Skip to content

Blog

How to use the Kimi K3.1 Pro API: key, snippets, and unlimited access

Kimi K3.1 Pro is live on CheapestInference: Moonshot’s new flagship — scoring 55 on the Artificial Analysis Intelligence Index (v4.3.2), #4 of every model AA tracks, 3 points off the leader (Claude Opus 5.5, 58) and one behind Claude Sonnet 5.5 at the same list price — served in the Flagship Pool next to Qwen3.8 Max (it replaced Kimi K3 in place on November 6, 2026), through an OpenAI- and Anthropic-compatible API with no token caps at a flat monthly price, from $199/mo (live availability — seats are very limited). This is the practical guide: get access, call it, understand its reasoning settings, wire it into your coding agent.

Two realistic routes to K3.1 Pro over an API today:

  1. Per-token — Moonshot’s list price is $2.00 per 1M input tokens ($0.20 cached) and $10.00 per 1M output. Ideal for evaluation and light use; it adds up fast for agent workloads, especially with reasoning on.
  2. Time-block subscription (this guide) — reserve one or more daily 8-hour blocks on the Flagship Pool and use K3.1 Pro with no token caps during your hours. From $199/mo per block ($169.15/mo billed annually); all three blocks = 24/7. The same subscription also serves Qwen3.8 Max, and existing kimi-k3 setups keep working until December 7, 2026.

For the subscription route: create an account, subscribe to the Flagship Pool, and mint an API key at cheapestinference.com/keys. Model id: kimi-k3.1-pro.

curl:

Terminal window
curl https://api.cheapestinference.com/v1/chat/completions \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "kimi-k3.1-pro", "reasoning_effort": "low", "max_tokens": 8192, "messages": [{"role": "user", "content": "Hello"}]}'

OpenAI SDK (Python):

from openai import OpenAI
client = OpenAI(
base_url="https://api.cheapestinference.com/v1",
api_key="sk-...", # your subscriber key
)
response = client.chat.completions.create(
model="kimi-k3.1-pro",
reasoning_effort="low", # low / high / max — omit for a direct answer with no reasoning
max_tokens=16000, # leave room for the reasoning when it is on
messages=[{"role": "user", "content": "Refactor this function..."}],
)
print(response.choices[0].message.content)

Anthropic SDK (Python) — the same key works on the Anthropic Messages endpoint:

import anthropic
client = anthropic.Anthropic(
base_url="https://api.cheapestinference.com/anthropic",
api_key="sk-...", # your subscriber key
)
message = client.messages.create(
model="kimi-k3.1-pro",
max_tokens=16000,
messages=[{"role": "user", "content": "Review this diff..."}],
)
print(message.content[-1].text)

Input is text and image — images as base64 data URLs in OpenAI-style image_url parts or Anthropic base64 image blocks; public http(s) image URLs are not accepted by this model, so download and inline the image (Qwen3.8 Max, in the same pool, takes both). The served context window is 1M tokens (1,048,576), with a maximum request size of 8 MB.

Reasoning works exactly as it did on Kimi K3, so existing K3 integrations carry over unchanged:

  • Off unless you ask. With nothing set (or "reasoning_effort": "none") the model answers directly, with no reasoning tokens. Turn it on per request with "thinking": {"type": "enabled"} or with reasoning_effort. (Moonshot’s own API documents K3.1 Pro as always reasoning; here the default is off, as it was for K3.)
  • reasoning_effort sets how much it thinks: low, high or max. Enabled without an effort it runs at max — the setting behind its benchmark scores. For interactive use (chat, code completion, quick edits) we recommend low: noticeably faster, and still the same model.
  • Give it a generous max_tokens. Reasoning tokens count toward the output budget, and K3.1 Pro is verbose by measurement — a tight cap truncates the answer mid-thought. Stream (stream: true) for responsive UIs.
  • Sampling parameters are fixed. temperature, top_p, presence_penalty and frequency_penalty are ignored if you send them, so existing client configs keep working unchanged.

Details: Reasoning · Kimi K3.1 Pro API docs.

The API speaks the Anthropic Messages format, so Claude Code works out of the box:

Terminal window
export ANTHROPIC_BASE_URL="https://api.cheapestinference.com/anthropic"
export ANTHROPIC_AUTH_TOKEN="sk-..." # your subscriber key
export ANTHROPIC_MODEL="kimi-k3.1-pro"
export ANTHROPIC_SMALL_FAST_MODEL="kimi-k3.1-pro"

Start claude as usual — every request now runs on K3.1 Pro with no per-token meter. Pointing the small/fast model at Kimi K3 (same pool, same key) keeps Claude Code’s background calls quick, since K3 can answer without reasoning. The same key works in OpenCode, Cline, Roo Code, Continue, and any client that accepts a custom OpenAI base URL — use https://api.cheapestinference.com/v1 and model kimi-k3.1-pro (Claude Code guide · OpenCode, Roo Code, Aider · Cline).

