GLM 5.2 vs Kimi 2.7: build a landing page, live.
Two models, one prompt, one job: build a complete landing page in a single shot. Both stream through the HYRE Gateway — one OpenAI-compatible endpoint, paid per call in USDC, no API keys for the agent. We recorded each run end-to-end, with the live token and cost meter on screen the whole way.
Watch them build it
Left panel streams the raw HTML from the gateway; right panel renders it live. The header shows the endpoint, the model, and a running latency / tokens / cost readout.
Result at a glance
| METRIC | GLM 5.2 | KIMI 2.7 |
|---|---|---|
| Time to first token | 9.73s | 3.75s |
| Total generation time | 40.9s | 55.1s |
| Output tokens | 2,402 | 1,660 |
| Throughput | 57 tok/s | 27 tok/s |
| HTML produced | 7,157 chars | 5,621 chars |
| Cost (live promo price) | $0.0059 | $0.0036 |
| Complete page | ✓ | ✓ |
Both models are 50% off on the gateway this week. Kimi 2.7's launch promo (hold 2M $HYRE → 1,000 free calls/week) runs through Jun 29.
The verdict
- Correctness: a tie. Both shipped a complete, structurally identical, polished dark landing page — nav, hero with dual CTA, a three-card feature grid, and a footer — from a single shot. No truncation, valid HTML, both reached </html>.
- Kimi 2.7 — value and a snappier start. ~39% cheaper this run and ~2.6× faster to first token, with a tighter, leaner page.
- GLM 5.2 — faster finish, richer output. ~2× the throughput, done ~14 seconds sooner, and a slightly more detailed page — at a higher price.
THE POINT
Two different open models, two real landing pages, billed by the token — for less than a cent each, with no key to manage and no plan to pick. That's the gateway: pick a model, pay per call, ship. Swapping GLM 5.2 for Kimi 2.7 is a one-line change.
Methodology
- Endpoint: POST gw.hyreagent.fun/api/inference/v1/chat/completions, streaming, temperature 0.4, max_tokens 8000. Both runs use the exact same system and user prompt.
- The task: one self-contained HTML landing page (inline CSS, no frameworks, no images) for a fictional autonomous trading agent. The model that writes the most complete, polished page wins on quality.
- Extended thinking disabled for both via chat_template_kwargs:{enable_thinking:false} — the benchmark compares output, not reasoning budget, and keeps the runs watchable. (Kimi K2.7 Code otherwise reasons for minutes before writing a line.)
- Cost = real token usage × the live launch-promo price (GLM 5.2 $0.60/$2.40, Kimi 2.7 $0.444/$2.10 per 1M in/out). Latency is noisy run-to-run; tokens and cost are the stable signals.
- The agent pays the gateway per call in USDC — no API key of its own, no subscription. The on-screen HUD reads straight from the live /models pricing.