Kimi K3.1 Pro vs K3: What Changed, What It Costs, and Who Should Switch

Kimi K3.1 Pro is the first update to Kimi K3, the 2.8-trillion-parameter open-weight model Moonshot AI shipped on 16 July 2026. It launched on 23–24 October 2026 in the Kimi app, Kimi Code and the Kimi API. Moonshot's launch post calls it "the first model in our new K3.1 series and our most capable 3T-class open-weight model to date": a post-trained version of K3, trained with much more post-training compute across "hundreds of downstream domains". Moonshot claims gains in coding, 3D creation, and financial research and modeling. A K3.1 Pro Highspeed variant has identical capabilities at about two to three times the speed, and the full weights are due by 4 November 2026.

The short answer: for API code, the switch is a model-ID change. kimi-k3.1-pro has the same parameter constraints as kimi-k3, so fixed sampling, the three effort levels and the max default all carry over. The change you'll notice first is quieter. In Kimi Code, K3 ran at high effort by default, and K3.1 Pro runs at max. Expect more thinking per task, and so more tokens or more of your plan's quota per task.

On price, K3.1 Pro is cheaper per token. AIHubMix lists it at $2.00 in and $10.50 out per million tokens, against $2.84 and $14.20 for K3, and a yuan price table circulating on launch day shows a similar cut at Moonshot. Moonshot's own USD list price couldn't be confirmed at the time of writing. Moonshot's self-reported benchmarks show clear gains over K3 on its own coding, hallucination and finance tests. So the real question is whether those gains, at a lower per-token price, still pay off for a model that thinks longer on your work.

K3 and K3.1 Pro side by side

The K3 column comes from Moonshot's Kimi K3 guide and model card. The K3.1 Pro column comes from Moonshot's K3.1 Pro quickstart and its Kimi Code model docs, which also give both Kimi Code defaults. The Highspeed IDs are kimi-k3.1-pro-highspeed on the API and k3.1-pro-highspeed in Kimi Code. In Kimi Code, the full 1M window comes with the higher plan tiers. Moonshot hasn't published K3.1 Pro's active-parameter count or said which licence its weights will use.

How K3.1 Pro surfaced

The K3.1 series leaked in pieces over a month. CometAPI's tracker compiled the leaks, and Wccftech, AIBase and posters on LINUX DO reported them. Moonshot confirmed none of them before its teaser video.

• 23 September: a config snippet posted on X showed a k3d1-agent identifier, three effort levels (Low, High, Max), an "Extra Long" context option up to 1M tokens, and flags for agent, swarm, search and batch modes.
• 28 September: developers found kimi-k3-1 in the model registry of Moonshot's API platform and reported that it answered calls. The open platform's preview listed the same price as K3.
• 21 October: Kimi Code's model_switch_confirm config began returning a rule that moves k3, K3, kimi-k3, k3-256k, kimi-for-coding and kimi-for-coding-highspeed to k3.1. Two days later the model shipped in Kimi Code as k3.1-pro.
• 22 October: an anonymous model called Kibitz appeared in Arena's Code Arena. A tokenizer fingerprint tied it to Kimi, and the name follows Moonshot's pattern (K3 was tested as Kivine). Testers' one-shot battles were favourable, and one rated a 3D build near Claude Opus 5.5 at medium effort. Arena published no score.
• 23 October: Kimi's official account posted a 24-second teaser. It shows a Kimi Code terminal reading "Model: Kimi K3.1" while the model writes a Three.js 3D scene.
• 23–24 October: Moonshot added K3.1 Pro and K3.1 Pro Highspeed to its Kimi Code docs, which say the model is "now fully rolled out in Kimi Code." The launch post went live briefly and was pulled, then came back that night with the full benchmark table, alongside API docs for kimi-k3.1-pro and kimi-k3.1-pro-highspeed.

Most of the leaked flags described things K3 already did: three effort levels, a 1M context window and Kimi's existing swarm mode. The launch bore that out. K3.1 Pro keeps K3's request contract; what's new is the extra post-training, the max-effort default in Kimi Code and the Highspeed twin.

