Within three weeks, three Chinese labs shipped three very different answers to the same question — what should a serious AI model cost, and who should be allowed to own it? DeepSeek's V4.1-Flash-1017 (October 17) is the efficiency play: open MIT weights and API prices measured in cents. Moonshot's Kimi K3.1 Pro (October 23) is the scale play: a reported 2.8 trillion parameters, a 1-million-token context, and open weights under a custom licence that landed this week. Meituan's LongCat-2.5 (November 2) is the endurance play: 1.6 trillion parameters trained for long agent runs, priced like a utility, and — for now — API-only.
All three undercut the closed Western frontier on price by an order of magnitude. All three claim coding crowns, on benchmarks chosen by their own makers. And they differ most where the marketing is quietest: the licence, the weights, and what happens when the launch discount ends. This comparison puts them side by side on the numbers that are actually public, flags every figure that is vendor-reported, and ends with the only verdict that matters — which one fits which job. We have covered each model in depth: the LongCat-2.5 guide, the Kimi K3.1 Pro guide and the DeepSeek V4.1-Flash guide.
The three models at a glance
| LongCat-2.5 | Kimi K3.1 Pro | DeepSeek V4.1-Flash-1017 | |
|---|---|---|---|
| Lab | Meituan | Moonshot AI | DeepSeek |
| Released | Nov 2, 2026 | Oct 23, 2026 | Oct 17, 2026 |
| Total parameters | 1.6T (~48B active) | 2.8T reported (104B active) | Flash-class (not headlined) |
| Context window | 1,048,576 tokens | 1M tokens | 1,048,576 tokens |
| Max output | 128K tokens | Not headlined | 384K tokens |
| Inputs | Text, image, video | Text, image | Text, image |
| API input / 1M | $0.30 ($0.006 cached) | $2.00 ($0.20 cached) | $0.15 off-peak ($0.003 cache hit) |
| API output / 1M | $1.20 | $10.00 | $0.60 off-peak |
| Weights | Not yet ("coming soon") | Out now (MXFP4, ~1.56 TB) | Out now (MIT) |
| Licence | None yet (API terms) | Kimi K3.1 Pro License (custom) | MIT |
Read the table's silences as carefully as its numbers. Only one of the three gives you the model outright under a permissive licence (DeepSeek, MIT). Only one is a true frontier-scale open release you can download today (Kimi, at 1.56 TB in a single 4-bit format). And the cheapest API of the three (LongCat, on headline rates) is the one you cannot download at all. Price, openness and scale form a triangle here — each model picks a different corner.
A note on method before the rounds begin. Every price in this comparison comes from the labs' published tariffs as recorded in our individual guides, checked against each other for internal consistency; every benchmark is labelled with the lab that reported it, because in this trio the source of a number is half its meaning. Licence terms for Kimi were read in the licence file itself, not summarised from coverage. Where a figure is a launch-window or reported rate rather than a settled tariff — LongCat's discount, Kimi's pricing as reported at launch — the text says so at the point of use. Models this new cannot be compared the way mature products are; the honest comparison is of their offers, their terms and their design intent, with capability claims held at arm's length until independent measurement catches up. That is the standard the rounds below apply, and it is why the verdict at the end is organised by job rather than by a single winner-takes-all score.
Round 1: Price — DeepSeek and LongCat leave Kimi behind
On API price there are two leagues. DeepSeek's off-peak rates — $0.15 per million input tokens on a cache miss, $0.003 on a cache hit, $0.60 per million output — are the lowest headline numbers of the three, with peak rates ($0.30/$1.20) that merely match LongCat's. LongCat-2.5's limited-time launch pricing ($0.30 input, $0.006 cached, $1.20 output, only successful calls billed) sits in the same league, and its flat structure — no peak/off-peak split, no length tiering — is simpler to budget than DeepSeek's two-clock tariff. Kimi K3.1 Pro plays a different sport: at a reported $2.00/$10.00 with $0.20 cached input, it costs roughly 7–17× more per token than the other two, in exchange for frontier-class capability — and even there, Moonshot's 33% price cut versus K3 is partly eaten by the new model generating about 14% more output tokens, leaving cost per finished task (per Artificial Analysis) essentially flat at $1.98 versus $2.00.
