Kimi vs Qwen — the reasoning lead, and the invoice that argues back.
Two Chinese labs at the frontier, priced in opposite directions. Qwen3.8 Max lists $2/$6 against Kimi K3's $3/$15, so raw output is 2.5× cheaper. But Kimi publishes a $0.30 cache-hit rate and Alibaba publishes none for this model, which flips our blended-workload figure to Kimi's side. Kimi also leads both neutral reasoning benchmarks. Neither appears on SWE-bench Verified, so nobody can honestly say which codes better.
| Category | Winner | Margin |
|---|---|---|
| Input price · per million tokens | BQwen | Qwen3.8 Max lists $2.00 against Kimi K3's $3.00 on each vendor's own page — a third cheaper before a single token comes back |
| Output price · per million tokens | BQwen | $6.00 against $15.00 — the widest gap on the page, and the one that decides any pipeline generating more than it reads |
| Cache pricing · repeated context | AKimi | Moonshot publishes a $0.30 cache-hit rate, a tenth of its miss rate, with caching automatic; Alibaba marks caching as discounted for this model but publishes no per-model rate |
| Blended workload cost · our agentic mix, 92/8 with cache | AKimi | $1.92 against $2.32 per million — Kimi wins on the blend precisely because it publishes a cache rate and Alibaba does not, so read this as a gap in disclosure as much as in price |
| Context window · how much fits in one request | ·Tie | Both bill a single band up to 1M input tokens; Kimi states 1,048,576 exactly and defaults max output to 131,072, configurable to the full window |
| Graduate-level reasoning · GPQA Diamond | AKimi | 93.5 against 92.6 on the llm-stats aggregate — a real lead, but on a benchmark saturated near the top, so treat it as a tie in practice |
| Hard reasoning · Humanity's Last Exam | AKimi | 56.0 against 43.6 — twelve points, and the widest quality gap either model has on a leaderboard that ranks both |
| Coding evidence · neutral leaderboards | ·Unproven | Neither model appears on SWE-bench Verified, LiveCodeBench or Terminal-Bench 2.1. Both vendors publish their own coding numbers on benchmarks the other does not run |
| Consumer app · chatting without an API key | AKimi | Kimi ships agents, swarms, scheduled tasks, docs and slides across five tiers; Qwen Studio is a free chat window with no paid consumer tier at all |
| Free tier · what $0 gets you | BQwen | Qwen Studio is free with no plan above it; Kimi's free Adagio caps you at 1 concurrent agent task, 2 scheduled tasks, 2 projects and 500MB of storage |
| Fixed-fee coding · subscription instead of tokens | BQwen | Alibaba's Coding Plan Pro is $50/mo for 90,000 requests, usable from Claude Code and 15+ tools; Kimi Code rides on consumer tiers whose quotas are described in credits |
| Buying it today · whether checkout works | BQwen | Kimi's pricing page currently shows 'Join Waitlist' on every paid tier while it splits Kimi and Kimi Code benefits; the Qwen API and Coding Plan are on sale normally |
| Price legibility · reading the rate card correctly | AKimi | Moonshot has one USD card; Alibaba shows International, Global and China cards on one page, and picking the wrong row is how $2/$6 becomes $1.65/$4.95 |
| Best overall · for most workloads | ·Depends | Qwen for anything output-heavy, where its $6 rate compounds; Kimi where context repeats enough for the cache rate to bite, or where the reasoning gap is the point |
CHOOSE A · KIMI
If the work is hard and the context repeats.
- Twelve points on HLE — 56.0 against 43.6, the largest quality gap either model can show on a leaderboard that ranks both
- A published cache rate — $0.30 per million on cache hits, automatic, with no cache ID or TTL to manage
- Reasoning is always on — K3 always reasons, with effort configurable rather than optional
- An actual product — agents, swarms, scheduled tasks, docs, sheets, slides and websites, not just an API
- One rate card — a single USD table, so the price you read is the price you pay
CHOOSE B · QWEN
If the tokens are many and the budget is real.