What stays the same, and what to re-check

Moonshot says kimi-k3.1-pro has "the same parameter constraints as kimi-k3", so code that runs on K3 needs only the new ID. The rules that carry over:

• Thinking is always on. Moonshot says requests that disable it are rejected with HTTP 400; AIHubMix's main IDs didn't reject them (see below).
• reasoning_effort takes low, high or max and defaults to max. There's no medium level, despite a pre-launch report of one: Kimi Code maps medium to high, xhigh and ultra to max, and minimum and light to low. In Kimi Code, off, none, disabled or an unknown value returns a 400.
• Sampling is fixed: temperature 1.0, top_p 0.95, n 1, and both penalties at 0. Leave them out; other values return an error on Moonshot's API.
• tool_choice accepts required, as on K3.
• max_completion_tokens defaults to 131,072 and goes up to 1,048,576, within a 1M context window.
• Changing the model or the effort level invalidates the cache. In Kimi Code, Moonshot advises against switching to K3.1 Pro inside an existing K3, K2.8 Preview or K2.7 Code Highspeed session, because output quality drops and cache hits fall.

As on K3, multi-turn and tool loops should pass back the full assistant message, including reasoning_content.

What AIHubMix measured. On 24 October, AIHubMix tested both models under their main IDs, kimi-k3.1-pro and kimi-k3.1-pro-highspeed, and under its coding channel, coding-kimi-k3.1-pro and coding-kimi-k3.1-pro-highspeed, which follows Kimi Code's effort mapping:

• Chat Completions (with reasoning_content) and Messages (thinking and text blocks) returned 200 on both. The Responses API returned 200 only on the coding channel; the main IDs returned a 404, "The model kimi-k3.1-pro is not available on this endpoint." So at launch, Codex, which uses Responses, needed the coding-channel ID on AIHubMix.
• Every reasoning_effort value tried (low, medium, high, max, xhigh) returned 200. The gateway accepts them all, but Moonshot documents only low, high and max, so don't read the extra values as extra levels. off and an unknown value returned 200 as well, and none or a thinking: {"type": "disabled"} field returned 200 with no reasoning at all: on AIHubMix, a thinking-off switch fails quietly, not with a 400.
• temperature=0.2 returned a 400 on the main IDs: "invalid temperature: only 1 is allowed for this model". The coding channel accepted it. A leftover K2.x thinking: {"type": "enabled"} field returned 200 on both.
• Structured output over the Messages API now works. A json_schema request whose prompt didn't ask for JSON came back as {"city":"Paris"}. In AIHubMix's July test, K3 had silently ignored the schema there.
• Caching is automatic: on the coding channel, a repeated prompt came back with the whole prompt counted as cached tokens.

The migration post in this series takes each of these as a failure you can hit, with the error text and the fix.

Benchmarks: against K3 and the tier above

Figures are Moonshot's self-reported launch numbers, from its K3.1 Pro launch post and, for K3, its GitHub model card. Harnesses differ by model (K3.1 Pro ran mostly in a Claude Code harness), and nobody has reproduced the numbers yet.

K3 to K3.1 Pro. These are the pairs Moonshot measured the same way on both models:

Kimi Code Bench, Kimi Hallu Bench and the two finance sets are Moonshot's internal benchmarks. The Kimi Code Bench 2.0 pair was run in a Claude Code harness (K3's headline 72.9 was on Kimi Code), and Hallu Bench Hard is run with tools. Every pair moved the right way: between 3 and 8 points on the coding sets, 7 and 12 points on the finance sets, and a hallucination rate on the hard set that fell from 23.4 to 8.3. Terminal-Bench, SWE Marathon and ProgramBench appear in both launches but under different versions or scoring, so K3's 88.3, 42.0 and 77.8 can't be set against K3.1 Pro's 65.7, 51.7 and 31.5.

Against the tier above. Kimi has nothing above K3.1 Pro, so the comparison that matters is with the closed models in Moonshot's table. Moonshot ran them at max effort and notes that "proprietary model testing may involve some fallback behavior".

