GLM-5.3 vs Kimi K3
GLM-5.3 scores higher on the Models at Work leaderboard (81 vs 71 out of 100). They cost about the same per million tokens ($2.15 vs $2.31 blended). On Artificial Analysis' independent Intelligence Index, GLM-5.3 leads 45 to 44.
| GLM-5.3 · Z.ai | Kimi K3 · Moonshot | |
|---|---|---|
| Leaderboard rank | #5 | #16 |
| Score (0–100) | 81 | 71 |
| Tier | Open weights | Open weights |
| Price per 1M tokens | $1.40 in · $4.40 out | $2.31 blended (estimate) |
| Context window | 1M | Not verified |
| Open weights | Yes | Yes |
| Artificial Analysis Intelligence Index | 45 | 44 |
| Arena Text score | 1478 (rank 27) | — |
| Scale SWE-Bench Pro V2 (mini-swe-agent) | 84.30% | — |
| Scale MCP Atlas | 84.20 | — |
| Arena WebDev score | 1622 (rank 22) | — |
| Arena text score | — | 1488 |
| Best for | Agentic coding in Claude Code / Pi-style harnesses, Security research and code auditing, Self-hosted frontier-ish reasoning | Self-hosting, Open-licence requirements, Arena-style chat quality |
Choose GLM-5.3 if…
- you need agentic coding in claude code / pi-style harnesses
- you need security research and code auditing
- you need self-hosted frontier-ish reasoning
The open model people actually ship coding agents on. Same 744B base as GLM-5.2 with heavy agentic post-training; drops into Claude Code harnesses and holds its own on SWE-Bench Pro V2 and MCP Atlas at a third of frontier prices.
Caveats: License is muddled: AA lists a custom 'GLM-5.3 License' with commercial restrictions while Arena lists MIT. Read it before you self-host. Permissive on offensive-security tasks; that is a feature for red teams and a governance problem for everyone else. Open weights shipped ~2 weeks after the API; HN worried the public weights were safety-tuned differently.
Choose Kimi K3 if…
- you need self-hosting
- you need open-licence requirements
- you need arena-style chat quality
The highest-placed open-licence model on Arena, and the one most teams have not tried yet.
Caveats: Price is Artificial Analysis' blended figure, not a vendor list price; context not verified from a primary source.
What practitioners say
GLM-5.3: One of the biggest HN launches of the year (1,171 points, 584 comments). Practitioners report it 'fits Claude Code as its own, zero issues', that '$5 in tokens' of Claude work costs '$0.50' on GLM via Pi, and that it 'routinely finds bugs missed by Fable and Sol' in audits. The cyber angle was the controversy: users bragged about adapting kernel exploits that 'Claude and Opus outright refused', and others flagged the risk of nerfed public weights.
Questions people ask
Which is better, GLM-5.3 or Kimi K3?
GLM-5.3 scores higher on the Models at Work leaderboard (81 vs 71 out of 100). They cost about the same per million tokens ($2.15 vs $2.31 blended). On Artificial Analysis' independent Intelligence Index, GLM-5.3 leads 45 to 44.
When should I choose GLM-5.3 over Kimi K3?
Choose GLM-5.3 for agentic coding in claude code / pi-style harnesses, security research and code auditing, self-hosted frontier-ish reasoning. The open model people actually ship coding agents on. Same 744B base as GLM-5.2 with heavy agentic post-training; drops into Claude Code harnesses and holds its own on SWE-Bench Pro V2 and MCP Atlas at a third of frontier prices.
When should I choose Kimi K3 over GLM-5.3?
Choose Kimi K3 for self-hosting, open-licence requirements, arena-style chat quality. The highest-placed open-licence model on Arena, and the one most teams have not tried yet.
Which is cheaper, GLM-5.3 or Kimi K3?
GLM-5.3 costs $1.40 per million input tokens and $4.40 per million output tokens on Z.ai's list price (Z.ai Pricing, read Oct 11, 2026). Moonshot does not publish a simple list price for Kimi K3; Artificial Analysis estimates a blended $2.31 per million tokens (three input to one output).
Scores come from the Models at Work leaderboard: Score is 0 to 100 and computed, not typed: 55% independent evals (Artificial Analysis, Arena, Scale SEAL, Epoch, each scaled against its natural floor and the best score in this table, then averaged), 15% blended price on a fixed log scale ($0.05 per million is 100, $60 is 0), 15% practitioner sentiment from Signals and community threads, 15% operational fit (context window, open weights). A missing component drops out and the rest are reweighted. A model with no independent eval yet is provisional and ranks below every measured one. Within three points is a tie. Vendor-published figures are labelled and never counted. Prices are vendor list prices where published; blended figures assume three input tokens per output token.