Kimi K3 vs GLM-5.3-Flash: Pricing, License & When Flash Is Enough
September 25, 2026 · kimi k3 vs glm 5.3 flash · glm 5.3 flash · glm 5.3 flash pricing · kimi k3 pricing · cheap ai coding model · open weight ai models · mit license ai model · ai model comparison
GLM-5.3-Flash is the cheaper model by an order of magnitude, and Kimi K3 is still the slightly smarter one. Zhipu’s 320B-parameter Flash lists at $0.15 input / $0.50 output per million tokens against K3’s $3 / $15, ships under a plain MIT license, and fits on a single GPU node. On Artificial Analysis’s current index K3 leads only 44 to 42, and Flash actually beats it on terminal and office-work tests. For most cost-conscious teams, Flash should be the default and K3 the escalation tier.
This comparison is current as of September 25, 2026, two weeks after Flash’s 50% launch promo ended and a week after Zhipu shipped the faster FlashX tier. Prices below are list prices from the official Z.ai pricing page and Kimi platform pricing. Benchmarks come from Artificial Analysis’s head-to-head comparison unless a vendor number is flagged as such.
The two models at a glance
| Kimi K3 | GLM-5.3-Flash | |
|---|---|---|
| Maker | Moonshot AI | Z.ai (Zhipu AI) |
| Released | July 16, 2026 | August 26, 2026 |
| Parameters | 2.8T MoE, 104B active | 320B MoE, 18B active |
| Context window | 1,048,576 tokens | 1M tokens (Z.ai docs) |
| Max output | Not publicly capped | 128K tokens |
| Input modalities | Text, image, video | Text, image, video, file |
| Thinking | Always on; low / high / max | Always on; low / high / max |
| Input / cached / output (per 1M) | $3.00 / $0.30 / $15.00 | $0.15 / $0.03 / $0.50 |
| License | Kimi K3 License | MIT |
| Weights on disk | ~1.56TB (MXFP4) | ~306 GiB (FP8) |
| AA Intelligence Index v4.3.2 | 44 | 42 |
| AA cost per task | $2.00 | $0.25 |
| AA output speed | 35 tok/s | 43 tok/s |
Two things stand out. First, a 9x difference in total parameters and a 6x difference in active parameters buys K3 only two index points. Second, K3’s old multimodal advantage is gone. When we compared it with GLM-5.2, Zhipu’s model was text-only. Flash is natively multimodal, and Z.ai’s model docs list video, image, text and file input.
What GLM-5.3-Flash actually is
Flash had an unusual launch. Zhipu ran it anonymously as “Ox Alpha” on OpenRouter and OpenCode before revealing it on August 26. According to the South China Morning Post, the stealth trial ran entirely on 100,000 domestically produced Chinese chips and processed 62 trillion tokens, and Zhipu’s Hong Kong shares closed more than 12% higher the day after the reveal.
Under the hood, per MarkTechPost’s launch breakdown:
- 320B total / 18B active mixture-of-experts, 45 layers, native FP8 weights with one multi-token-prediction draft layer.
- Hybrid attention that interleaves linear-attention layers with sparse attention layers, plus a technique Zhipu calls IndexPool, claimed to give about 3x less attention compute and a 4.4x smaller KV cache than GLM-5.3.
One naming trap worth flagging: GLM-5.3-Flash is not GLM-5.3. Zhipu’s text-only flagship GLM-5.3 launched on August 14 and costs $1.40 / $4.40 per million tokens, the same as GLM-5.2. Flash is the smaller, multimodal, roughly 10x cheaper sibling, so check the model ID (glm-5.3-flash) before you trust a quote.
Then on September 18 Zhipu added GLM-5.3-FlashX, a faster serving tier. Phemex and other outlets report up to 200 tokens per second, up from roughly 30 to 50 on standard Flash.
