Netlify ran one prompt through 11 models: Claude Opus 5 burned 519 credits, DeepSeek V4 Flash did it for 2.4

Netlify published an AXIS-framework evaluation on August 14 that I've been thinking about all day. Same task, 11 models, three runs each, scored on functional correctness rather than aesthetics. The task was deliberately boring: a static one-page coffee shop site with hours, address, menu and a photo. Explicitly no CMS.

Average credit cost, cheapest to most expensive: DeepSeek V4 Flash 0731 at 2.4. Kimi K2.7 Code at 19. GLM 5.2 at 27. GPT 5.6 Terra at 39. Gemini 3.1 Pro at 53. Kimi K3 at 102. GPT 5.6 Sol at 141. Claude Sonnet 5 at 143. Claude Opus 5 dead last at 519.

That's a 216x spread on an identical task where all of them produced a working page. Opus 5 spent 216 times what DeepSeek V4 Flash spent to build the same coffee shop website.

I've done this to myself. I have a default in my head that says "use the best model, it's worth it," and for hard architectural work it usually is. For scaffolding a static page it's just setting money on fire. Frontier reasoning models overspend on simple work because they're built to explore, and simple work has nothing to explore.

The fix arrived the same week. LLMRouter from Tao Feng, Jiaxuan You and colleagues at UIUC (arXiv 2608.06867) hit 2,340 HuggingFace upvotes: learned routers outperform the strongest fixed-model baseline by 14.6% relative, and lightweight routers get more competitive as cost constraints tighten. They open-sourced 16+ representative routers and an xRouteBench evaluation platform covering single-turn, multi-turn and personalized routing. The problem and its answer landed within a week of each other.

Urgency on this just went up. DeepSeek raised API prices 50% to more than 1,100% depending on model, token type and time of day, effective 16:00 UTC on August 16. V4-Pro cache-miss input goes from $0.435 to $1.32 per million at peak, with peak windows at 01:00–04:00 and 06:00–10:00 UTC and off-peak at half. If you have V4 in a production loop, you have two days to re-price.

Do this today: pick the three dumbest, highest-volume tasks in your pipeline. Boilerplate generation, file scaffolding, commit message writing. Route them to a cheap model and diff the output against what your expensive model produces. If you can't tell the difference, you just found your margin.