On Moonshot's own scorecard, K3.1 Pro leads all four on long sustained coding (SWE Marathon v1.1) and on HLE with tools, and edges ahead on DeepSWE v1.1. On Terminal-Bench 4.0 it beats Fable 5.1 and GPT-6 Astra and lands within a point of Opus 5.5, and on GDPval-AA it sits just behind Opus 5.5. It trails clearly on ProgramBench, where Opus 5.5 scores about twice as high, on HLE without tools, and on Moonshot's internal Design Bench, where GPT-6 Astra leads. K3 opened at the top of Arena's Frontend Code leaderboard with 1,679; K3.1 Pro had no Arena score at launch.

Price and real cost

Prices are per million tokens. AIHubMix's K3.1 Pro price is 30% below its K3 price on input and 26% below on output; Highspeed costs exactly twice K3.1 Pro on input and output, with the same cache-hit price. Moonshot's own USD list for K3.1 Pro couldn't be confirmed at the time of writing, so the official column shows a dash. OpenAI is running a promotional $4 / $20 rate on GPT-5.6 Sol, reported to last at least until 21 November 2026. Cache hits on AIHubMix cost $0.20 per million for K3.1 Pro and Highspeed, against $0.284 for K3 ($0.30 at Moonshot). Since Moonshot's billing update, K3 cache writes appear as a separate line: $3.00 per million for the default 5-minute lifetime, $6.00 for one hour. On Moonshot's API, K3.1 Pro follows the same pattern: flat pricing across context length, cache writes billed by lifetime, and reasoning tokens billed as output at every effort level. AIHubMix's model pages list input, output and cache-hit prices only, with no separate cache-write rate.

Moonshot's reported yuan list. A price table circulating on launch day shows Moonshot's Chinese list for K3.1 Pro and K3.1 Pro Highspeed, including cache writes. It hasn't been confirmed against Moonshot's pricing page:

Per million tokens, in CNY. The K3 row matches Moonshot's Chinese list. If the table holds, Moonshot cut K3.1 Pro by 35% on input, 32% on output and 35% on cache hits, a little more than AIHubMix's cut. Highspeed costs twice as much as K3.1 Pro on input, output and cache writes, with the same cache-hit price, which puts it 30% above K3 on input and 36% above on output. Moonshot's USD and CNY lists aren't tied by a fixed rate, so the yuan figures don't give its dollar price.

One task, costed. In July, AIHubMix asked K3 to build a single-file HTML game from one prompt and a reference image. The run took 2,541 seconds (about 42 minutes) and produced 1,275 lines of working code. It spent 74,994 completion tokens, and 54,486 of them (73%) were thinking. It ran at max effort, the only level K3 accepted at the time and K3.1 Pro's default today. The input size wasn't recorded, so the estimates below assume 5,000 input tokens. Prices are AIHubMix's.

• K3: 5,000 × $2.84/M + 74,994 × $14.20/M ≈ $1.08
• GPT-5.6 Sol, same token counts: ≈ $1.52
• Claude Fable 5.1, same token counts: ≈ $4.18
• K3.1 Pro, same token counts: 5,000 × $2.00/M + 74,994 × $10.50/M ≈ $0.80
• K3.1 Pro Highspeed, same token counts: 5,000 × $4.00/M + 74,994 × $21.00/M ≈ $1.59

Same token counts is an assumption, and real counts move the result both ways. In a hands-on build test by Chase AI, K3 used 21.5 million tokens where Claude Fable 5 used 3.5 million, about six times as many. At that ratio, Fable 5 would finish the game in about 12,200 tokens for about $0.73, less than K3's $1.08 and K3.1 Pro's $0.80, even though its output price is nearly four times higher. Across Artificial Analysis's index tasks, the gap ran the other way: Fable 5.1 averaged 78,000 output tokens per task to K3's 48,000, and cost $7.63 per task to K3's $2.00.

