GLM-5.2 vs Kimi K3: Cheaper and Faster, or Bigger and Better

GLM-5.2 and Kimi K3 arrived a month apart in mid-2026 and invert each other almost perfectly. Kimi K3, released 2026-07-16 with open weights following on 2026-07-27, is the bigger and on paper the better model: 2.8 trillion parameters, 104 billion of them active, and the higher score on essentially every benchmark the two have been run against. GLM-5.2 has been out since 2026-06-16, is the smaller of the two at 753 billion parameters with 40 billion active, and costs roughly a third as much per output token, answers faster at the median, and ships under MIT rather than a bespoke licence with revenue triggers.

That is the decision in one paragraph. What follows is the evidence behind it — the price arithmetic on a job you might actually run, the measurements where the two are close enough that the difference is noise, and one popular number that is not comparable to the number printed beside it.

Where the two stand

Both are generally available through their vendors' own APIs, both are routable on OrcaRouter, and both have downloadable weights. The differences that matter before you get anywhere near a benchmark:

• Released — GLM-5.2 on 2026-06-16 vs Kimi K3 on 2026-07-16 (Artificial Analysis model pages; OrcaRouter's own model page for Kimi K3 dates it a day earlier, 2026-07-15)

• Weights — GLM-5.2 open under MIT vs Kimi K3 open under the Kimi K3 License, which requires a separate agreement above $20M in annual model-as-a-service revenue and prominent attribution above 100M monthly users (internal research and development is exempt)

• Size — GLM-5.2 753B total / 40B active vs Kimi K3 2.8T total / 104B active

• Context — GLM-5.2 1,000,000 tokens vs Kimi K3 1,048,576 tokens

• Input — GLM-5.2 text in, text out vs Kimi K3 text and images in, text out

• Published max output — GLM-5.2 128,000 tokens vs not published on Kimi K3's OrcaRouter page

On OrcaRouter both sit behind one key on one OpenAI-compatible endpoint at https://api.orcarouter.ai/v1, and we pass the provider's list price straight through with no markup added — so every figure below is the vendor's number, not ours.

What a million tokens costs, on a job that costs something

Vendor rate cards first, read from the vendors' own pricing pages on 2026-09-23. Z.ai lists GLM-5.2 at $1.40 per million input tokens, $0.26 per million cached input tokens, and $4.40 per million output tokens. Moonshot lists Kimi K3 at $3.00 input, $0.30 cache read, $15.00 output — plus a cache write charge of $3.00 per million for a five-minute cache and $6.00 per million for a one-hour cache. Those are published prices, not independently audited ones, and either vendor can change them.

Put them on a job that is mostly reading: a 600,000-token codebase review with an 8,000-token written answer. On GLM-5.2 that is 0.6 x $1.40 = $0.84 of input plus 0.008 x $4.40 = $0.035 of output, or about $0.88. On Kimi K3 it is 0.6 x $3.00 = $1.80 plus 0.008 x $15.00 = $0.12, or about $1.92. Roughly 2.2 times the cost for the same job.

Now run it again with the codebase already sitting in the provider's cache. The input line collapses to the cache-read rate, and the two prices that were 2.1x apart become $0.26 against $0.30 per million — a 15% gap. This is the least-discussed number in the comparison and the one that matters most to an agent that re-reads the same context every turn. It also has an asterisk: Kimi K3 bills separately to write the cache and GLM-5.2's page does not advertise a write charge at all, so a cold start costs meaningfully more on Kimi K3 than the $3.00 headline suggests.

• A 600k-token read with an 8k answer — GLM-5.2 about $0.88 vs Kimi K3 about $1.92

• The same cached input line — GLM-5.2 $0.16 vs Kimi K3 $0.18 per 600k tokens

The independent model agrees on direction but not on magnitude. Artificial Analysis blends cache-hit, input and output tokens at a 7:2:1 ratio and adds the reasoning tokens a model actually burns, and its figures for these two — read 2026-09-23 under its v4.3.2 index — put GLM-5.2 at $0.902 per million blended tokens against Kimi K3's $2.31, and the cost of completing one of its evaluation tasks at $0.96 against $2.00. Running the full Intelligence Index once costs $1,559 on GLM-5.2 and $3,658 on Kimi K3.