At list price, a coding agent that burns 100M tokens a month on K3.1 Pro costs roughly $150–300 per-token, depending on your cache-hit and output mix — and with reasoning on at max the output share climbs: Artificial Analysis measured 150M output tokens to run its index on K3.1 Pro, against a median of 87M. Light or cache-heavy users can come in under $199, and for them per-token is the cheaper route. Heavy agentic users go far beyond 100M. A Flagship block at $199/mo ($169.15/mo annual) covers your working day; all three blocks cover 24/7. The trade-offs: during your reserved hours there are no token caps, each subscription runs one request at a time (fair use), and outside your blocks the key doesn’t serve. Full pricing detail: Plans & Limits · Kimi K3.1 Pro API docs · the full analysis: specs, benchmarks, license.

Seats on the Flagship Pool are very limited — when a block sells out it’s gone until someone leaves. Check live availability →


CheapestInference serves Kimi K3.1 Pro and Qwen3.8 Max (Flagship Pool), GLM 5.3 and MiniMax M3 (Frontier Pool) and DeepSeek V4.1 Flash and MiMo V2.6 Flash (Core Pool) through one OpenAI- and Anthropic-compatible API on time-block subscriptions with no token caps. See the pools or get started.

Kimi K3.1 Pro: specs, benchmarks, pricing & API options

Kimi K3.1 Pro is Moonshot AI’s new flagship, released on October 23, 2026, and the headline is a number we have not been able to write about an open-weights model before: Artificial Analysis measures it at 55 on its Intelligence Index (v4.3.2) — #4 of every model it tracks, 3 points off the index leader (Claude Opus 5.5, 58), one point behind Claude Sonnet 5.5 at the same $2 / $10 list price, and 9 points above what was, until its own weights shipped, the best open-weights model on the index (MiMo-V2.6-Pro, 46).

Update — November 5, 2026: the weights are out. Moonshot published the full Kimi K3.1 Pro weights on Hugging Face on November 4, the day it promised (moonshotai/Kimi-K3.1-Pro) under the Kimi K3.1-Pro License, a Modified-MIT-style text that permits commercial use and serving outright (details below). It is now the top open-weights model on the Artificial Analysis index, 3 points off the closed leader. Our status is unchanged: confirmed for the Flagship Pool, with capacity validation the one remaining gate — live status.

Update — November 6, 2026: Kimi K3.1 Pro is live in the Flagship Pool, replacing Kimi K3 in place next to Qwen3.8 Max, on the same time-block subscriptions from $199/mo with no token caps during your reserved hours. Model id kimi-k3.1-pro (kimi-k3 is served by K3.1 Pro until December 7, 2026); 1M-token context as served, text and image input (base64), reasoning on demand (reasoning_effort low / high / max; off unless enabled). Kimi K3.1 Pro API docs → · How to use the Kimi K3.1 Pro API → · Subscribe →

At release the “open-weights” part was a promise, not a download: Moonshot said the full weights would ship by November 4, 2026, and they landed on November 4, the day promised (see the update above). Until then K3.1 Pro was API-only. It is a post-trained Kimi K3 — same 2.8T-parameter MoE, same 1M-token context, “substantially larger post-training compute” in Moonshot’s words — released alongside a K3.1 Pro Highspeed variant with the same capabilities at 2–3× the output speed.

And to answer the question this blog exists to answer: yes, we serve it. Since November 6, K3.1 Pro is live in the Flagship Pool, where it replaced Kimi K3 in place next to Qwen3.8 Max — the weights and a license that permits commercial serving both landed on November 4, and capacity was validated two days later. Model id, limits and snippets are on the Kimi K3.1 Pro API docs page; the step-by-step is How to use the Kimi K3.1 Pro API.

ArchitectureMixture-of-Experts, ~2.8T total parameters (“3T-class” in Moonshot’s wording) — a post-trained Kimi K3, not a new base model
Context window1M tokens (1,048,576); no long-context price surcharge
InputText, image and video; text output
ReasoningReasoning model — thinking is always on. reasoning_effort accepts low, high and max (default max); Moonshot’s docs state that requests with thinking disabled are rejected with HTTP 400
Variantskimi-k3.1-pro · kimi-k3.1-pro-highspeed (same capabilities, 2–3× the output speed, 2× the price)
Open weightsPublished November 4, 2026 on Hugging Face — 2.8T MoE / 104B active, MXFP4 weights (quantization-aware training), MoonViT-V2 vision encoder — under the Kimi K3.1-Pro License (commercial serving permitted; see below). Cyber-related data was removed from the open-weight training stages
List price (API)$2.00 in / $10.00 out per 1M, cached input $0.20 — Highspeed $4.00 / $20.00. Kimi K3 was $3.00 / $15.00
Measured speed52.2 output tok/s, 2.97s to first token on Moonshot’s API (Artificial Analysis)

Two things are new relative to K3 beyond the scores. The price went down a third for a better model ($2/$10 vs $3/$15 — the first time a Kimi flagship has been cheaper than its predecessor), and the reasoning configuration got stricter on Moonshot’s own API: K3 let you turn thinking off for fast direct answers; K3.1 Pro, as Moonshot documents it, does not (served here, reasoning stays on demand — off unless you enable it, as on K3).

