GLM-5.3 vs Kimi K3.1 Pro: Benchmarks, Price, Vision, and Licenses

TL;DR

Last updated November 9, 2026.

Kimi K3.1 Pro is the stronger model; GLM-5.3 is the cheaper one. Moonshot's Kimi K3.1 Pro (2.8T total parameters, 104B active, API launch October 23, 2026) leads GLM-5.3 on every row where Moonshot's benchmark table lists both, including Terminal-Bench 4.0 (65.7 vs 41.8) and DeepSWE v1.1 (75.5 vs 66.9). The independent check points the same way: Artificial Analysis Intelligence Index v5 scores K3.1 Pro 46 and GLM-5.3 (max) 32. K3.1 Pro also accepts images; GLM-5.3 is text-only. GLM-5.3 (Z.ai, 753B total, August 2026) wins on price: $1.40 input and $4.40 output per million tokens on Z.ai, against $2 and $10 for K3.1 Pro on Moonshot. Its license is looser for API hosts: Z.ai asks for a security review above $10 billion in revenue, while Moonshot requires a separate agreement above $20 million. Both have open weights and 1M-token context. On Morph, morph-glm53-744b costs $1.19/$6.00 and morph-kimik31 costs $2.50/$14.00, and standby halves both.

The 30-Second Verdict

The two models split cleanly on capability versus cost. Find the axis your workload lives on and the choice is usually obvious.

• Optimizing for capability? Kimi K3.1 Pro. It wins every shared row of Moonshot's table, and the independent Intelligence Index v5 has it 14 points ahead.
• Optimizing for cost per token? GLM-5.3. Kimi K3.1 Pro's output price is about 2.3x GLM-5.3's at list ($10 vs $4.40) and 2.3x on Morph ($14.00 vs $6.00).
• Need vision? Kimi K3.1 Pro only. GLM-5.3 takes text only.
• Self-hosting? GLM-5.3 is about half the download (756 GB vs 1.56 TB on Hugging Face).
• Building a hosted API business? Read both licenses. GLM-5.3's extra clause starts at $10B in revenue; Kimi K3.1 Pro's starts at $20M.
• Paying a flat rate? Morph Ultra ($200/month) includes unlimited Kimi K3.1 within per-subscription rate limits. GLM-5.3 is not in Ultra, so on that plan K3.1 Pro is the cheaper of the two to run.

GLM-5.3 vs Kimi K3.1 Pro: Spec-by-Spec

Kimi K3.1 Pro is about 3.7 times GLM-5.3's total size and runs roughly 2.6 times the active parameters per token. Total size sets memory and self-host cost. Active size sets compute per token.

The active-parameter figure for GLM-5.3 comes from The Batch, not from Z.ai, so treat it as reported. Everything else in the GLM-5.3 column is from Z.ai's docs or the Hugging Face metadata. The Kimi K3.1 Pro column is from Moonshot's model card and API quickstart. For each model on its own, see the GLM-5.3 and Kimi K3.1 pages.

GLM-5.3 vs Kimi K3.1 Benchmarks: Vendor vs Independent

There is no clean head-to-head from either lab. Moonshot's Kimi K3.1 Pro table includes a GLM-5.3 column. Z.ai's GLM-5.3 table came out two months earlier and compares against the original Kimi K3, not K3.1. Every number below says who reported it.

Moonshot's table: Kimi K3.1 Pro vs GLM-5.3

Read this table with two filters. First, harnesses differ. Moonshot ran K3.1 Pro in Claude Code, while many GLM-5.3 cells are imported. The DeepSWE 66.9, for example, is Z.ai's own number, which Z.ai ran on mini-swe-agent at a single pass. Second, where Moonshot imported a figure it matches Z.ai's: DeepSWE 66.9, SWE Marathon 42.5, ExploitBench 54.4, and HLE with tools 62.5 all appear in both vendors' tables. Where the version or runner changed, GLM-5.3 looks worse in Moonshot's table: PostTrain Bench is 34.5 on Moonshot's v1.2 run against 39.8 in Z.ai's, and GDPval-AA is 1650 on v2.1 against 1769 on v2. Swap in Z.ai's own figures and Kimi K3.1 Pro still leads every row except PostTrain Bench, where Z.ai's 39.8 would top K3.1 Pro's 36.5 on a different version of the benchmark. The direction is solid. The size of the gap on any single row is less certain than it looks.