Two warnings apply to the whole round. LongCat's prices are an introductory discount with no published end date and no published list price behind it — the rate you build on may not be the rate you keep. And per-token prices mislead on agent work, where output volume and cache behaviour dominate the bill; DeepSeek's own cost-per-task chart (vendor-reported) puts V4.1-Flash at just under $0.20 per finished DeepSWE task, the kind of number that decides real budgets. Round to DeepSeek, on price you can verify today and weights that make the price permanent — you can always self-host your way out of a tariff change. LongCat is a close second while the discount lasts.
One workload, three bills
Per-token tables hide the shape of real usage, so here is one concrete workload priced three ways: a busy coding-agent deployment processing 40 million input tokens and generating 8 million output tokens in a month, with half the input served from cache — a realistic cache ratio for an agent that re-reads the same repository context every turn.
| Model | Fresh input | Cached input | Output | Monthly total |
|---|---|---|---|---|
| DeepSeek V4.1-Flash (off-peak) | 20M × $0.15 = $3.00 | 20M × $0.003 = $0.06 | 8M × $0.60 = $4.80 | $7.86 |
| DeepSeek V4.1-Flash (peak) | 20M × $0.30 = $6.00 | 20M × $0.006 = $0.12 | 8M × $1.20 = $9.60 | $15.72 |
| LongCat-2.5 (launch pricing) | 20M × $0.30 = $6.00 | 20M × $0.006 = $0.12 | 8M × $1.20 = $9.60 | $15.72 |
| Kimi K3.1 Pro (reported) | 20M × $2.00 = $40.00 | 20M × $0.20 = $4.00 | 8M × $10.00 = $80.00 | $124.00 |
The coincidences in that table are the story. LongCat's all-in launch bill lands exactly on DeepSeek's peak bill — Meituan priced its discount to sit precisely at the incumbent's ceiling. DeepSeek's off-peak window halves it again, which rewards teams who can schedule batch and background agent work into cheap hours. And Kimi's bill is roughly 8–16× higher, before its extra verbosity: at 14% more output tokens, the Kimi total rises to about $135. Cache behaviour is the silent term in every row — at these ratios, an agent framework with poor prefix caching pays the fresh-input rate on everything, and the ranking between DeepSeek and LongCat can flip on that alone. Whichever model you choose, measure cost per finished task on your own traffic within the first week; the tariff sheet is a hypothesis, not a bill.
Round 2: Coding and agent claims — three crowns, no shared ring
Each lab arrives wearing a crown it made itself. Meituan reports LongCat-2.5 at 88.4 on Terminal-Bench v2.1 and 64.6 on DeepSWE v1.1, with agent-endurance scores (BrowseComp 91.9, MCP Atlas 83.1) that are the release's real thesis: this model was RL-trained on multi-hundred-step tasks to finish work, not to chat. DeepSeek reports V4.1-Flash at 77.0 on its own DeepSWE chart at a fifth of the old Flash's cost per task — and, in Moonshot's own comparison table, ahead of Kimi K3.1 Pro on that same-named test, 77.0 to 75.5. Moonshot reports K3.1 Pro fourth on the independent AI Index and strongest of the three on finance research and 3D creation, with a faster Highspeed sibling for latency-sensitive work.
| Claim | LongCat-2.5 | Kimi K3.1 Pro | DeepSeek V4.1-Flash |
|---|---|---|---|
| Headline coding score | Terminal-Bench v2.1: 88.4 (Meituan) | 4th on the AI Index (independent) | DeepSWE: 77.0 (DeepSeek) |
| DeepSWE, someone's table | v1.1: 64.6 (Meituan) | 75.5 (Moonshot's table) | 77.0 (DeepSeek's chart) |
| Agent endurance | BrowseComp 91.9, MCP Atlas 83.1 (Meituan) | Long-horizon agent focus (Moonshot) | Agentic coding focus (DeepSeek) |
| Cost per task | Not reported | $1.98 (Artificial Analysis) | Under $0.20 on DeepSWE (DeepSeek) |
Every cell in that table is a different vendor's ruler, a different benchmark version, or a different harness — which is precisely the point of printing it this way.
The honest problem: almost none of these numbers share a ring. DeepSWE appears in all three stories but in different versions and different harnesses; Terminal-Bench versions differ; the finance and 3D wins are single-vendor tables. The only genuinely independent cross-vendor instrument to launch in this window is SemiAnalysis's GDPx, and none of these three models appears on its reported leaderboard — where, for calibration, the best open-weight model of any lab scored 2.7% on expert work, as our LifeVal vs GDPx comparison details. Until an independent harness runs all three on identical tasks, the defensible ranking is by design intent, not by decimal: Kimi for peak capability, LongCat for long-horizon agent endurance, DeepSeek for cost per finished coding task. Any stronger claim is marketing wearing a lab coat — including, especially, the labs'.