Moonshot’s own reported numbers (vendor harness — not independently reproduced), against the two closed models it chose to compare with. The third row is a hallucination rate, so lower is better there:

Kimi K3.1 ProClaude Opus 5.5GPT-6 Astra
Terminal-Bench 4.0 BenchCAD (tools) Hallucination rate (lower = better) Kimi K3.1 Pro — Terminal-Bench 4.0: 65.7 Claude Opus 5.5 — Terminal-Bench 4.0: 66.4 GPT-6 Astra — Terminal-Bench 4.0: 58.2 Kimi K3.1 Pro — BenchCAD with tools: 92.7 Claude Opus 5.5 — BenchCAD with tools: 96.2 GPT-6 Astra — BenchCAD with tools: 95.9 Kimi K3.1 Pro — Kimi Hallu Bench Hard, claim-level hallucination rate with tools: 8.3% (lower is better) Claude Opus 5.5 — Kimi Hallu Bench Hard, with tools: 15.0% (lower is better) GPT-6 Astra — Kimi Hallu Bench Hard, with tools: 5.5% (lower is better) 65.766.458.2 92.796.295.9 8.3%15.0%5.5%
Benchmark (Moonshot, vendor-run)Kimi K3Kimi K3.1 ProClaude Opus 5.5GPT-6 Astra
Terminal-Bench 4.0—65.766.458.2
Kimi Code Bench 3.0 (long-horizon, internal)55.9561.2——
BenchCAD, with tools—92.796.295.9
BenchCAD, without tools—72.873.0—
OfficeQA Pro v2—64.4 (claimed state of the art)——
Kimi Hallu Bench Hard, with tools (claim-level hallucination rate, lower = better)23.4%8.3%15.0%5.5%
Internal finance evals: Financial Modeling · Financial Research · Business Operation50.4 · 58.8 · 8.957.6 · 70.4 · 37.0——

Read the pattern rather than any single cell. On the agentic benchmarks Moonshot leads with, K3.1 Pro lands within a point of Claude Opus 5.5 (Terminal-Bench 4.0, BenchCAD without tools) and clearly ahead of GPT-6 Astra on terminal work; on tool-augmented CAD it trails both closed models by 3–4 points. Against its own predecessor the jumps are large everywhere Moonshot measured them — and the hallucination number is the one to hold loosely: it is Moonshot’s own adversarial stress test (the company itself says it does not reflect everyday error rates), but a drop from 23.4% to 8.3% on the same test is a real change in behaviour. Terminal-Bench scores also move by tens of points with scaffolding, so the 7.5-point gap to GPT-6 Astra is a direction, not a measurement.

The independent signal, from Artificial Analysis on Intelligence Index v4.3.2 — every score below re-checked today on that edition (AA rescored the whole field in September, so none of these numbers compare with scores quoted from earlier editions):

Kimi K3.1 ProServed in a CheapestInference poolReference models (not served here)
Claude Opus 5.5 Claude Sonnet 5.5 Kimi K3.1 Pro GPT-6 Astra · Fable 5.1 MiMo-V2.6-Pro GLM-5.3 · Qwen3.8 Max Kimi K3 Claude Opus 5.5 (max): 58 — current index leader, $4/$20 per 1M Claude Sonnet 5.5 (max): 56 — $2/$10 per 1M Kimi K3.1 Pro: 55 — $2/$10 per 1M; open weights published 2026-11-04 GPT-6 Astra (max) and Claude Fable 5.1 (max): 53 — $10/$50 per 1M MiMo-V2.6-Pro: 46 — top open-weights model on the index until K3.1 Pro's weights shipped; under evaluation here GLM-5.3 (max) — Frontier Pool — and Qwen3.8 Max — Flagship Pool: 45 Kimi K3 (max) — Flagship Pool until 2026-11-06: 44 5856 5553 4645 44
ModelAA Intelligence Index v4.3.2List price in / out per 1MWeights
Claude Opus 5.5 (max)58$4 / $20closed
Claude Sonnet 5.5 (max)56$2 / $10closed
Kimi K3.1 Pro55$2 / $10open (Kimi K3.1-Pro License)
GPT-6 Astra (max) · Claude Fable 5.1 (max)53$10 / $50closed
MiMo-V2.6-Pro46$0.43 / $0.87open (MIT)
GLM-5.3 (max) · Qwen3.8 Max45$1.40 / $4.40 · $2 / $6open
Kimi K3 (max)44$3 / $15open

Scores from the current index edition; prices are vendor list prices for the variants AA evaluates. Full field and history in our monthly LLM Pareto Frontier report.

The honest read. K3.1 Pro does not beat the closed frontier: Claude Opus 5.5 leads at 58, and Claude Sonnet 5.5 edges it by one point at the identical $2/$10 list price — which is why, on our Pareto chart, K3.1 Pro sits one point under the frontier line rather than on it. What is remarkable is the column on the right. Every model within 3 points of the top of the index is closed except this one — and since November 4 this one is downloadable. The open-vs-closed gap on the index went from 12 points (MiMo-V2.6-Pro’s 46 against 58) to 3 that day, and the best open-weights score on the current index jumped 9 points in a single release, more than the 6-point jump K3 itself delivered in July.

Two more facts from the same AA page keep the enthusiasm calibrated: K3.1 Pro is verbose (150M output tokens to complete the index, against a median of 87M — expect that at $10/M), and it is not fast (52.2 tok/s, in the lower range for its price tier). The Highspeed variant exists for the second problem, at twice the price.

