Open Weights vs. APIs

Show notes

For the first time, you can download a model that trades punches with the best API models on the planet and run it on hardware you own. Rocco and Ruby run the actual math — API price ladders, open-weight API pricing, self-host economics — plus the four-question decision framework, the license traps hiding behind the word "open," and the API deprecation dates that prove you're renting.

Segments

• The Build — October's open-weight deluge: Meta Muse Spark Open (816B/42B-active, 1M context), Reflection Beam (501B, Apache 2.0), DeepSeek V4.1-Flash beating Claude Opus 5 on coding benchmarks, NVIDIA Nemotron 3.5. The price ladder: GPT-6 Astra $10/$50, Sol $2/$10, Luna $0.10/$0.50 per million tokens — versus gpt-oss-120b at $0.15/$0.60 and $0.04/$0.17. The 84%-of-tokens-go-non-frontier stat. The four-question framework (volume, data residency, swapability, license). The Muse Model License $20M revenue clause, custom licenses vs. MIT/Apache 2.0. And the shutdown calendar: GPT-4.1 nano gone Oct 23, gpt-5 snapshots Dec 11, GPT-4o Jan 14.
• Loud Links — aetherneural.com: the sovereign AI research lab. Free weekly Friday notes on AI operations; November's series is all about agents.
• Hot Take — "The frontier API is becoming a luxury good, and luxury goods don't win infrastructure wars." Three receipts and the steelman: the API survives as the lab, the weights become the factory. Rent the frontier to learn; own the weights to earn.
• Demo Derby — Muse Spark Open shipping real iOS changes on linked Mac Studios, DeepSeek open-sourcing its DeepInfer inference engine plus AWS TorchNeuron's native PyTorch support, Reflection Beam as an "open operating system for agents" with Dell/HPE on-prem boxes, NVIDIA's Nemotron Bolt models on RTX Spark AI PCs — plus the anti-demo of the week: the $50 ProjectDiscovery backdoor, and why provenance comes before behavior.

Links & references

• Meta releases open-weight Muse Spark modelshashi.co →
• Muse Spark Open: 816B parameters under the Muse Model Licenseainotis.com →
• Reflection releases Beam's open weights under Apache 2.0unite.ai →
• DeepSeek V4.1-Flash beats Claude Opus 5 on some coding testspbxscience.com →
• AI Brief, October 20: Nemotron 3.5, TorchNeuron, DeepInfersean-weldon.com →
• ChatGPT API pricing 2026: token costs and rate limitsintuitionlabs.ai →
• Claude API pricing per million tokens, all models (Oct 2026)dev.to →
• Hugging Face for business: inference, endpoints and costforasoft.com →
• Backdooring open-weight AI modelsschneier.com →
• With open-weight AI, more of the security burden falls on enterpriseshelpnetsecurity.com →
• Aether Neural — this week's Loud Linksaetherneural.com →
• Showcase — demos and projects on our radaraiextrovert.com →

Coming up next

Episode 5 — "AI Video's Creator Moment" — generative video moves from novelty clips to real pipelines. Out now.