Kimi K3 arrived with numbers that look capable of changing coding-agent economics. GLM-5.2 provides a useful counterexample: it has a directly comparable coding-agent result, open weights, and strong long-context credentials, but its measured hosted task cost is much less attractive.
We added both models to the Aeon Model Economics Lab. They appear differently for a reason:
- GLM-5.2 is a measured circle because score, task cost, and output tokens come from the same Artificial Analysis Coding Agent Index run.
- Kimi K3 is a provisional diamond because its current public score and economics come from a different Artificial Analysis model suite.
The visual distinction prevents an exciting launch from becoming a false ranking.
What the updated plot shows
The permanent plot compares coding score on the vertical axis, completed-task cost on the horizontal axis, and output tokens through bubble area. Every point carries a model and reasoning-effort label.
| Model and setting | Score | Cost per task | Output tokens | Evidence |
|---|---|---|---|---|
| Luna medium | 58.7 Coding Agent Index | USD 0.47 | 15.2k | Measured, comparable |
| GLM-5.2 Max | 57.9 Coding Agent Index | USD 6.47 | 40.6k | Measured, comparable |
| Grok 4.5 high | 76.4 Coding Agent Index | USD 2.59 | 40.0k | Measured, comparable |
| Terra max | 77.4 Coding Agent Index | USD 2.76 | 60.7k | Measured, comparable |
| Kimi K3 max | 76.2 Coding Index | USD 0.94 | 23.4k | Provisional, cross-suite |
The K3 row is intentionally not a like-for-like comparison. Its score is the Artificial Analysis Coding Index, while its cost and token figures come from the general Intelligence Index task suite. The other measured rows use the Artificial Analysis Coding Agent Index and its task economics.
GLM-5.2: comparable, open, and expensive in this hosted run
Z.ai released GLM-5.2 for long-horizon coding and agentic work with High and Max thinking settings, a 1 million token context window, and open weights under an MIT license. Its published API pricing is USD 1.40 per million input tokens, USD 0.26 per million cache-hit tokens, and USD 4.40 per million output tokens.
Artificial Analysis measured GLM-5.2 in Claude Code across DeepSWE, Terminal-Bench v2, and SWE-Atlas-QnA:
- Coding Agent Index: 57.9
- Cost per task: USD 6.47
- Output tokens per task: 40.6k
- Total token usage per task: approximately 6.5 million
- Active agent time per task: approximately 25.2 minutes
The benchmark row displays GLM-5.2 without an effort suffix. We label the point Max to align with Artificial Analysis's reasoning-variant naming and Z.ai's published effort options. It remains one measured configuration, not a measured High-to-Max curve.
The economic comparison is direct. Luna medium reaches 58.7 at USD 0.47 with 15.2k output tokens. GLM-5.2 reaches a similar 57.9 score at roughly 13.7 times the task cost and 2.7 times the output tokens in the same index.
That does not make GLM-5.2 irrelevant. It changes the buying thesis. GLM-5.2 is interesting when open weights, private deployment, a 1 million token context window, provider choice, or model ownership creates value that a hosted benchmark bill does not capture. It is not the obvious hosted API route for routine coding work on this evidence.
Sources: Z.ai GLM-5.2 release and Artificial Analysis coding-agent results.
Kimi K3: the visual frontier is exciting, but provisional
Kimi K3 launched on July 16, 2026 with max reasoning effort, a 1 million token context window, and published API pricing of:
| Token category | Price per million tokens |
|---|---|
| Cache-hit input | USD 0.30 |
| Cache-miss input | USD 3.00 |
| Output | USD 15.00 |
Kimi reports max-effort coding results of 67.5 on DeepSWE, 77.8 on Program Bench, 88.3 on Terminal-Bench 2.1, and 81.2 on FrontierSWE. Artificial Analysis independently measured K3 at 57.1 on its Intelligence Index and 76.2 on its Coding Index, with approximately USD 0.94 and 23.4k output tokens per general-intelligence task.
On the plot, that provisional marker sits near the measured 76 to 77 point cluster while remaining below USD 1 per task and using a relatively small token bubble. If a same-harness Coding Agent Index run confirms that location, K3 would materially change the cost-capability frontier.
But the current marker does not prove that K3 beats Grok high, Terra max, GPT-5.5 xhigh, Fable max, Luna max, or GLM-5.2. It combines a different score and task suite. It is best read as a high-upside research hypothesis.
K3 also exposes an important cost tension. Its cached-input price is low, but output costs USD 15 per million tokens. Completed-task economics will depend on cache behavior, reasoning length, retries, tool calls, supervision, and rework. Cheap input does not guarantee a cheap completed task.
Sources: Kimi K3 official launch and Artificial Analysis Kimi K3 measurements.
The routing conclusion
The updated plot supports four different decisions:
- Use Luna and similar efficient measured points for routine implementation when they clear the quality bar.
- Reserve Terra, Grok, or Sol frontier settings for work that can repay the additional cost.
- Evaluate GLM-5.2 when open weights, private deployment, long context, and provider control matter enough to offset weaker hosted task economics.
- Test Kimi K3 now, but hold the harness and workload constant before treating its provisional visual position as a routing decision.
For K3, the next decisive evidence is a same-harness coding-agent run with score, cost, output tokens, and execution configuration reported together. Low and High effort results would then reveal whether K3 has an economically useful routing curve rather than one exceptional launch point.
This is the core Model Economics Lab thesis: model selection is not a leaderboard purchase. It is a routing and control problem. The best operating point is the least expensive configuration that clears the quality, data-control, and failure-cost threshold for the work.
Explore the updated Aeon Model Economics Lab, filter all 29 operating points, and download the CSV or JSON dataset. For a workload-specific benchmark and routing analysis, contact info@airiskmanagement.ca.