How long a model thinks depends on the task. In a small AIHubMix spot check on 24 October, a 600-word explainer prompt at low effort made K3.1 Pro reason for 5,561 to 7,356 tokens before answering, and K3 for 3,013 to 7,051: similar amounts. The difference was speed. K3.1 Pro streamed about 86 tokens a second against K3's 22 to 26, so it answered in 80 to 99 seconds where K3 took 154 to 309. That's one prompt and three runs per model on AIHubMix's main IDs, not a benchmark. Still, a quarter off the output price is a quarter off the bill only if K3.1 Pro thinks no longer than K3 does on your work, so measure tokens per task before you count the saving.

Where K3.1 Pro is not the obvious pick

• Latency-bound work. K3.1 Pro always thinks, defaults to max, and doesn't support turning thinking off. K3's game run took 42 minutes in one request, and in the spot check above, even at low effort, K3.1 Pro's first answer token came after 71 to 91 seconds (K3's after 132 to 280). Highspeed makes output about 2–3× faster, but Moonshot says it "only speeds up model output", so tool calls and scripts run as before. If users wait on each answer, a faster non-reasoning model fits better, even at the same quality.
• Pipelines that need temperature control. K3.1 Pro keeps K3's fixed sampling at 1.0, and on AIHubMix's main IDs temperature 0.2 returns a 400. The coding channel accepted it, but a 200 doesn't show the value was applied. If you rely on temperature 0 for repeatable outputs, K3.1 Pro won't give you that.
• Work that looks like its weak benchmarks. On Moonshot's own table it trails Claude Opus 5.5 by half on ProgramBench, trails all three closed models on HLE without tools, and trails GPT-6 Astra on Design Bench. Test those kinds of tasks before switching.
• Tasks where the tier above thinks less. As the Chase AI build shows, a pricier model that thinks less can cost less per task. Measure your own workload before assuming K3.1 Pro is the cheaper option.
• Lower Kimi Code plans. K3.1 Pro needs Plus or above on the new plans (Moderato or above on the legacy ones), and its 1M window needs Pro (Allegretto). Highspeed needs Pro or above (Allegretto).

Who should switch, who should wait

• Kimi Code users: K3.1 Pro is live as k3.1-pro. Start a new session with it rather than switching an existing K3, K2.8 Preview or K2.7 Code Highspeed session; Moonshot says a mid-session switch lowers output quality and cache hits. Expect max effort by default where K3 ran at high, and watch your quota for the first week. If it drains faster, drop routine work to high, but don't expect low to be cheap: in AIHubMix's spot check, K3.1 Pro at low still reasoned for 5,561 to 7,356 tokens.
• K3 API users: the switch is kimi-k3 to kimi-k3.1-pro, with the same parameter constraints. On Moonshot's own API it unlocks after a $1 top-up; AIHubMix serves it under the same ID. Run a one-day eval on your own tasks first, and compare completion and thinking tokens, not just pass rates. Moonshot's own USD list wasn't confirmed at the time of writing; on AIHubMix, K3.1 Pro is $2.00 / $10.50 per million.
• K2.6 and K2.7 Code users: the jump is bigger, and Moonshot has set a deadline. On Moonshot's platform, K2 series models are closed to newly registered users from 31 October 2026, and Moonshot plans to retire all of them by the end of December 2026. AIHubMix still listed K2.6 and K2.7 Code on 24 October, with no retirement date. K3.1 Pro doesn't support thinking off, and reasoning is set by effort level rather than K2.x's thinking switch. Read the migration post before you change anything.
• If speed matters more than price: K3.1 Pro Highspeed has the same capabilities at about 2–3× the speed, for about twice the quota in Kimi Code and twice the per-token price ($4.00 / $21.00 on AIHubMix). The Highspeed post in this series weighs that trade.
• Wait if you buy from Moonshot directly and want to see its USD list before committing.
• Stay put if your workload is short, high-volume requests where always-on thinking is overhead you don't need.