Kimi K3Kimi K3.1 ProK3.1 Pro Highspeed
Input, per 1M tokens$3.00$2.00$4.00
Cached input, per 1M$0.30$0.20$0.20
Output, per 1M tokens$15.00$10.00$20.00
AA Intelligence Index4455— (same model, faster serving)

At K3’s launch we called it “the first open flagship priced like a closed mid-tier model”. K3.1 Pro keeps that bracket and moves the bar: $10 per million output tokens for an index score that only two closed models beat. The per-token shape is still the one where a flat fee pays off fastest — reasoning-heavy when you turn it on, verbose by measurement, 1M context — which is exactly the profile the Flagship Pool exists for: since November 6, K3.1 Pro is served there in K3’s place with no token caps from $199/mo.

The weights: published November 4, under the Kimi K3.1-Pro License

Section titled “The weights: published November 4, under the Kimi K3.1-Pro License”

Moonshot’s commitment was full weights by November 4, 2026; they went public on Hugging Face on November 4, the day promised — not gated, as MXFP4 safetensors from quantization-aware training (2.8T MoE / 104B active, 93 layers, MoonViT-V2 vision encoder, 1M context). The card lists text and image input for the open release.

The license is the part that matters for anyone who serves models, and it is better news than the K3 precedent suggested. K3 shipped under Moonshot’s own Kimi K3 License; K3.1 Pro ships under the Kimi K3.1-Pro License, which reads like the K2 line’s Modified MIT: use, copy, modify, distribute, sell, run and deploy, all granted without restriction, with three conditions — (1) keep the notice; (2) a Model-as-a-Service operator whose aggregate revenue exceeds US$20M over any 12 months must sign a separate agreement with Moonshot before commercial use (the definition explicitly excludes end-user products that merely embed the model and “mere relaying of requests to models hosted by others”); (3) products above 100M monthly active users or US$20M in monthly revenue must display “Kimi K3.1-Pro” on their UI. Internal use and access through Moonshot’s own products or certified inference partners are exempt from (2) and (3). For a flat-rate API like ours that is a green light; the row is in our State of Open Weights report.

One detail worth noting for anyone evaluating models to serve: Moonshot removed cyber-related data from the open-weight training stages, and published the cost — on its xdev-bench-hard cyber benchmark the final K3.1 Pro scores 81.0 (pass@3) where a checkpoint without that removal scored 90.5. That is a deliberate capability trade made before the weights go public, and a first for a frontier-class open release.

Kimi K3.1 Pro is served in the Flagship Pool since November 6, 2026, where it replaced Kimi K3 in place next to Qwen3.8 Max — one subscription and one key cover both, switch by model id. Three things gated the launch, in order: the weights publishing — done November 4 — a license that permits commercial serving — done, the Kimi K3.1-Pro License — and capacity validation for a 2.8T-parameter model — done November 6. Existing keys keep working: requests with kimi-k3 are served by K3.1 Pro until December 7, 2026. As served: model id kimi-k3.1-pro, 1M-token context (1,048,576), 8 MB maximum request, text and image input (images as base64), reasoning on demand (reasoning_effort low / high / max — off unless enabled; send low for interactive use), sampling parameters fixed. Setup and snippets: Kimi K3.1 Pro API docs · How to use the Kimi K3.1 Pro API · changelog.

What is Kimi K3.1 Pro? Moonshot AI’s flagship model, released October 23, 2026: a post-trained version of Kimi K3 (~2.8T-parameter MoE) with a 1M-token context window, text, image and video input, and reasoning that Moonshot’s own API keeps always on (served here, it is on demand). Artificial Analysis scores it 55 on its Intelligence Index (v4.3.2), #4 overall and 3 points off the leader. The full open weights were published on November 4, 2026.

How much does the Kimi K3.1 Pro API cost? Moonshot lists $2.00 per 1M input tokens, $0.20 for cached input and $10.00 per 1M output tokens — a third below Kimi K3’s $3 / $15, with no surcharge across the 1M context. The faster K3.1 Pro Highspeed variant is $4.00 / $20.00. Artificial Analysis measures $2.11 per Intelligence-Index task.

Is Kimi K3.1 Pro open-weights? Yes, since November 4, 2026: the full weights are on Hugging Face under the Kimi K3.1-Pro License, a Modified-MIT-style license that permits commercial use and serving (a separate agreement is needed only for Model-as-a-Service operators above US$20M in 12-month revenue, and UI attribution only above 100M MAU or US$20M monthly revenue). Before that date the model was available only through Moonshot’s API and apps.

How does Kimi K3.1 Pro compare to Kimi K3? Same architecture and context window, much more post-training. On the AA index it scores 55 against K3’s 44 (both on v4.3.2); on Moonshot’s own tests it gains 5 points on Kimi Code Bench long-horizon tasks, cuts the adversarial hallucination rate from 23.4% to 8.3%, and improves 7–12 points on internal finance evals. It is also cheaper per token ($2/$10 vs $3/$15), but thinking can no longer be switched off.

Is Kimi K3.1 Pro better than Claude Sonnet 5.5 or Opus 5.5? On the AA Intelligence Index, no — narrowly. Opus 5.5 scores 58 and Sonnet 5.5 scores 56 to K3.1 Pro’s 55, and Sonnet 5.5 costs the same $2/$10. On Moonshot’s vendor benchmarks K3.1 Pro is within a point of Opus 5.5 on Terminal-Bench 4.0 (65.7 vs 66.4) and BenchCAD without tools (72.8 vs 73.0). The difference the index does not show: K3.1 Pro’s weights are public, which no other model above 46 on the index can say.