- Output at 40% of the price — $6 against $15 per million, which decides any workload that writes more than it reads
- $50/mo buys 90,000 requests — the Coding Plan Pro is a fixed fee against a metered alternative
- Free chat with no upsell — Qwen Studio has no consumer subscription because there is nothing to sell you
- Thinking costs the same — Non-Thinking and Thinking modes are priced identically, so reasoning is not a surcharge
- A deep line — Max, Plus, Flash, Coder and open-weight releases share one API, so downgrading a task is a model-id change
| Aspect | Kimi | Qwen |
|---|---|---|
| API · inputper 1M tokens · from snapshot | $3.00 | $2.00 B wins |
| API · outputper 1M tokens · from snapshot | $15.00 | $6.00 B wins |
| Effective costblended 92/8 · 82% cache | $1.92 A wins | $2.32 |
| Context windowmax input tokens | 1M | 1M |
| Real cost / 1M charstokenizer-adjusted (prose) | $0.779 | $0.383 B wins |
| Free consumer appwhat a non-paying user gets | $0 · Adagio Kimi on web and mobile with 1 concurrent agent task, 2 scheduled tasks, 2 projects, 500MB of storage and 15+ plugin types | $0 · Qwen Studio Free chat at chat.qwen.ai with no paid tier above it; the signed-out app defaults to Qwen3.7-Plus rather than the flagship B wins |
| Entry subscriptionthe cheapest paid consumer plan | $19/mo · Moderato Or $15/mo billed as $180 a year; adds Deep Research, Swarm with 2 sub-agents, scheduled tasks and Kimi Code. Every paid tier currently shows 'Join Waitlist' | None sold Alibaba sells no consumer subscription for Qwen Studio; heavier use goes to the metered API or the Coding Plan |
| Heavy use, fixed feewhen metering gets uncomfortable | $99–$199/mo Allegro at $99 ($79 annual) unlocks 1M-token conversations and 8 Swarm sub-agents; Vivace at $199 ($159 annual) tops the credit quotas | $50/mo · Coding Plan Pro 90,000 requests a month, capped at 6,000 per 5 hours and 45,000 per week, usable from Claude Code, Cursor, Cline and a dozen more tools B wins |
| Teamsseats, admin, data terms | $600/year/seat Kimi Business Premium, minimum 2 seats, annual billing only with no discounts; business data is not used for training by default A wins | API or Token Plan No per-seat plan for the app; teams buy Model Studio tokens, a Token Plan subscription or Coding Plan seats through Alibaba Cloud |
| Capability | Kimi | Qwen |
|---|---|---|
| Flagship model compared | Kimi K3 | Qwen3.8 Max |
| Lab | Moonshot AI | Alibaba Cloud |
| Context window | 1,048,576 tokens | 1M billing band |
| Max output tokens | 131,072 default, up to 1M | Not separately published |
| Reasoning mode | ✓ always on, effort configurable | ✓ Thinking and Non-Thinking |
| Reasoning priced as a surcharge | ✗ one rate | ✗ both modes priced identically |
| Published cache-hit rate | ✓ $0.30 / 1M | ~ discounted, rate not published |
| Cache management required | ✗ automatic, no cache ID or TTL | ~ undocumented for this model |
| Batch discount | ~ not published | ~ 50% where marked; not on the International card for this model |
| Vision input | ✓ | ✓ |
| Video input | ✓ | ✓ |
| Tool calling | ✓ with tool_choice constraints | ✓ |
| Structured outputs | ✓ JSON mode + structured outputs | ✓ |
| Built-in web search | ~ vendor flags it as under revision | ✓ in the app |
| GPQA Diamond | 93.5 | 92.6 |
| Humanity's Last Exam | 56.0 | 43.6 |
| SWE-bench Verified | ✗ not ranked | ✗ not ranked |
| Terminal-Bench 2.1 | ✗ not submitted | ✗ not submitted |
| Consumer chat app | ✓ kimi.com, five tiers | ✓ Qwen Studio, free only |
| Agent features in the app | ✓ Swarm, scheduled tasks, Claw | ~ chat and search |
| Fixed-fee coding subscription | ~ Kimi Code inside consumer tiers | ✓ $50/mo, 90,000 requests |
| Business seats | ✓ $600/yr/seat, min 2 | ✗ no per-seat app plan |
| Paid tiers purchasable today | ~ waitlist on consumer tiers | ✓ |
| Regional rate cards on one page | ✓ one USD card | ✗ International, Global and China |
| Model line depth | K3, K2.7 Code, K2.6 | Max, Plus, Flash, Coder, VL, open weights |
| Open-weight releases in the family | ~ earlier K2 generations | ✓ Qwen3 family on Hugging Face |
| Runs locally through Ollama or llama.cpp | ~ older open weights only | ✓ many sizes, GGUF available |
| Taxes | Excluded, added at checkout | Excluded from list rates |
§ 04 / BENCHMARKS
The numbers, not the spin.
Reasoning · GPQA Diamond
Hard reasoning · Humanity's Last Exam
A · MOONSHOT AI
Kimi
A lab that sells a product as much as an API, and prices its flagship like the frontier model it benchmarks as.