Z.ai's table: GLM-5.3 vs the original Kimi K3

Z.ai's own table shows GLM-5.3 roughly level with the original Kimi K3 on coding and ahead on cyber. Kimi K3.1 Pro moved that line. Do not compare the Terminal-Bench rows across the two tables: Z.ai reports versions 2.1 and 3.0, Moonshot reports 4.0, and they are different task sets. Z.ai ran Terminal-Bench 3.0 in Claude Code at max effort, averaged over three runs.

The independent numbers

Artificial Analysis is the one evaluator that scores both models on the same version of the same index. On Intelligence Index v5, released October 27, 2026, Kimi K3.1 Pro scores 46 and GLM-5.3 at max reasoning scores 32. Use only v5 for this comparison. Kimi K3.1 Pro launched at 55 on the earlier index (October 23, 4th overall), and v5 rescored every model, so mixing versions produces nonsense. Artificial Analysis also ran Terminal-Bench 4.0 itself and got 67% for K3.1 Pro, the highest of any model at launch. That is a separate measurement from Moonshot's 65.7. Neither lab reports SWE-bench Verified for these models; a number you see quoted for either is not from the vendor.

GLM-5.3 vs Kimi K3.1 Pricing and Token Economics

At list price Kimi K3.1 Pro costs about 1.4x GLM-5.3 on input and 2.3x on output. On Morph the ratios are 2.1x and 2.3x. The one leg where K3.1 Pro is cheaper at list is cache reads: $0.20 at Moonshot against $0.26 at Z.ai.

Two honest notes on Morph's rates. For Kimi K3.1 Pro, Morph's per-token price is above Moonshot's on every leg. For GLM-5.3, Morph is below Z.ai on input and cached input and above it on output. The reasons to run either on Morph are standby at 50% off, the flat Ultra plan for K3.1, measured speed, and one key for both models. They are covered in the Morph section below.

Both models think before they answer, and both default to max effort. Each vendor publishes what the lower setting costs you, on its own benchmark:

• Kimi K3.1 Pro (Moonshot's Terminal-Bench 4.0 cost chart): 52.8% at high effort for $4.16 per task, 65.7% at max for $5.25 per task. Dropping to high saves about a fifth of the cost and gives up 13 points.
• GLM-5.3 (Z.ai Code Bench, an in-house benchmark): 31.4% at high effort with about 50K output tokens per task, 34.5% at max with about 75K. Dropping to high cuts a third of the output tokens and gives up 3 points.

These are different benchmarks, so they do not rank the two models against each other. They do say where the effort lever pays. One more Kimi K3.1 Pro detail: Moonshot requires the full assistant message, including reasoning_content and tool_calls, to be passed back on every turn. Long agent runs replay a lot of reasoning as input, so the cache-read rate matters more than the fresh-input rate. On Morph, a session key (prompt_cache_key or x-session-id) keeps those turns on cache. See the caching docs.

Open Weights and Licenses

Both models are downloadable today. GLM-5.3 shipped first on Z.ai's GLM Coding Plan on August 14, 2026, and the zai-org/GLM-5.3 repo on Hugging Face was created on August 25. Kimi K3.1 Pro reached Moonshot's API on October 23, and the moonshotai/Kimi-K3.1-Pro repo was created on November 2; Moonshot had promised the weights by November 4.

Both licenses start from an MIT-style grant: use, copy, modify, fine-tune, and redistribute. Each adds terms for companies that sell model access, which both define as giving a third party inference or fine-tuning access with meaningful control over inputs, parameters, or training data.

Kimi K3.1 Pro's license adds one more rule: a product or service with more than 100 million monthly active users, or more than $20 million in monthly revenue, must display "Kimi K3.1-Pro" prominently in its interface.