Pricing: every tier side by side
| Model (per 1M tokens) | Input | Cached input | Output | Context |
|---|---|---|---|---|
Kimi K3 (kimi-k3) | $3.00 | $0.30 | $15.00 | 1,048,576 |
Kimi K2.7 Code (kimi-k2.7-code) | $0.95 | $0.19 | $4.00 | 262,144 |
| GLM-5.3 (flagship, text-only) | $1.40 | $0.26 | $4.40 | 1M |
| GLM-5.3-Flash | $0.15 | $0.03 | $0.50 | 1M |
| GLM-5.3-FlashX | $0.37 | $0.075 | $1.25 | 1M |
| Flash launch promo (ended Sep 9) | $0.075 | $0.015 | $0.25 | 1M |
On list price, K3 costs 20x more for fresh input, 10x more for cached input and 30x more for output. Even FlashX, the premium speed tier, is 8x cheaper than K3 on input and 12x cheaper on output.
A few pricing details that matter in practice:
- The promo is over. The 50% launch discount ended at 24:00 on September 9, 2026 (UTC+8), per eesel’s pricing tracker. Quotes of $0.075 / $0.25 are out of date.
- No free tier for Flash. Z.ai’s free API models are older: GLM-4.7-Flash, GLM-4.5-Flash and the vision model GLM-4.6V-Flash. Cloudflare’s Workers AI page requires the Paid plan or prepaid AI Gateway credits.
- Kimi’s cache has a write cost. Moonshot bills cache writes at $3.00 per million tokens for a 5-minute TTL (same as normal input) or $6.00 for a 1-hour TTL; hits then cost $0.30. Z.ai lists cache storage as “limited-time free.”
If you want to stay inside Moonshot’s ecosystem on a budget, kimi-k2.7-code at $0.95 / $4.00 is the closest thing Moonshot has to a Flash tier. It is still more than 6x Flash’s input price and 8x its output price.
What real workloads cost
Sticker prices exaggerate the gap, because the two models spend tokens differently:
| Workload | Kimi K3 | GLM-5.3-Flash | GLM-5.3-FlashX |
|---|---|---|---|
| One coding-agent session: 2M input (90% cached), 150K output | $3.39 | $0.16 | $0.40 |
| Monthly team usage: 1B input (85% cached), 50M output | $1,455 | $73 | $182 |
| Artificial Analysis cost per benchmark task | $2.00 | $0.25 | n/a |
| Cost to run AA’s full Intelligence Index | $3,658 | $280 | n/a |
The first two rows assume both models use the same number of tokens, which gives a roughly 20x gap. The last two rows are measured, and the gap narrows to about 8x per task and 13x for the full suite. The reason is verbosity. Artificial Analysis counts about 69K output tokens per task for Flash (47K of them reasoning) against 48K for K3 (32K reasoning). Flash thinks longer to reach its answers, and because thinking cannot be switched off (Z.ai’s docs say the thinking parameter “only supports enabled”), you pay for those tokens on every call.
An 8x saving is still enormous. But plan budgets on measured per-task cost, not on the 30x output-price ratio.
The GLM Coding Plan option
For individual developers, Zhipu’s GLM Coding Plan is often cheaper than the API: Lite is $18 per month, Pro $80 and Max $168 per eesel’s tracker. The plan runs on credits, and Flash burns them at about a third of GLM-5.3’s rate (multipliers of 2.3 input / 8 output versus 6.9 / 24). Through October 7, paid plans also get unlimited Flash use between 23:00 and 09:00 daily. FlashX is not on the plan yet.
On the Kimi side, consumer memberships cover K3 in the Kimi app and Kimi Code. For API buyers, the provider comparison lists every K3 host, including cheaper third-party endpoints.