Round 3: Openness and licences — where the real differences live
This is the round the launch posts talk about least and the one that will matter longest.
DeepSeek V4.1-Flash-1017 is the straightforward case: weights on Hugging Face and ModelScope under the MIT licence — download, run, fine-tune, redistribute, build a business on it, no revenue gate, no attribution trigger. The caveats are practical, not legal: no Jinja chat template ships with the release, so self-hosters follow DeepSeek's encoding reference and recommended sampling settings.
Kimi K3.1 Pro is open weights with a fence. The weights landed this week at moonshotai/Kimi-K3.1-Pro — we verified the drop and read the licence in full for our Kimi K3.1 Pro open-weights guide. The Kimi K3.1 Pro License is substantively the K3 licence: commercial use is broadly allowed, but a Model-as-a-Service business above a $20M revenue gate needs Moonshot's separate permission, and products passing 100M monthly users or $20M monthly revenue must display "Kimi K3.1 Pro" attribution. The download itself is a commitment: MXFP4 only, 96 shards, roughly 1.56 TB, served via vLLM, SGLang or TokenSpeed — a data-centre model in every practical sense.
LongCat-2.5 is, today, not open at all. Its GitHub and Hugging Face links are marked "coming soon," Meituan has announced no weights date and no licence, and the lineage cuts both ways: LongCat-Flash and LongCat-2.0 were released openly (2.0 under MIT), which suggests openness is the family habit — but a suggestion is not a licence, and the 2.5 API discount could be the strategy instead. Round to DeepSeek, outright: in a comparison full of reported numbers, MIT is the only claim in this article that is fully verifiable and permanently binding.
Round 4: Context, output and modalities — a quiet three-way split
All three models advertise a 1M-token context window, and all three deserve the same footnote: capacity is not recall, and no independent long-context audit has graded any of them yet. The differences are at the edges. DeepSeek allows the longest single responses — 384K tokens of output, a genuine advantage for whole-codebase generation and long documents — and pairs the window with vision input and both OpenAI- and Anthropic-format APIs, the easiest integration story of the three. LongCat's distinction is input-side: vision is encoded as tokens in the same stream as text (and video is accepted), which is why its showcase use cases are visual coding, interface-from-mockup work and 3D generation rather than chat. Kimi's distinction is ecosystem: a Highspeed sibling for latency, presence on routers and harnesses from day one, and the finance-research tooling Moonshot has been building around the K series. For teams wiring models into agent frameworks, our best AI coding assistants roundup covers the tools that sit on top of all three.
Round 5: Availability — where you can actually run each one
Specs are theoretical until a model answers your API call. DeepSeek is the most available of the three today: first-party API and app, listings on OpenRouter through multiple third-party providers (nine at last count in our guide's research, at marketplace prices that float around $0.08/$0.50), and weights for anyone who would rather be their own provider. Kimi K3.1 Pro listed on OpenRouter on launch day and is built into Moonshot's own Kimi products and the major agent harnesses; its Highspeed sibling covers the latency-sensitive tier, and now that the weights are out, third-party hosting of K3.1 Pro itself will follow the usual pattern of price competition among inference providers. LongCat-2.5 is the narrowest: Meituan's own API (in both OpenAI and Anthropic formats, which is what makes it a drop-in for Claude Code, Codex and the other harnesses Meituan names), plus listings on Vercel AI Gateway and AnyRouter — notably not on OpenRouter as of November 4. For a team standardising on a router, that gap is a practical argument that no benchmark captures: the best-value model is the one your stack can already call. Round to DeepSeek again, on breadth — availability is a feature, and it is the one most comparison tables forget to score.
What would change this verdict
Comparisons of three-week-old models have a short half-life, and three dated events would rewrite this one. First, LongCat's weights. Meituan's "coming soon" page is the single biggest unknown in the table above: a LongCat-2.5 release under an MIT-style licence would move it from the price corner to the openness corner overnight, and the licence it ships under will say more about Meituan's strategy than any benchmark. Second, LongCat's list price. The launch discount has no published end date and no published undiscounted rate; the day either appears, Round 1 gets re-scored — a discount that quietly doubles is a different product than the one reviewed here. Third, a shared independent measurement. Artificial Analysis already tracks Kimi's cost per task; SemiAnalysis updates GDPx every three to six months; the first neutral harness to run all three models on identical agent tasks will replace three vendor narratives with one ranking, and history suggests at least one lab's favourite number will not survive contact with it. We will update this comparison when any of the three lands — the same standard we applied when the Kimi K3.1 Pro weights dropped and this comparison gained its licence facts.
