Three vendors, 48 hours, three different ways to stop paying frontier prices

September 15 and 16, from three companies with nothing in common:

Salesforce announced Koa, its first CRM reasoning model, built by post-training NVIDIA Nemotron 3 Super on a synthetic dataset derived from three decades of CRM deployments. Salesforce holds the weights and runs inference on its own infrastructure, so customer data never crosses the trust boundary. It claims parity or better against leading models on its internal CRM action benchmark with three times fewer errors, with pilots at 1-800Accountant, Baxter Credit Union, Formula 1, UChicago Medicine and Xero.

TypeSafe AI released Jev, a model that returns typed probabilistic values instead of text. Diogo Almeida frames it as "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out." Pricing is $0.042 per million input tokens with output free, and the company claims 70-500ms end-to-end latency and hallucination that is structurally impossible because the output type is guaranteed.

Weave Router 2.0 took the top Product Hunt slot on September 16 with a Go proxy that sits in front of Claude Code, Codex, Cursor or OpenCode and routes each request across multiple provider subscriptions by which one has quota left. Weave's own analysis says 60-70% of Claude Code requests are short, structurally simple completions an open model handles at parity for about a fortieth of the cost.

Every benchmark number in that list is vendor-reported and unaudited. Discount them accordingly. The structural claim survives the discount: three unrelated companies decided within two days that inference cost is a line you engineer down rather than a bill you accept.

Pair this with CobbleDB. Perplexity engineered down its database bill using agents. Salesforce is engineering down its inference bill by owning a model. Both are the same move at different layers, and both were previously considered obviously not worth the engineering time.

The practical version for a small team: instrument what fraction of your agent traffic is routing, classification, and extraction versus actual reasoning. If it looks anything like Weave's 60-70%, you're paying frontier prices for work a 7B model does at parity. You don't need a router product to act on that. You need to read your own logs.