Benchmarks: closer than the price gap implies
Independent head-to-head
Artificial Analysis runs both models through the same harness, which makes it the fairest comparison available. Scores are from Intelligence Index v4.3.2, with K3 at max reasoning effort:
| Evaluation | Kimi K3 (max) | GLM-5.3-Flash | Winner |
|---|---|---|---|
| Intelligence Index v4.3.2 | 44 | 42 | K3 |
| Humanity’s Last Exam | 47% | 40% | K3 |
| SciCode | 59% | 52% | K3 |
| AA-LCR v1.1 (long context) | 89% | 80% | K3 |
| AA-Omniscience index | 20 | 7 | K3 |
| CritPt | 23% | 15% | K3 |
| GDP.pdf | 22% | 15% | K3 |
| AA-Briefcase v1.1 | 1,505 | 1,451 | K3 |
| Terminal-Bench 4.0 | 13% | 33% | Flash |
| GDPval-AA v2.1 | 1,524 | 1,641 | Flash |
| AutomationBench-AA | 58% | 60% | Flash |
The split is clean. K3 wins where raw knowledge and deep reasoning matter: frontier science questions, scientific coding, long-document reasoning and the Omniscience index, which rewards correct answers and penalizes hallucinated ones. Flash wins the agentic, get-the-job-done tests: terminal tasks, GDPval’s real-world office deliverables and automation workflows.
The Terminal-Bench 4.0 result is the surprise, with Flash scoring 2.5x what K3 does. It is a new and hard version of the test, and K3’s 13% looks out of line with its strong results on earlier Terminal-Bench releases. Treat one test as a reason to run your own trial, not as a verdict.
Vendor numbers, and why you can’t line them up
Z.ai’s launch chart did not include Kimi K3. It compared Flash with Claude Opus 4.8, GPT-5.6 Terra and GLM-5.2, reporting Terminal-Bench 2.1 at 84.3 (Opus 4.8: 85.0; GPT-5.6 Terra: 87.4), DeepSWE v1.1 at 63.4 (GLM-5.2: 46.2), AutomationBench at 48.8 (GLM-5.2: 26.2) and Z.ai Code Bench at 29.0 versus Opus 4.8’s 29.5, per OfficeChai’s writeup.
It is tempting to put those next to Moonshot’s launch numbers for K3 (Terminal-Bench 2.1 at 88.3, DeepSWE at 67.5). Be careful. Moonshot’s suite scored GLM-5.2 at 12.9 on AutomationBench, while Z.ai’s scored the same model at 26.2. Different harnesses produce different numbers for the same weights. DeepSWE is the one partial exception: both vendors report GLM-5.2 at exactly 46.2, which suggests comparable setups, and on that anchor K3 (67.5) sits about four points above Flash (63.4). That is a vendor-to-vendor inference, not an independent result.
Speed and latency
Flash is smaller, but it is not dramatically faster on Z.ai’s own API. Artificial Analysis measures 43 tokens per second for Flash versus 35 for K3, with near-identical time to first token (3.67s vs 3.81s). Because Flash reasons for longer, its average time per benchmark task (1,388s) is actually a little longer than K3’s (1,266s).
The speed picture changes once you leave the first-party endpoint. eesel reports third-party hosts serving the same Flash weights at 86 tokens per second on Together AI and 273 on Databricks, and FlashX is rated at up to 200. An 18B-active model is simply cheap to serve fast; K3 has far less room for that.
Licensing: MIT vs the Kimi K3 License
| Kimi K3 | GLM-5.3-Flash | |
|---|---|---|
| License | Kimi K3 License | MIT |
| Commercial use | Yes | Yes |
| Fine-tuning and redistribution | Yes, with conditions | Yes |
| Attribution trigger | >100M MAU or >$20M monthly revenue | None beyond the MIT notice |
| Model-as-a-Service clause | Separate agreement above $20M revenue in 12 months | None |
| Where to download | Hugging Face moonshotai/Kimi-K3 | Hugging Face and ModelScope, BF16 and FP8 |
The Kimi K3 License is permissive for almost everyone. You only need to display “Kimi K3” in your interface once your product passes 100 million monthly active users or $20 million in monthly revenue. The clause that actually bites is for companies that resell K3 as an API: above $20 million in total revenue over any 12 consecutive months, a Model-as-a-Service operator needs a separate agreement with Moonshot. Internal use is excluded.
GLM-5.3-Flash’s model card says MIT, full stop. If you are an inference provider, a platform that wants to white-label a model, or a legal team that simply doesn’t want to track thresholds, Flash is the simpler answer. We break down the K3 terms in more detail in our Kimi K3 license explainer.