There is a counter-intuitive detail buried in that cost model. Kimi K3 uses fewer output tokens per task than GLM-5.2 — 48,000 against 64,000 — and fewer reasoning tokens, 32,000 against 51,000. It is the more economical model per unit of work and still costs roughly twice as much per task, because its per-token price is 2.1x on input and 3.4x on output. Cheap-per-task and cheap-per-token are different properties, and this is a pairing where they point in opposite directions.

The context window, and what actually fits in it

The two are within 5% of each other on paper: 1,000,000 tokens against 1,048,576. Reporting that as a Kimi K3 win would be silly. Both are million-token models and the window is not a differentiator; the honest question is what each one can still do with text buried in the middle of it.

That is what Artificial Analysis's long-context reasoning evaluation measures, and it is one of the wider gaps in the dataset: Kimi K3 scores 89% against GLM-5.2's 78%. If your job is to load a large corpus and ask questions that require joining facts from opposite ends of it, the extra 48,576 tokens are not the reason to pick Kimi K3 — the eleven-point long-context margin is.

The floor under the window differs too. GLM-5.2 publishes a 128,000-token maximum output; Kimi K3's page does not publish an equivalent figure, so we are not going to supply one. If you are generating long documents rather than reading them, that asymmetry is worth checking against your own workload before you buy either.

Speed: the one axis GLM-5.2 owns outright

Two independent measurements point the same way here, which is worth saying precisely because they do not agree on everything.

Our own routing data, over the seven days to 2026-09-23, shows GLM-5.2 answering at a median 3.28 seconds to first token against Kimi K3's 8.09 seconds, and streaming 68.1 tokens per second against 43.4. Both models' p95 time to first token sits at exactly 10.00 seconds. Artificial Analysis, measuring separately, has GLM-5.2 at 68.2 output tokens per second against Kimi K3's 36.7, at 41.29 seconds end-to-end against 72.13, and at 33.95 seconds to first answer against 58.52.

The two sources diverge on one point, and it is worth flagging rather than smoothing over. Artificial Analysis puts Kimi K3 slightly ahead on raw time to first token, 4.08 seconds against 4.62, while our own routing has GLM-5.2 nearly two and a half times faster at p50. Different endpoints, different upstream providers, different windows. Treat the throughput direction — GLM-5.2 comfortably faster — as solid, and treat any single time-to-first-token figure, ours or theirs, as a property of a particular week.

There is also a number on our GLM-5.2 page that we are not going to quietly leave out: over the same seven days it shows a 9.2% error rate, against 0.26% for Kimi K3. A gap that size is about as likely to be a bad week for the upstream providers serving GLM-5.2 as a property of the model, and seven days is not a long enough window to tell the difference. It is the kind of problem routing exists to solve — when a provider is failing one request in eleven, automatic failover moves the traffic without anyone paging on-call.

Benchmarks: a real sweep, narrower than the headline

Kimi K3 wins almost everything both models have been run on. These are Artificial Analysis figures under index revision v4.3.2, both models in their maximum-reasoning configuration, read 2026-09-23:

• Intelligence Index — Kimi K3 44 vs GLM-5.2 34

• AA-Briefcase v1.1 — 1510 vs 1233

• GDPval-AA v2.1 — 1524 vs 1358

• AutomationBench-AA — 58% vs 28%

• Terminal-Bench 4.0 — 13% vs 1%

• SciCode — 59% vs 51%

• Humanity's Last Exam — 47% vs 41%

• GDP.pdf — 22% vs 10%

• AA-LCR v1.1 — 89% vs 78%

• CritPt — 23% vs 21%

Two caveats before anyone screenshots that list.

First, the index revision. Kimi K3's July launch coverage quoted it at 57 on the Intelligence Index with GLM-5.2 at 51. The live pages today say 44 and 34. Those are not the same measurement — Artificial Analysis revises the index and re-scores models against it, and putting a July figure beside a September figure is the easiest mistake to make on this pairing. The direction is stable across every snapshot; the absolute numbers are not comparable across revisions.

Second, the levels. Terminal-Bench 4.0 at 13% and 1% is a hard evaluation, and at that altitude the ratio tells you more than either number. AA-Omniscience, which probes whether a model knows the limits of its own knowledge, has GLM-5.2 at 4 against Kimi K3's 20 — a floor worth knowing about if you plan to run either one unsupervised.

