in the CSV.

Key numbers

What	Number	Source, checked Oct 24 2026
Kimi K3.1 Pro price (Moonshot, standard)	$2.00 in / $10.00 out per 1M tokens ($4.00 blended)	OpenRouter endpoints API
Claude Sonnet 5.5 price	$2.00 / $10.00 ($4.00 blended)	OpenRouter
Intelligence index: Sonnet 5.5 vs K3.1 Pro	56.0 vs 55.1	Artificial Analysis, as listed by OpenRouter
Providers selling K3.1 Pro	1 (Moonshot; Highspeed tier $4 / $20)	OpenRouter endpoints API
Providers selling Kimi K3 (open weights)	21; 13 of them under Moonshot's price	OpenRouter endpoints API
Kimi K3: Moonshot vs cheapest seller	$6.00 vs $2.28 blended (62% less, Sail Research, fp4)	OpenRouter endpoints API
Providers per model, median	open 3, closed 1	215 models, OpenRouter
Cheapest open vs closed, index 40 to 50	$0.45 vs $3.00 blended (6.7x)	median of cheapest seller per model
Open models scoring above 50	0 (K3.1 Pro weights promised "by Nov 4")	Moonshot, Hugging Face (no K3.1 repo yet)
Best value near the top: GLM-5.3	index 44.8, $0.41 blended, 37 providers	OpenRouter

Download the data (CSV, 215 rows: every model with its intelligence score, listed price, number of providers, cheapest provider and its quantization, median endpoint price and the lab's own price. Blended = 3 parts input to 1 part output.)

Why this came up today

Moonshot's Kimi K3.1 Pro launch post is the biggest thread on Hacker News right now (927 points, 550 comments when I checked). The pitch: 65.7% on Terminal-Bench 4.0, "narrowly behind Claude Opus 5.5 (66.4%)", a 1-million-token context window, and "full model weights will be released by Nov 4, 2026".

Half the replies read it as "open source caught up, closed labs are cooked". The other half asked the boring question that actually decides whether you switch: what does it cost? So I stopped reading takes and pulled the price list for the whole market.

How I priced 215 models in 4 minutes

OpenRouter is a marketplace that sells hundreds of models from dozens of providers through one API. Its public models endpoint lists every model with its price, context window, and two fields that matter here: hugging_face_id (filled in when the weights are public, so it is a clean open vs closed label) and the Artificial Analysis Intelligence Index, a composite score of many benchmarks. Then a second endpoint lists every provider serving a given model, with its own price and quantization.

curl -s https://openrouter.ai/api/v1/models | jq '.data | length'
curl -s https://openrouter.ai/api/v1/models/moonshotai/kimi-k3/endpoints \
  | jq '.data.endpoints[] | [.provider_name, .pricing.prompt, .pricing.completion, .quantization]'

I pulled the full catalog (469 models), kept the 215 that carry an intelligence score, and fetched the provider list for each one, about one request a second. To compare models with one number I used a blended price: 3 parts input to 1 part output, the same weighting Artificial Analysis uses. Most real apps send more tokens than they get back, so this is closer to a real bill than the output price alone. Try your own ratio in our free LLM API cost calculator.

Kimi K3.1 Pro is priced like a closed model

Put every model on one chart, score across, price up, and the picture is blunt. The open-weight dots (green) sit a whole order of magnitude below the closed ones at the same score. Kimi K3.1 Pro (orange) does not sit with the green dots. It sits in the closed-model cluster at the top right, at $4.00 blended, the same as Sonnet 5.5, Gemini 4 Argon and GPT-6.1 Sol at list price.

That is not a scandal, it's a business decision. But it kills the "open model, so it's cheap" reflex for the next 11 days. On price per point of intelligence, K3.1 Pro ($0.073) is a hair worse than Sonnet 5.5 ($0.071) and far better than Opus 5.5 ($0.139). If you are on Sonnet today, switching saves you nothing.

The open-weight discount is not a feature of the model. It's what happens when 20 strangers are allowed to sell it.

Open weights are 3x to 11x cheaper at every tier, until the top

Group the 215 models by score and take the median cheapest price in each group. Open beats closed everywhere open exists: 4.6x under an index of 20, 3.1x from 20 to 30, 11.3x from 30 to 40, and 6.7x from 40 to 50 ($0.45 vs $3.00).

Above 50 there are 8 closed models with a median of $6.00 and zero open ones. K3.1 Pro will be the first. Until Nov 4 it is an open model in name only: you can't download it, so nobody else can host it, so Moonshot is the only seller and sets the price.

The discount comes from strangers hosting the model

Here's the part I didn't expect to be this clean. Kimi K3 shipped in July. Its weights are public, so anyone with GPUs can serve it. Today 21 different providers sell K3 through OpenRouter. Moonshot's own endpoint charges $3 / $15 ($6.00 blended). 13 of the 21 undercut it, and the cheapest, Sail Research, charges $1.20 / $5.50: $2.28 blended, 62% less. The median seller is $5.15.

Zoom out to every model and the pattern holds. Open-weight models have a median of 3 providers (24 of them have 10 or more; GLM-5.3 has 37). Closed models have a median of 1, and when a closed model does have several sellers (Claude on Anthropic, Bedrock, Vertex and Azure), they all charge the same list price or more. Across the 40 open models whose own lab also sells them on OpenRouter, the cheapest seller undercuts the lab in 23 cases, by a median of 25.5%.