What is Kimi K3.1 Pro Highspeed? The same model served 2–3× faster, at twice the price ($4 / $20 per 1M). Moonshot’s documentation describes identical capabilities; it exists because the standard model’s measured 52 tok/s is slow for interactive use.

Is there an unlimited Kimi K3.1 Pro API subscription? Yes, since November 6, 2026: Kimi K3.1 Pro is served in the Flagship Pool from $199/mo ($169.15/mo billed annually) per daily 8-hour block — flat monthly fee, no token caps during your reserved hours, OpenAI- and Anthropic-compatible, and it works in Claude Code. It replaced Kimi K3 in place (kimi-k3 keeps working until December 7, 2026), and the same subscription covers Qwen3.8 Max. Seats are very limited. Kimi K3.1 Pro API docs · setup guide.


CheapestInference serves Kimi K3.1 Pro and Qwen3.8 Max (Flagship Pool), GLM 5.3 and MiniMax M3 (Frontier Pool) and DeepSeek V4.1 Flash and MiMo V2.6 Flash (Core Pool) through one OpenAI- and Anthropic-compatible API on unlimited time-block subscriptions. See the pools or get started.

DeepSeek V4.1 Flash-1017: what changed, and how to use it over the API

On October 17, DeepSeek shipped V4.1-Flash-1017 — not a new generation, an updated checkpoint of V4.1 Flash. Same architecture (552B MoE, causal encoder–decoder), same 1M-token context, same native image input, same price bracket — but refreshed post-training, and DeepSeek reports gains on all eleven benchmarks it publishes, with the biggest jumps on terminal and program-automation agent work.

If you use the Core Pool, there is nothing to migrate: we rolled the 1017 build out to the pool the same day, the model id is still deepseek-v4.1-flash, and your existing keys already serve it. Unlimited, flat-rate access starts from $22/mo ($18.70/mo billed annually) — live pricing on /pools.

Same model, new checkpoint. This is DeepSeek’s own reported before/after (vendor harness — no independent verification of these specific numbers yet):

V4.1-Flash-1017V4.1-Flash (September checkpoint)
Terminal-Bench 4.0 HLE (no tools) ProgramBench V4.1-Flash-1017 — Terminal-Bench 4.0: 42.3 V4.1-Flash September — Terminal-Bench 4.0: 31.2 V4.1-Flash-1017 — HLE (no tools): 42.6 V4.1-Flash September — HLE (no tools): 36.8 V4.1-Flash-1017 — ProgramBench: 29.1 V4.1-Flash September — ProgramBench: 20.3 42.331.2 42.636.8 29.120.3

Full vendor-stated table for the 1017 build, next to the September checkpoint it replaces:

BenchmarkFlash-1017Flash (Sep)
Terminal-Bench 4.042.331.2
HLE (no tools)42.636.8
ProgramBench29.120.3
Apex Bench v1.172.164.6
Automation-Bench v1.0.660.554.8
Agents’ Last Exam35.431.8
DeepSWE v1.177.074.2
NL2Repo-Bench68.265.4
Chartography (with tools)83.078.9
BabyVision (with tools)92.589.6
ZeroBench-main (with tools)51.049.0

The pattern mirrors what the 0731 retrain did for V4 Flash: the architecture and price stay put while agentic reliability — long terminal sessions, multi-step program edits, tool-driven research — takes the jump. The weights are open under MIT: deepseek-ai/DeepSeek-V4.1-Flash-1017.

  • Nothing to change. The Core Pool serves the 1017 build since October 17 under the same id, deepseek-v4.1-flash, on the same subscriptions and keys.
  • Same flat rate. Time-block subscriptions are unaffected — your cost does not scale with tokens, so a stronger checkpoint at the same price is pure upside.
  • Same limits. 1M-token context and image input are unchanged; see the model page for current specifics.
Terminal window
curl https://api.cheapestinference.com/v1/chat/completions \
-H "Authorization: Bearer sk-..." \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-v4.1-flash",
"messages": [{"role": "user", "content": "What build are you?"}]
}'

It also works over the Anthropic-style endpoint (/anthropic/v1/messages) and with any OpenAI- or Anthropic-compatible SDK or agent — point the base URL at https://api.cheapestinference.com and use your subscriber key. Setup guides: getting started, DeepSeek V4.1 Flash model page.

Is there an unlimited DeepSeek V4.1 Flash-1017 API plan? Yes — the Core Pool serves it on unlimited, flat-rate 8-hour daily time blocks from $22/mo ($18.70/mo billed annually), with no token caps.

MiMo-V2.6 Pro & Flash: specs, benchmarks, pricing & API options

Update — September 23, 2026: MiMo V2.6 Flash is now served in the Core Pool as mimo-v2.6-flash, replacing MiMo v2.5 in place — same pool, same price, subscriptions uninterrupted; the old id mimo-v2.5 keeps working until October 23, 2026. Model page: MiMo V2.6 Flash API. MiMo-V2.6-Pro remains under evaluation (live status).