Self-hosting: one node vs a cluster
This is where the size difference becomes decisive:
| Kimi K3 | GLM-5.3-Flash | |
|---|---|---|
| Weights | ~1.56TB (MXFP4) | ~306 GiB (FP8) |
| Minimum realistic serving | 8x B300 / 8x MI355X, or 16x H200 | One 8-GPU Hopper-class node, or a GB200 tray at TP4 |
| 8x H100 (640GB) | Impossible | Fits, with room for KV cache |
| Frameworks | vLLM, SGLang | SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth |
Flash’s FP8 checkpoint is about 306 GiB before KV cache, per MarkTechPost, and needs Hopper or newer GPUs: hardware a mid-size company can rent by the hour. K3 needs roughly 1.7 to 1.84TB of GPU memory once you add KV cache and overhead, which is why Moonshot’s own production guidance talks about 64+ accelerator supernodes. Our VRAM requirements guide has the full K3 hardware math.
If you need air-gapped deployment or fine-tuning on a normal budget, Flash is the realistic open-weight choice.
When Flash is enough, and when you need K3
| Your situation | Pick | Why |
|---|---|---|
| High-volume coding agents, CI bots, bulk refactors | GLM-5.3-Flash | ~8x cheaper per task; wins Terminal-Bench 4.0 |
| Terminal and shell automation | GLM-5.3-Flash | 33% vs 13% on Terminal-Bench 4.0 |
| Office deliverables, spreadsheets, reports | GLM-5.3-Flash | Higher GDPval-AA (1,641 vs 1,524) |
| Extraction, classification, screenshot QA | GLM-5.3-Flash | Multimodal at $0.15 input |
| Self-hosting on one node, or fine-tuning | GLM-5.3-Flash | 306 GiB, MIT |
| Reselling a model as an API | GLM-5.3-Flash | No MaaS clause |
| Hard science, math and research questions | Kimi K3 | HLE 47% vs 40%, CritPt 23% vs 15% |
| Knowledge-heavy answers where hallucination hurts | Kimi K3 | AA-Omniscience 20 vs 7 |
| Reasoning over very long documents | Kimi K3 | AA-LCR 89% vs 80% |
| Tasks Flash already failed | Kimi K3 | Higher overall ceiling |
| Latency-critical chat | FlashX or a fast Flash host | Up to 200 tok/s |
Put simply: a Flash model is enough when the task is well-specified and checkable. Agent loops with tests, terminal commands with exit codes, extraction with a schema and office tasks with a clear deliverable all give the model feedback, and Flash uses that feedback well enough that its lower ceiling rarely shows. You need K3 when the answer depends on what the model knows or how deeply it reasons, and nothing downstream will catch a confident mistake.
A routing setup for cost-conscious teams
The pattern we recommended for GLM-5.2 works even better with Flash, because the price gap is wider and the capability gap is narrower:
- Default every request to GLM-5.3-Flash at
reasoning_effort: high, or FlashX for interactive surfaces. - Escalate to K3 on explicit triggers: a failed test run after two attempts, a research or math question, a document over a few hundred thousand tokens, or a user-facing answer that has to be factually right.
- Log per-task cost and success rate for both. K3 costs about 8x more per task, so if it rescues fewer than one in eight escalated tasks, your escalation rules are too loose.
Both models expose OpenAI-compatible APIs, so routing is a model-ID change. For another take on the same small-vs-flagship question, see our Kimi K3 vs DeepSeek V4.1 Flash comparison.
Honest caveats
- Index versions move. Z.ai’s launch materials cited a score of 57 on the older Intelligence Index v4.1.1. The 44 vs 42 gap is from the harder v4.3.2, so only compare models on the same version.
- Z.ai’s launch benchmarks are vendor-run and compare against Claude Opus 4.8, which Artificial Analysis now marks as deprecated. None of them include K3.
- Context specs disagree slightly. Z.ai’s docs say 1M tokens, OpenRouter’s listing shows about 1.3M, and the Hugging Face model card notes that its evaluations used a 300K maximum. Test at your real context length before relying on the top end.
- Hosting geography. Flash was trialed on Chinese domestic chips and Z.ai’s API is a Chinese provider, the same situation as Moonshot’s first-party API. Teams with residency rules should use a US or EU host, or self-host, for either model.