Vision, Context, Cyber, and Self-Host

• Vision. Kimi K3.1 Pro has a 401M-parameter MoonViT-V2 vision encoder and takes text and images. Moonshot's API also takes video through its Files API, and it does not accept public image URLs: send base64 or an uploaded file id. Z.ai's docs say GLM-5.3 "currently supports text-only inputs." Moonshot's table shows K3.1 Pro at 84.1 on MMMU-Pro, with no GLM-5.3 entry.
• Context. Both read 1M tokens. Kimi K3.1 Pro can also write up to 1,048,576 tokens on Moonshot's API (131,072 by default); GLM-5.3 tops out at 128K. Moonshot replaced K3's full-attention layers with Gated DeepSeek Sparse Attention "for more efficient long-context modeling," and prices its window flat, with no tier by context length.
• Cyber. GLM-5.3 is the cyber-positioned release. Z.ai reports 84.5 on CyberGym and says the model found 2,436 vulnerabilities, 1,097 of them medium or high severity. Moonshot went the other way: for the open-weight release it says it "identified and removed cyber-related data from both training stages." Even so, Moonshot's own table puts K3.1 Pro ahead on ExploitBench (74.7 vs 54.4).
• Self-host. GLM-5.3's repo is about 756 GB, almost all FP8. Kimi K3.1 Pro's is about 1.56 TB even at native MXFP4, so plan on roughly twice the GPU memory before KV cache. Moonshot recommends vLLM, SGLang, or TokenSpeed for K3.1 Pro.

Which One to Pick, by Workload

Many teams route instead of picking. Run GLM-5.3 as the default for routine text work and send hard, long, or image-bearing tasks to Kimi K3.1 Pro. On Morph that is a model-string change on the same key. For sub-tasks that need speed more than depth, DeepSeek V4.1 Flash runs on the same key too.

Running GLM-5.3 and Kimi K3.1 Pro on Morph

Morph serves GLM-5.3 as morph-glm53-744b and Kimi K3.1 Pro as morph-kimik31, both with a 1M context window. Send service_tier: "standby" to pay 50% less on either (see the standby docs). The older ids morph-kimik3 and morph-glm52-744b still work and now serve the new weights; use the new ids in new code.

The Ultra plan caps subscriptions at 1,000 and rate-limits each one: 24,290 uncached input tokens and 6,162 output tokens per minute on Kimi K3.1. Inside those limits there is no per-token bill for K3.1. Morph is also an OpenRouter provider for moonshotai/kimi-k3.1-pro.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.morphllm.com/v1",
    api_key="YOUR_MORPH_API_KEY",
)

def ask(model: str, prompt: str, standby: bool = False) -> str:
    resp = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": prompt}],
        extra_body={"service_tier": "standby"} if standby else None,
    )
    return resp.choices[0].message.content

# Routine text work: GLM-5.3
print(ask("morph-glm53-744b", "Write unit tests for this parser.", standby=True))

# Hard or long-horizon work: Kimi K3.1 Pro
print(ask("morph-kimik31", "Plan the migration of this service to async I/O."))

Claude Code connects through Morph's Anthropic Messages endpoint. This setup runs Kimi K3.1 Pro as the main model; set ANTHROPIC_MODEL to morph-glm53-744b to run GLM-5.3 instead.

export ANTHROPIC_BASE_URL="https://api.morphllm.com"
export ANTHROPIC_AUTH_TOKEN="YOUR_MORPH_API_KEY"
export ANTHROPIC_MODEL="morph-kimik31"
export ANTHROPIC_SMALL_FAST_MODEL="morph-dsv41flash"
claude

Codex works too: add a provider with base URL https://api.morphllm.com/v1 and wire_api = "responses". To set up your coding agents in one step, run curl -fsSL https://www.morphllm.com/install.sh | bash && morph install. Get a key from the API keys page.

Kimi K3.1 Pro on Morph has three documented agent features. Dynamic tool loading adds tools mid-conversation. Agent Runs tag each run's requests with a run_id. Session caching uses prompt_cache_key or x-session-id to keep a conversation on its cache. For first-party setup and Moonshot's own endpoints, see the Kimi K3.1 API guide.

GLM-5.3 vs Kimi K3.1 Pro: The Trade-Offs

Frequently Asked Questions