Xiaomi’s MiMo team shipped two models on September 21, 2026, and each one matters for a different reason:

  • MiMo-V2.6-Pro is the new #1 open-weights model on the Artificial Analysis Intelligence Index: 46, ahead of GLM-5.3 and Qwen3.8 Max (45) and Kimi K3 (44), at a blended $0.18 per 1M tokens.
  • MiMo-V2.6-Flash is the direct successor to MiMo v2.5: the same size class (309B total / 15B active) and the same list price, with agent-benchmark scores that land within a few points of Pro.

Both ship MIT-licensed weights, a 1M-token context window and omnimodal input (text, image, audio, video). A naming note for searchers: this is MiMo-V2.6, September 2026. It is not the older MiMo-V2-Flash from late 2025, which had a 256K context and is a different model.

MiMo-V2.6-ProMiMo-V2.6-Flash
Architecture1.02T total / 42B active MoE309B total / 15B active MoE, hybrid sliding-window + global attention
DecodingReasoning modelReasoning model, 5-layer multi-token prediction (7 tokens per pass)
InputText, image, audio, videoText, image, audio, video
Context window1M tokens, up to 128K output1M tokens, up to 128K output
Open weightsXiaomiMiMo/MiMo-V2.6-Pro-RL, MIT, FP8XiaomiMiMo/MiMo-V2.6-Flash-RL, MIT, BF16 + FP8
List price (API)$0.435 in / $0.87 out per 1M, cached input $0.0036$0.14 in / $0.28 out per 1M, cached input $0.0028
Measured speed124.5 output tok/s, 2.34s to first token (AA)Not yet measured independently

The “-RL” suffix in the repository names confuses people. These are full checkpoints from the final reinforcement-learning run, not adapters. Pro and Flash are also separate training runs with their own benchmark tables, not one model cut down to two sizes.

These are Xiaomi’s own reported numbers from its vendor harness, not independently reproduced. The baseline is last generation’s largest model, MiMo-V2.5-Pro:

MiMo-V2.5-Pro (previous generation)MiMo-V2.6-Flash (15B active)MiMo-V2.6-Pro (42B active)
Terminal-Bench 2.1 DeepSWE v1.1 CyberGym MiMo-V2.5-Pro — Terminal-Bench 2.1: 65.2 MiMo-V2.6-Flash — Terminal-Bench 2.1: 87.6 MiMo-V2.6-Pro — Terminal-Bench 2.1: 89.9 MiMo-V2.5-Pro — DeepSWE v1.1: 19.0 MiMo-V2.6-Flash — DeepSWE v1.1: 67.9 MiMo-V2.6-Pro — DeepSWE v1.1: 71.9 MiMo-V2.5-Pro — CyberGym: 40.0 MiMo-V2.6-Flash — CyberGym: 95.1 MiMo-V2.6-Pro — CyberGym: 94.0 65.287.689.9 19.067.971.9 40.095.194.0
Benchmark (Xiaomi, vendor-run)MiMo-V2.5-ProMiMo-V2.6-FlashMiMo-V2.6-Pro
Terminal-Bench 2.165.287.689.9
DeepSWE v1.119.067.971.9
CyberGym40.095.194.0
AutomationBench16.052.353.1
MiMo Code Bench40.461.263.2
JobBench—61.262.0
MiMo Visual Coding—71.572.3

The gap between the two new models is small. Flash is within 4.0 points of Pro on every benchmark, and it beats Pro on CyberGym. The generational jump is large: on DeepSWE, Flash scores 67.9 where last generation’s Pro scored 19.0. Xiaomi credits scaled RL for this: about 750,000 trajectories across more than 7,000 environments, and it published the runs live while they trained. That jump is the headline, and it is also the reason to wait for independent reproductions before trusting every point.

The independent signal comes from Artificial Analysis, using Intelligence Index v4.3.2. Every score below was checked today on that edition. AA rescored the whole field in this edition, so these numbers are not comparable with scores quoted from earlier editions.

MiMo-V2.6-ProServed in a CheapestInference poolClosed frontier reference
GPT-6 Astra · Fable 5.1 MiMo-V2.6-Pro GLM-5.3 Qwen3.8 Max Kimi K3 DeepSeek V4.1 Flash GPT-6 Astra: 53 — top of the index, tied with Claude Fable 5.1 MiMo-V2.6-Pro: 46 — #1 open-weights model, blended $0.18/1M (AA) GLM-5.3 (max) — Frontier Pool: 45, blended $0.90/1M (AA) Qwen3.8 Max — Flagship Pool: 45, blended $1.18/1M (AA) Kimi K3 (max) — Flagship Pool: 44, blended $2.31/1M (AA) DeepSeek V4.1 Flash — Core Pool: 39, blended $0.18/1M (AA) 5346 4545 4439
ModelAA Intelligence Index v4.3.2AA blended $/1MAA output tok/s
GPT-6 Astra / Claude Fable 5.153——
Claude Opus 551—56.4
MiMo-V2.6-Pro46$0.18124.5
GLM-5.3 (max)45$0.9061.0
Qwen3.8 Max45$1.1839.2
Kimi K3 (max)44$2.3140.6
DeepSeek V4.1 Flash39$0.18225.6

Blended = AA’s 7:2:1 cache-hit/input/output mix. The full field and its history are in our monthly LLM Pareto Frontier report.

The honest read. Pro does not beat the closed frontier: GPT-6 Astra and Claude Fable 5.1 lead at 53, 7 points ahead. Its lead over the next open-weights models is a single point, which is within the noise of any index. The notable part is the rest of the row. Against the three open-weights models it edges out:

  • it is 5–13× cheaper on AA’s blended price;
  • it is about 2–3× faster on measured output speed.