- Prices change. Both vendors have repriced before. Check the official pages linked above before you commit a budget.
FAQ
Is GLM-5.3-Flash better than Kimi K3? Not overall, but closer than the size gap suggests. On Artificial Analysis’s current Intelligence Index (v4.3.2), Kimi K3 at max effort scores 44 and GLM-5.3-Flash scores 42. K3 wins on Humanity’s Last Exam (47% vs 40%), SciCode (59% vs 52%), long-context reasoning AA-LCR (89% vs 80%) and the AA-Omniscience knowledge index (20 vs 7). Flash wins on Terminal-Bench 4.0 (33% vs 13%), GDPval-AA (1,641 vs 1,524) and AutomationBench-AA (60% vs 58%).
How much cheaper is GLM-5.3-Flash than Kimi K3? On list price, 20x cheaper on input ($0.15 vs $3.00 per million tokens), 10x cheaper on cached input ($0.03 vs $0.30) and 30x cheaper on output ($0.50 vs $15.00). Per completed task the gap shrinks, because Flash writes more tokens: Artificial Analysis measures about $0.25 per task for Flash versus $2.00 for K3, roughly 8x.
Is GLM-5.3-Flash free? No. Z.ai’s pricing page lists no free tier for GLM-5.3-Flash, and the 50% launch discount ended at 24:00 on September 9, 2026 (UTC+8). Z.ai’s free API models are older ones: GLM-4.7-Flash, GLM-4.5-Flash and GLM-4.6V-Flash. On Cloudflare Workers AI, Flash requires the Workers Paid plan or prepaid AI Gateway credits.
What license does GLM-5.3-Flash use compared with Kimi K3? GLM-5.3-Flash ships under the plain MIT license on Hugging Face and ModelScope. Kimi K3 uses the Kimi K3 License, which requires visible Kimi K3 attribution above 100 million monthly active users or $20 million in monthly revenue, and requires a separate agreement for Model-as-a-Service businesses above $20 million in revenue over 12 consecutive months. Most teams never hit those thresholds, but resellers and API hosts should read it.
Can I self-host GLM-5.3-Flash more easily than Kimi K3? Yes, by a wide margin. GLM-5.3-Flash’s native FP8 checkpoint is about 306 GiB before KV cache and fits one 8-GPU Hopper-class node or a GB200 tray. Kimi K3’s weights are about 1.56TB and a full serving setup needs roughly 1.7 to 1.84TB of GPU memory, meaning 8x B300 or 16x H200 class hardware at minimum.
What is GLM-5.3-FlashX? GLM-5.3-FlashX is a faster serving tier Zhipu launched on September 18, 2026, rated at up to 200 output tokens per second versus roughly 30 to 50 for standard Flash. Z.ai prices it at $0.37 per million input tokens, $0.075 cached and $1.25 output, still far below Kimi K3’s $3 and $15. As of late September it is not included in the GLM Coding Plan.
When should I use Kimi K3 instead of GLM-5.3-Flash? Use K3 when a wrong answer costs more than the token bill: hard reasoning and science questions, knowledge-heavy work where hallucination hurts, analysis across very long documents, and tasks where Flash has already failed once. Use Flash as the default for high-volume coding agents, terminal automation, extraction and office-style tasks, where independent tests show it matching or beating K3 at a fraction of the cost.
The bottom line
GLM-5.3-Flash has made “good enough” very cheap. At $0.15 / $0.50 per million tokens, MIT-licensed and small enough for one GPU node, it lands within two index points of Kimi K3 and beats it on the agentic tests that make up most production traffic. K3 still has the higher ceiling on reasoning, knowledge and long-context work, and it scores higher on knowledge reliability, but you now pay roughly 8x per task for that edge.
For a cost-conscious team the move is clear: run Flash as the default, keep K3 behind explicit escalation rules, and measure what the escalations actually fix. If you are choosing between Zhipu’s flagship and K3 instead, start with our Kimi K3 vs GLM-5.2 breakdown. For the K3 side of the stack, see the full Kimi K3 specifications and the open weights release guide.