One honest catch: the cheapest sellers often run the model at lower precision. Sail Research's K3 is fp4 (4-bit numbers instead of the 8 or 16 the model was trained with), which is cheaper to serve and can cost some accuracy. Moonshot's own K3 endpoint is mxfp4 too, so for this model the gap is smaller than it sounds. The CSV has the quantization of every cheapest seller. If you need the full-precision model, filter by provider in your OpenRouter request, and if you want to know what it takes to run one yourself, our LLM VRAM calculator does the math.

Moonshot has been raising the price as the score climbs

Look only at Moonshot's own endpoints. K2.6 (April) and K2.7 Code (June) cost $0.95 / $4.00 with scores around 26 to 27. K3 (July) jumped to $3 / $15 at 43.6. K3.1 Pro (Oct 23) is $2 / $10 at 55.1. Twice the score, 2.5x the output price in six months.

The launch coverage (OrcaRouter, for one) calls K3.1 Pro a 33% price cut, and against K3's $3 / $15 it is. But it's priced against the closed labs, not against the open market. K3 itself is already cheaper than K3.1 Pro on most third-party hosts.

My read: this is the playbook. Launch closed-only at a frontier price, collect the API revenue from everyone who can't wait, then open the weights and let the hosting market race the price down. That's not shady, it's just the model. Nobody at Moonshot has said that's the plan; it's what the price list says.

What to use this week, by job

Your job	Pick	Index	Cheapest blended $ / 1M
Best score money can buy	Claude Opus 5.5	57.6	$8.00
Near-top score, sane price	Claude Sonnet 5.5 or Kimi K3.1 Pro (tie on price, Sonnet scores higher)	56.0 / 55.1	$4.00
Want Kimi, want it cheap	Kimi K3 via a third-party host	43.6	$2.28
Best value, open weights	GLM-5.3	44.8	$0.41
Cheap and decent, open weights	MiniMax M3.1 Flash	42.7	$0.19
Cheap and decent, closed	Claude Haiku 5.5	43.4	$0.20

Two of those surprised me. Haiku 5.5 is a closed model priced like an open one, so "closed = expensive" isn't a law either. And GLM-5.3 gets 81% of K3.1 Pro's score for about 10% of the price. For most agent loops and chatbots, that's the trade to test first. Before you switch anything, check how many of your calls are repeats: caching duplicate prompts cuts calls on any model, and the token counter shows what a prompt really costs.

My take, and a bet

"Open source caught up" is half true. The score caught up, nearly. The price catches up only when the weights do, and Moonshot decides when that is. For 11 days, K3.1 Pro is a closed model with an IOU attached.

The bet: by Nov 11, a week after the promised weights date, at least one third-party host will sell K3.1 Pro on OpenRouter for 40% or more below Moonshot's $4.00 blended. K3's cheapest host is 62% under Moonshot today, and 21 hosts already know how to serve this model family. (What it really takes to serve one of these giant models is in our DeepSeek inference-engine piece.) If the weights slip past Nov 4, that's the story instead. I'll rerun the same script on Nov 11 and update this page either way.

Your turn: is a 1-point gap on a benchmark index worth paying 10x for, or are you already running something like GLM-5.3 in production?

Caveats

• OpenRouter is one marketplace. Some providers sell cheaper direct, and big contracts get discounts nobody publishes.
• The intelligence index is one composite from Artificial Analysis. Moonshot's own launch numbers (Terminal-Bench, BenchCAD, OfficeQA) are separate and self-reported.
• "Cheapest" includes discounted tiers some labs list as separate endpoints (OpenAI and Google AI Studio show $1 / $5 endpoints for GPT-6.1 Sol and Gemini 4 Argon). In the text I quote list prices for those; the tier chart uses the cheapest listed endpoint for every model, open or closed.
• Cheapest open sellers often use fp4 or fp8 quantization. Quality can differ from the lab's endpoint.
• Blended price assumes 3 input tokens per output token. Reasoning models that think out loud spend far more output tokens, which favors cheap output prices.
• A one-day snapshot. Prices on OpenRouter move every week.

The numbers so far

Item	Today (Oct 24, 7 AM ET)	Rerun
K3.1 Pro sellers on OpenRouter	1	Nov 11
K3.1 Pro cheapest blended price	$4.00	Nov 11 (bet: $2.40 or less)
Kimi-K3.1 repo on Hugging Face	not public	Nov 4
Open models scoring above 50	0	Nov 11

Know your bill before the switch

Price per million tokens means nothing until it's your traffic. The LLM API Cost Calculator ($7 once, Google Sheets and Excel) turns your prompt size, reply size and requests a day into your bill per request, per month and per year across 26 models side by side, so you see what a switch saves before you make it. You get your AI bill per request, month and year for every model in it. If you don't, we keep working on it at no extra cost until you do. Email support@tensorust.com; a real person replies within one business day.

Follow for one checked, money-saving fix for AI builders a day.

Sources

• Moonshot AI, Kimi K3.1 Pro and K3.1 Pro Highspeed (benchmarks, 1M context, weights "by Nov 4, 2026")
• Hacker News discussion, 927 points, 550 comments, Oct 24 2026
• OpenRouter models API and per-model endpoints API, pulled Oct 24 2026, 7 AM ET
• Artificial Analysis Intelligence Index (George Cameron and Micah Hill-Smith's team), as listed by OpenRouter
• Moonshot AI on Hugging Face (Kimi-K3 public, no K3.1 repo as of Oct 24)
• Our dataset: data.csv