It costs the same blended $0.18 as DeepSeek V4.1 Flash while scoring 7 points higher. That combination puts Pro on AA’s intelligence-vs-cost Pareto frontier at $0.13 per Index task.

MiMo-V2.6-Flash: a new generation at the old price

Section titled “MiMo-V2.6-Flash: a new generation at the old price”

Flash has no AA score yet. Here is why it may be the more important release for high-volume workloads:

  • The cost profile is unchanged. Flash keeps the exact list price of MiMo v2.5 ($0.14 in / $0.28 out, cached input $0.0028) and the same 15B-active footprint. Agent loops, sub-agents and long-context summarization get a generation’s worth of capability at the same per-token cost.
  • It lands close to Pro. On Xiaomi’s own table, Flash sits within 4.0 points of the 1T Pro everywhere. At the size class, this is the part to check once independent numbers land.
  • 1M context stays usable. The hybrid sliding-window + global attention keeps long inputs affordable to serve, and the 7-token multi-token-prediction head is built for fast decoding.

Pro and Flash both ship under plain MIT on Hugging Face. Pro is in FP8; Flash is in BF16 and FP8. Alongside them, Xiaomi published:

  • a technical report;
  • more than 7,000 RL environments;
  • its end-to-end RL framework;
  • a 9B distilled model.

The full task datasets have not been released yet. For anyone evaluating models to serve rather than just call, the top of the open-weights index now carries the most permissive license there is.

MiMo-V2.6-Flash is served. Since September 23, 2026 it is the Core Pool’s MiMo model, as mimo-v2.6-flash, next to DeepSeek V4.1 Flash, unlimited on flat monthly time blocks. It replaced MiMo v2.5 in place: same pool, same price, nothing to change on existing keys, and requests with the old id mimo-v2.5 are served by V2.6 Flash until October 23, 2026. Image input works on the same endpoint; reasoning is off by default and switched on per request. MiMo-V2.6-Pro is under evaluation — its live status is on the models-under-review page, and any addition lands in the changelog the day it ships.

What is MiMo-V2.6? Xiaomi’s September 2026 model generation, in two open-weights models. MiMo-V2.6-Pro is a 1.02T-total / 42B-active MoE. MiMo-V2.6-Flash is a 309B-total / 15B-active MoE. Both have a 1M-token context window, omnimodal input (text, image, audio, video) and MIT-licensed weights. Xiaomi also offers a Pro-UltraSpeed variant on its own API, advertised at up to 20× Pro’s output speed.

What is the difference between MiMo-V2.6-Pro and MiMo-V2.6-Flash? Size and price. Pro is about 3× larger in active parameters and costs about 3× more per token ($0.435 / $0.87 vs $0.14 / $0.28 per 1M). On Xiaomi’s vendor benchmarks, Flash sits within 4.0 points of Pro on every task and beats it on CyberGym (95.1 vs 94.0).

Is MiMo-V2.6-Pro the best open-weights model? On the Artificial Analysis Intelligence Index v4.3.2 it is #1 among open-weights models, at 46. That is one point ahead of GLM-5.3 and Qwen3.8 Max (45) and two ahead of Kimi K3 (44). It is also the cheapest and fastest of that group: $0.18/1M blended and 124.5 tok/s. GPT-6 Astra and Claude Fable 5.1 lead the overall index at 53.

How much does the MiMo-V2.6 API cost? Xiaomi lists MiMo-V2.6-Pro at $0.435 input / $0.87 output per 1M tokens, with cached input at $0.0036. MiMo-V2.6-Flash is $0.14 / $0.28, with cached input at $0.0028, which is the same as MiMo v2.5.

Does MiMo-V2.6 have open weights? Yes. Both are on Hugging Face under the MIT license: XiaomiMiMo/MiMo-V2.6-Pro-RL and XiaomiMiMo/MiMo-V2.6-Flash-RL. The “-RL” suffix marks the final RL checkpoint; these are full models, not adapters.

Is MiMo-V2.6-Flash the same as MiMo-V2-Flash? No. MiMo-V2-Flash is Xiaomi’s late-2025 model with a 256K context. MiMo-V2.6-Flash is the September 2026 successor to MiMo v2.5, with a 1M context and the RL-scaled post-training described above.

Is there an unlimited MiMo-V2.6 API subscription? Yes, for Flash. Since September 23, 2026, MiMo V2.6 Flash is served unlimited in the Core Pool as mimo-v2.6-flash: flat monthly fee, no token caps during your reserved hours, OpenAI- and Anthropic-compatible, image input included. It replaced MiMo v2.5 in place, and the old id mimo-v2.5 keeps working until October 23, 2026. MiMo-V2.6-Pro is under evaluation (live status) and is not served today.


CheapestInference serves Kimi K3 and Qwen3.8 Max (Flagship Pool), GLM 5.3 and MiniMax M3 (Frontier Pool) and DeepSeek V4.1 Flash and MiMo V2.6 Flash (Core Pool) through one OpenAI- and Anthropic-compatible API on unlimited time-block subscriptions. See the pools or get started.

Asking a model "who are you?" proves nothing. Here is what the evidence says.