Strengths
- The reasoning lead is real — first on both leaderboards that rank the pair, and twelve points clear on Humanity's Last Exam
- Caching is priced and automatic — $0.30 per million on hits against $3.00 on misses, with no cache ID, TTL or extra parameter to manage
- The full million — 1,048,576 tokens of context, with max output configurable up to the same figure
- A real application layer — Swarm sub-agents, scheduled tasks that run and push results, docs, sheets, slides, websites and Kimi Claw
- One price card — a single USD table with taxes stated separately, which is rarer among Chinese labs than it should be
- Business terms exist — $600 per seat per year with business data excluded from training by default
Weaknesses
- Output costs 2.5× more — $15 against $6 per million, and on our blended mix output is 62% of Kimi's cost against 21% of Qwen's
- Paid tiers are on a waitlist — the pricing page is mid-restructure while Kimi and Kimi Code benefits are separated
- No neutral coding score — absent from SWE-bench Verified, LiveCodeBench and Terminal-Bench, so its coding claims are self-reported
- Business billing is annual only — no monthly option and, per its own FAQ, no promotions or discounts
Best for
- Research and analysis where the hard-reasoning gap is worth paying for
- Long documents re-read across many turns, where the cache rate does the work
- Anyone who wants the agent features rather than raw API access
- Teams that need a written no-training-by-default commitment
B · ALIBABA
Qwen
A model line priced to be used in volume, sold through a cloud console rather than a consumer subscription.
Strengths
- Output at $6 — 40% of Kimi's rate, which decides any workload that generates more than it reads
- Thinking is not a surcharge — Thinking and Non-Thinking modes sit on the same rate, in a single billing band to 1M tokens
- A fixed-fee coding path — $50/mo for 90,000 requests, working from Claude Code, Cursor, Cline and a dozen other tools
- Depth of line — Max, Plus, Flash, Coder, VL and open-weight releases, so moving a task down a tier is a model-id change
- Runs locally too — Qwen3 weights are on Hugging Face in GGUF, so the same family works through Ollama or llama.cpp
- Free chat with no upsell — Qwen Studio charges nothing and has no plan to sell you
Weaknesses
- Three rate cards on one page — International, Global and China; reading the wrong row turns $2/$6 into $1.65/$4.95
- No published cache rate — caching is marked as discounted for this model, but the discount is not quantified
- Batch is not on the card — the 50% batch discount applies only where a model is marked for it, and the International row for Max is not
- The app is thin — a chat window against Kimi's agents, and it defaults to Qwen3.7-Plus rather than the flagship when signed out
Best for
- Output-heavy pipelines where the per-million rate compounds
- Coding work that fits a fixed monthly fee instead of a meter
- Anyone who wants one API across frontier, mid and cheap tiers
- Teams that would rather self-host a smaller model from the same family
01
Summarising thousands of support tickets every night
Short inputs, short outputs, enormous volume, and a finance team that reads the invoice line by line.
Reasoning: Every ticket is a fresh prompt, so nothing caches and Kimi's cache rate never engages — which is exactly the case where the list price is the whole story. Qwen is a third less on input and 60% less on output, and nothing in the reasoning gap between these two changes what a ticket summary looks like.
Picked
Qwen
Runner-up: Kimi, if the summaries feed a downstream analysis that needs the reasoning
02
A research assistant over a 400-page document set
The same corpus re-read across dozens of questions in a session, with answers that have to be defensible.
Reasoning: Two things favour Kimi. Its cache-hit rate of $0.30 per million turns a repeatedly re-sent corpus into a tenth of its list price, and Alibaba publishes no equivalent figure for Qwen3.8 Max. The twelve-point Humanity's Last Exam gap also lands squarely on this kind of work.
Picked
Kimi
Runner-up: Qwen, if the corpus changes every request and caching never engages
03
A solo developer replacing a coding subscription
Daily agent use from Claude Code, a preference for one predictable charge instead of a meter.
Reasoning: Alibaba's Coding Plan Pro is exactly this product: $50 a month, 90,000 requests, and documented support for Claude Code among 15+ tools. Kimi Code exists but rides on consumer tiers whose quotas are expressed in credits, and those tiers are currently on a waitlist while the plans are restructured.
Picked
Qwen
Runner-up: Kimi Moderato at $19/mo, once its checkout reopens
04
Choosing a model for a hard-reasoning eval
A benchmark-shaped workload — science questions, multi-step deduction, no tolerance for a confident wrong answer.
Reasoning: This is the one case where the published evidence points clearly. Kimi K3 is ahead on both leaderboards that rank the pair, and the Humanity's Last Exam margin is wide enough to survive the usual caveats about configuration. Pay the output rate and take the accuracy.
Picked
Kimi
Runner-up: Qwen3.8 Max, which is within a point on GPQA at a lower price
05
Deciding which one writes better code
A team that wants a defensible answer before standardising an agent on one of these models.
Reasoning: There is no defensible answer from public data. Neither model appears on SWE-bench Verified, LiveCodeBench or Terminal-Bench 2.1, and each vendor reports coding scores on benchmarks the other does not run. The honest move is a bake-off on your own repository, with both vendors' cheapest tiers, before committing.
