← sources

Same agent loop, 14 models, hard caps: what I learned building a photo-to-Blender agent

source: reddit (reddit:AI_Agents)published 2026-10-02 10:40ingested 2026-10-03 02:35

retained text

I'm building an agent that turns a photo into an editable Blender scene by writing and running code, and I benchmarked 14 models as its brain. Every run got the same brief and hard caps: 20 minutes, $4 and 60 requests per scene. Lessons that apply beyond 3D:

1. **Enforce limits outside the agent, and make it save early.** The brief asks for a usable scene early and improvements after, because the caps end the run by terminating it. A run that saved nothing scores 0.
2. **Models disagree about when they're done.** GPT-5.6 Sol and Terra declared the job finished after about 8 minutes with most of the budget unspent. GPT-6 Astra never stopped: it refined until the $4 cap on every scene, and won all three.
3. **Heavy reasoners can lose to the clock.** DeepSeek V4.1 Flash and Qwen3.8 Max spent 83% and 68% of their output reasoning and saved nothing in 20 minutes. Given 60 to 120 minutes, both built some of the scenes.
4. **Provider plumbing changes results.** Anthropic needed explicit cache markers: my first runs were 0% cached and burned the budget in 8 to 18 requests (96–97% once fixed). A Google endpoint rejected requests that ended on the model's own message until the gateway added a closing user turn. Kimi K3 once returned an empty first answer, so the bench now restarts a run that ends within a minute with nothing saved.
5. **Judge outside the agent.** The bench re-renders each delivered scene and scores it with deterministic code, renders a second view swung 35° to catch flat stand-ins, and checks that the photo wasn't reused as a texture.

Top of the board: GPT-6 Astra 66 ($3.91 an attempt), GPT-6.1 Sol 61 ($0.36), Claude Opus 5.5 60 ($1.19), Claude Sonnet 5.5 56 ($0.50). One run per model per photo. The full write-up with every render is in the comments.

document fetched 2026-10-03 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1wvqi05/same_agent_loop_14_models_hard_caps_what_i