One more thing Artificial Analysis records about the GLM-5.2 listing: it marks the maximum-reasoning configuration as deprecated in favour of the newer GLM-5.3. That changes none of the figures above, which are a snapshot of GLM-5.2 as it was measured, but it does mean Z.ai's roadmap has moved past this model. On a routed endpoint that costs a model string rather than a migration, which is the main argument for not hard-coding either one of these names into your application.

Agents and tool use: where the gap is widest

If one block of that table should decide the purchase, it is this one. Long-horizon agentic work is where Kimi K3's margin is largest and most consistent:

• AutomationBench-AA — 58% vs 28%, a 30-point gap

• GDPval-AA v2.1 — 1524 vs 1358

• AA-Briefcase v1.1 — 1510 vs 1233

• Terminal-Bench 4.0 — 13% vs 1%

Moonshot's own release material leans on the same story — the K3 weights drop shipped with a technical report covering the vendor's expert-parallel training stack and an agent-environment harness — but vendor material is vendor material, and the figures above are the independent ones.

The third-party aggregator LLM Stats, which runs its own composite over a shared benchmark set, reaches the same conclusion from a different direction: of the eleven benchmarks it scores for both models, Kimi K3 takes all eleven, and its category scores put Kimi K3 ahead on tool use (30.8 vs 19.6), agents (38.3 vs 29.6) and coding (42.3 vs 34.8). Those are LLM Stats's own composites under its own weighting, on a shared set that is not large, so treat them as corroboration rather than as a lab result. Eleven out of eleven is still not a close call.

Coding

Same direction, smaller margin. On the coding evaluations Artificial Analysis scores for both, Kimi K3 leads SciCode 59% to 51%; on LLM Stats's shared set it takes DeepSWE, DeepSWE 1.1, FrontierSWE, SWE-Marathon, Program Bench and Terminal-Bench 2.1. GLM-5.2's flattering numbers — 99.2% on AIME 2026, 94.4% on HMMT 2025, 91.2% on GPQA as carried by third-party aggregators — are vendor-reported and unreproduced, and most of them measure competition mathematics rather than software engineering.

The practical read: for a coding agent that runs for a long time, edits many files and has to recover from its own mistakes, Kimi K3's agentic margin is the more relevant signal than either model's maths scores. For a fast, cheap, largely single-shot code assistant, GLM-5.2's throughput advantage is worth more than the benchmark gap costs.

Where they are close enough that it does not matter

• Cached input price — $0.26 vs $0.30 per million, a 15% gap against a 3.4x gap on output

• Context window — 1,000,000 vs 1,048,576 tokens

• p95 time to first token — 10.00 s vs 10.00 s on our own routing

• CritPt — 23% vs 21%

• Mathematics — LLM Stats puts GLM-5.2 marginally ahead on its maths composite (41.4 vs 40.9) while Kimi K3 leads on maths-with-images; on the evidence available neither model owns arithmetic

Anyone telling you one of these models is twice the other across the board is reading a single row of the table.

Pick GLM-5.2 if…

You are paying the bill and the bill is the constraint. GLM-5.2 is roughly a third the output price, cheaper on the cached line as well, MIT-licensed with no revenue trigger to check against, faster at the median and much faster when streaming, and its million-token window is close enough to Kimi K3's that the difference will never decide anything. If your workload is text in and text out, and it is mostly reading rather than long chains of tool calls, the extra benchmark points on Kimi K3 are points your application will never collect.

Pick Kimi K3 if…

The work is agentic and long. The 30-point AutomationBench gap, the leads on GDPval and AA-Briefcase, the eleven-point long-context reasoning margin and the sweep of LLM Stats's shared set all point the same way: Kimi K3 is the better model to hand a multi-step task and walk away from. It is also the only one of the two that accepts images. You will pay roughly twice as much per task for that, and if your working set is heavily cached you will pay less than twice — but the case for it is not price, and it is not speed.

Both are routable on OrcaRouter from one key against one OpenAI-compatible endpoint, at the providers' list prices with nothing added on top, and both sit behind the same automatic failover. That is the cheapest way to find out which one your workload actually wants, because switching between them is a model string rather than a contract.