Picked
Qwen
Runner-up: Kimi, if the trial shows it — the price gap means Qwen starts ahead
06
A small team that needs a written data commitment
Five people, work product they cannot have trained on, and a procurement checklist that asks about it in writing.
Reasoning: Kimi sells a Business plan at $600 per seat per year, minimum two seats, that states business data is not used for training by default. Qwen has no per-seat app plan at all — the equivalent path is a cloud account with Model Studio terms, which is a different conversation with a different signer.
Picked
Kimi
Runner-up: Qwen through Alibaba Cloud, if the team already has an account there
Frequently asked.
Common questions about this comparison, with sources where they matter.
Q · 01 Is Kimi or Qwen cheaper? +
It depends on the shape of the work, and the answer flips. On list rates Qwen is cheaper on both sides — $2.00 input and $6.00 output per million against Kimi K3's $3.00 and $15.00. But on the blended agentic mix this site uses everywhere (92% input, 8% output, most of the input cached) the order reverses: $1.92 for Kimi against $2.32 for Qwen. The reason is disclosure, not generosity — Moonshot publishes a $0.30 cache-hit rate and Alibaba publishes none for this model, so our formula has to charge Qwen full price on cached tokens. If your work re-reads the same context, Kimi is likely cheaper than its headline. If it generates long outputs from short prompts, Qwen wins outright.
Q · 02 Why do I see $1.65 and $4.95 for Qwen3.8 Max elsewhere? +
Because Alibaba prints more than one rate card on the same page. The $2.00 / $6.00 figures are the International (Singapore) deployment; the $1.65 / $4.95 pair belongs to the Global and China cards. Aggregators frequently quote whichever row they scraped first. We record the International card, read in a rendered browser, because mixing regional cards is exactly how a batch of prices goes wrong — it produced fourteen incorrect figures in our own snapshot once, which is why the rule now exists.
Q · 03 Which is better at coding? +
Nobody can honestly say from public data. Neither Kimi K3 nor Qwen3.8 Max appears on SWE-bench Verified, LiveCodeBench or the Terminal-Bench 2.1 leaderboard. Each vendor publishes coding results, but on benchmarks the other does not run, which makes the numbers unjoinable rather than merely inconvenient. If coding is the decision, run both against your own repository — and note that the price gap means Qwen only has to match Kimi to win on cost.
Q · 04 How much better is Kimi at reasoning, really? +
On the two neutral leaderboards that rank both models, Kimi K3 leads on each. GPQA Diamond is 93.5 against 92.6, which is a lead in the table and noise in practice, because the benchmark is saturated near the top. Humanity's Last Exam is 56.0 against 43.6, and twelve points is a real difference. HLE forks on whether tools were allowed, so treat that gap as a direction rather than a precise margin — but it points the same way on both tests.
Q · 05 Do both have a consumer app I can just use? +
Yes, and they are very different products. Qwen Studio is a free chat window with no paid tier above it, and when signed out it defaults to Qwen3.7-Plus rather than the flagship. Kimi is a five-tier product — free Adagio, then Moderato $19, Allegretto $39, Allegro $99 and Vivace $199 a month, cheaper annually — with agent swarms, scheduled tasks, documents, slides and websites. As of late August 2026 every paid Kimi tier shows Join Waitlist while Moonshot separates Kimi and Kimi Code benefits.
Q · 06 Which one has a team plan? +
Kimi does: a Business Premium tier at $600 per seat per year, minimum two seats, billed annually with no promotions or discounts, and a commitment that business data is not used for training by default. Alibaba sells no per-seat plan for the Qwen app — teams buy Model Studio tokens, a Token Plan subscription or Coding Plan seats through an Alibaba Cloud account instead. If a signed per-seat agreement is the requirement, only one of these two is selling one.
Q · 07 Can I run either of these models on my own hardware? +
Not these two specifically — Kimi K3 and Qwen3.8 Max are served through their vendors' APIs. But Alibaba publishes a broad Qwen3 family as open weights, in sizes from 8B upward and in GGUF, so the same lineage runs locally through Ollama or llama.cpp. Moonshot has released weights for earlier K2 generations. If self-hosting is the goal, Qwen is the family with the most rungs on the ladder.
Q · 08 Does thinking mode cost extra on either? +
No, on both. Alibaba prices Qwen3.8 Max identically in Thinking and Non-Thinking modes within a single billing band that runs to 1M input tokens. Kimi K3 always reasons — it is not a mode you switch on — with reasoning effort configurable and one rate on the card. That is worth noting because several Western vendors do charge a separate, higher rate for reasoning output, which makes a naive rate comparison flatter than the bills turn out to be.