To compare K3 and K3.1 Pro on your own prompts, run both through AIHubMix with one key. The quick start covers keys and base URLs, and the K3.1 Pro model page lists the price. Use Chat Completions or Messages with kimi-k3.1-pro; for the Responses API, use coding-kimi-k3.1-pro for now, which AIHubMix labels an evaluation preview, not for production.

import os
from openai import OpenAI

client = OpenAI(base_url="https://aihubmix.com/v1", api_key=os.environ["AIHUBMIX_API_KEY"])
prompt = "Write a single-file HTML page that renders a rotating 3D cube with Three.js."

for model in ["kimi-k3", "kimi-k3.1-pro"]:
    r = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": prompt}],
        timeout=3600,
    )
    print(model, r.usage.prompt_tokens, r.usage.completion_tokens)

The script uses Chat Completions and leaves out temperature: K3 and K3.1 Pro reject any value other than 1.

FAQ

Is Kimi K3.1 Pro out?
Yes. Moonshot launched it on 23–24 October 2026 in the Kimi app, Kimi Work, Kimi Code and the Kimi API, as kimi-k3.1-pro on the API and k3.1-pro in Kimi Code. Business and Enterprise admins have to enable it for their teams.

Does K3.1 Pro cost more than K3?
Per token, no. AIHubMix lists it at $2.00 input and $10.50 output per million tokens, against $2.84 and $14.20 for K3. A yuan table circulating on launch day shows a cut of about a third at Moonshot, but Moonshot's own USD list wasn't confirmed at the time of writing. It still thinks a lot, though: max is the default in Kimi Code, and in one AIHubMix spot check it reasoned for 5,561 to 7,356 tokens even at low effort. The cost per task can still rise.

Will my K3 code work on K3.1 Pro?
Yes, with the model ID changed. Moonshot says kimi-k3.1-pro has the same parameter constraints as kimi-k3. Keep leaving out temperature and the other sampling parameters, keep passing back reasoning_content in multi-turn requests, and don't try to turn thinking off.

What changes in Kimi Code?
Pick K3.1 Pro in a new session rather than mid-session, as Moonshot's docs advise. Its default effort is max, where K3's was high. It needs the Plus plan or above, and the full 1M window needs Pro or above.

What is K3.1 Pro Highspeed?
The same model capabilities at about two to three times the output speed, as kimi-k3.1-pro-highspeed on the API. In Kimi Code it uses about twice the quota, and per token it costs twice as much: $4.00 / $21.00 on AIHubMix. Only the model's output gets faster; tool calls and scripts don't.

Is K3.1 Pro as good as Claude Fable 5.1, Claude Opus 5.5 or GPT-6 Astra?
On Moonshot's own table it leads all three on SWE Marathon and HLE with tools, beats Fable 5.1 and GPT-6 Astra on Terminal-Bench 4.0, and trails on ProgramBench, HLE without tools and Design Bench. The numbers are self-reported, harnesses differ, and nobody has reproduced them yet.

Will K3.1 Pro have open weights?
Yes. Moonshot says the full weights will be released by 4 November 2026. It hasn't said which licence they'll use; K3's weights are under the Kimi K3 licence.

Keep reading: the Kimi K3.1 Pro series

• If you've decided to switch, this is where the 400s and the quiet bill increases come from, from the max-effort default to a mistyped Highspeed ID: Migrating to Kimi K3.1 Pro: Fix the 400s, Mind the Max-Effort Default, and Don't Mistype Highspeed
• If you're weighing K3.1 Pro against alternatives rather than upgrading, here it is against Qwen 4, which surfaced in Arena the same week, and Claude Fable 5.5, the model both are chasing: Kimi K3.1 Pro vs Qwen 4 vs Claude Fable 5.5: Front-End Quality, Cost per Task, and the Hype Checked
• If the 2–3× speed tempts you, this one works out when Highspeed is worth paying twice the per-token price: Kimi K3.1 Pro vs K3.1 Pro Highspeed: Is 2–3× the Speed Worth 2× the Price?

Sources

• Moonshot AI: Kimi K3.1 Pro tech blog
• Moonshot AI: Kimi K3.1 Pro quickstart
• Moonshot AI: Kimi Code model configuration
• AIHubMix: Kimi K3.1 Pro model page
• Chase AI: Kimi K3 vs Claude Fable 5 vs GPT-5.6, real cost