Which one should you pick?
- Pick DeepSeek V4.1-Flash-1017 if the API bill or self-hosting control decides. Volume coding agents, cost-sensitive pipelines, and anyone who wants the legal certainty of MIT plus prices that start at $0.15 per million input tokens. It is the default recommendation of this comparison for working developers.
- Pick Kimi K3.1 Pro if you need the strongest capability of the three and can pay frontier-adjacent prices for it: long-document finance and research work, hard agentic tasks, and teams who want a downloadable frontier-scale model and can house 1.56 TB of it. Read the $20M MaaS gate before you build a product on the weights.
- Pick LongCat-2.5 if your workload is long-horizon agent work — terminal, browser and tool chains that run for hundreds of steps — and you want it at utility prices while the launch discount holds. Go in clear-eyed: API-only, vendor-only benchmarks, and a discount with no visible expiry.
- Pick none of them yet if your buying decision needs independent proof. The shared, neutral measurement of this trio does not exist as of November 4, 2026; the rational move for a large commitment is a one-afternoon bake-off on your own tasks, which at these prices costs less than the meeting that would approve it.
FAQ
Which of the three models is cheapest?
DeepSeek V4.1-Flash-1017, on verified headline rates: $0.15 per million input tokens off-peak ($0.003 on a cache hit) and $0.60 per million output. LongCat-2.5's introductory pricing ($0.30/$1.20, $0.006 cached) matches DeepSeek's peak rates and is simpler, but it is a limited-time discount with no published list price behind it. Kimi K3.1 Pro is the premium option at a reported $2.00/$10.00.
Which models can I download and self-host?
Two of the three. DeepSeek V4.1-Flash-1017 is available now under the MIT licence with no commercial gates. Kimi K3.1 Pro's weights are out now too — MXFP4 format, about 1.56 TB — under Moonshot's custom Kimi K3.1 Pro License, which adds a $20M Model-as-a-Service revenue gate and an attribution trigger at scale. LongCat-2.5 is API-only as of November 4, 2026; its weights are marked "coming soon" with no announced date.
Which is best for coding?
It depends on which claim you weight. DeepSeek's own numbers put V4.1-Flash ahead on cost per finished coding task; Moonshot's table has DeepSeek edging Kimi on DeepSWE (77.0 vs 75.5) while Kimi leads on broader capability; Meituan's numbers make LongCat-2.5 the agent-endurance pick (Terminal-Bench v2.1 88.4, DeepSWE v1.1 64.6). These are different benchmark versions from different vendors — no independent test has run all three head-to-head yet, so serious adopters should benchmark their own workload.
Are any of them actually "open source"?
DeepSeek's release is the only one that meets the strict definition: MIT-licensed weights you can use, modify and redistribute commercially without asking anyone. Kimi K3.1 Pro is open weights under a custom licence with commercial conditions — a meaningful release, but not open source in the MIT sense. LongCat-2.5 is currently a closed API product, whatever its predecessors did.
Why are all three from Chinese labs?
Because the open-and-cheap frontier is currently being built there. DeepSeek competes on efficiency, Moonshot on scale and context, and Meituan — a consumer-platform giant rather than a pure AI lab — on agent training drawn from real platform workloads, all three priced to pressure the closed Western APIs. The practical effect for buyers everywhere is the same: frontier-adjacent capability at a tenth of incumbent prices, with the licence, not the benchmark, as the thing to read first.
Verdict
Three models, three corners of the triangle. DeepSeek V4.1-Flash-1017 wins this comparison for most working developers: the lowest verified prices, the longest output window, and the only licence that asks nothing of you. Kimi K3.1 Pro is the capability pick — the biggest, most broadly skilled model here, priced like it, with open weights that come with a fence serious builders must read. LongCat-2.5 is the specialist: the most interesting agent-training story of the season at utility prices, held back only by being the one model you cannot yet own. The benchmarks will keep arguing; the licences have already decided. Choose by the job, verify on your own tasks, and re-run this comparison the day LongCat's weights page stops saying "coming soon." If this trio proves anything as a group, it is that the price of serious AI capability has decoupled from the price of the companies that used to define it — and that from here on, the comparison that matters most is not model against model, but terms against terms.
Top comments (0)