Every week someone sends a provider the same two screenshots: the request asked for model A, the reply says “I am model B”. The conclusion feels obvious, and it is usually wrong. A subscriber sent us this pair this week, verbatim:

// request
{ "model": "deepseek-v4.1-flash",
"messages": [{ "role": "user",
"content": "Who are you? Who made you? What's your model name, give me a precise model name" }] }
// response
{ "model": "deepseek-v4.1-flash",
"choices": [{ "message": { "role": "assistant",
"content": "I'm DeepSeek, an AI assistant created by the Chinese company DeepSeek (深度求索).\n\nMy precise model name is **DeepSeek-V3**." } }] }

Same subscriber, other runs: the model called itself an Anthropic model. Same key, same endpoint, same served model every time. The model was not lying and the platform was not switching anything. Language models simply do not know what they are, and the research on this is now large enough to put numbers on it.

Study / reportDateWhat was testedResult
TechCrunch: “Why DeepSeek’s new AI model thinks it’s ChatGPT”Dec 2024DeepSeek V3, asked what it isClaimed to be ChatGPT (GPT-4) in 5 of 8 generations; gave OpenAI API instructions when asked about DeepSeek’s API
”I’m Spartacus, No, I’m Spartacus” (Shandong Univ., Drexel, UMass Lowell)Nov 202427 LLMs, systematic identity probing25.93 % exhibit identity confusion; output-distribution analysis attributes it to hallucination, not copied weights
Same paper, documented real-world cases2024Production modelsGemini-Pro says it is Baidu’s Wenxin when asked in Chinese; GPT-4 over the API says it is GPT-3; ByteDance Seed says it is GPT
”Know Thyself? On the Incapability and Implications of AI Self-Recognition” (Univ. of Chicago)Oct 202510 frontier models (GPT-4.1, GPT-5, Claude Sonnet 4, Gemini 2.5 Flash, Kimi K2, DeepSeek V3, GLM-4.5, Qwen3-235B, Grok 4), 20,000 predictionsExact self-identification accuracy 10.3 – 10.9 % against a 10 % random baseline. GPT and Claude receive 97.7 % of all attributions while producing 40 % of the text. GLM-4.5 identified itself as Claude in half of its runs

The Chicago study is the cleanest one: each model was shown text and asked which model wrote it, including its own text. Only four of the ten ever named themselves. The models that were most often named as authors were the ones with the most public visibility, not the ones that actually wrote the sample.

Identity lives in the system prompt, not in the weights. When a vendor’s own chat app says “I am X, made by Y”, that sentence comes from instructions the vendor injects in front of every conversation. Over a raw API, with no system prompt, the model falls back on whatever its training data says about “AI assistants”, and that corpus is dominated by a few famous names.

Training corpora contain other models’ output. The DeepSeek V3 case was traced to exactly this: public datasets full of GPT-4 generated text. Mike Cook, research fellow at King’s College London, described the effect to TechCrunch as “taking a photocopy of a photocopy”. Synthetic data and distillation are standard practice across the industry, so this is not a property of one vendor; the Chicago paper found the same bias in every family it tested.

Version numbers are the least reliable part. A model trained before its own release cannot know its final product name. The Chicago authors report GPT-5 dismissing “gpt-5” as a fake model name because, from the model’s point of view, it had not been released yet. A DeepSeek model calling itself “V3” is the same failure: it names the latest version it read about.

The model field is written by the server. The name in the JSON response is metadata attached by the API, not something the model produced. It tells you what the endpoint claims to have served. It cannot be cross-checked by asking the model, because the model never sees it.

Signals that are more reliable than the model’s own answer

Section titled “Signals that are more reliable than the model’s own answer”

None of these is proof on its own, and each takes some work. All of them beat a “who are you?” prompt, because they measure behaviour the model cannot talk its way around:

  • Tokenizer behaviour. Every model family counts tokens differently, and the tokenizers of open-weights models are public. The usage.prompt_tokens reported for a fixed input is a fingerprint of the tokenizer that actually processed it.
  • Knowledge and capability boundaries. Training cutoffs, supported languages, context length behaviour and native modalities (vision, audio) differ between models in ways that a system prompt cannot fake.
  • Behaviour with structured features. How a model handles tool calls, JSON mode, reasoning fields, stop sequences or unusual parameters is specific to the model and to the software serving it.
  • Response metadata patterns. ID formats, error message wording, header sets and latency profiles tend to be stable per serving stack and change when the stack changes.
  • Consistency across many samples. A single answer at default temperature is noise. The studies above used hundreds or thousands of samples per model; a verification worth trusting does the same and looks at the distribution.
  • Independent, reproducible attestations. Published test harnesses with recorded runs let a third party repeat the measurement. That is how the September 2026 CrofAI investigation established that a reseller was silently routing sixteen model ids to four cheaper models: tool signatures, response ID formats, token counting and canary strings, recorded in public CI runs, and not one “who are you?” prompt.

If a model tells you it is ChatGPT, Claude, or an older version of itself, you have learned something about its training data and nothing about what is being served to you. Treat self-identification the way the research does: as a hallucination category with a measured rate of roughly one model in four, and at chance level when the question gets specific.

When it matters, measure behaviour instead. The signals above are the ones the published investigations relied on, and the ones we would point anyone to, including for our own endpoints.