AI Vendors Now Sell the Meter, Not Just the Model

A gateway for AI traffic

Snowflake, the company says, has released its Cortex AI Gateway into public preview as a central entry point for AI traffic, letting administrators define an explicit list of models the gateway will serve. m:…5916

Usage history, the company says, is retained for 365 days with a granular credit breakdown for requests that touch multiple models, and a monitor role can query that data without permission to send inference requests. m:…5916

Shared resource budgets, the company says, can trigger stored procedures at set threshold percentages, and per-user quotas can block AI requests within minutes of reaching a limit. m:…5916

Budget checks themselves run on a schedule: evaluation occurs periodically, up to 6.5 hours apart, or one hour with a low-latency option, so a budget alert can lag the spending it describes. m:…5916

The distinction matters because budgets alone do not stop spending: a Snowflake cost guide explains that budgets notify users and run attached procedures but do not block spend, and budget thresholds can take up to eight hours to act, or two hours with latency optimisation. m:…5880

Only quotas actually stop the spend

Per-user quotas are the only AI cost control with built-in enforcement, according to the same guide, and quota blocks apply to AI functions, Cortex Agents, Snowflake CoWork and the AI Gateway. m:…5880

Even quotas are not instant: blocks occur within minutes rather than in real time, so some overshoot remains possible, and anomaly detection monitors spend only once per day and needs at least 30 days of history. m:…5880

Warehouse compute and AI credits are also distinct units that cannot be added together, the guide notes, so finance teams that merge them into one AI line will misread where the money went. m:…5880

Other controls have their own ceilings: anomaly monitors are limited to 20 per account, and changes to tags on objects can take up to six hours to show up in tag-based budgets, so a reorganisation of cost centres does not reattribute spending immediately. m:…5880

The practical lesson is procedural: set quotas before agents go wider, treat budgets as alerts rather than brakes, and keep an audit trail long enough to attribute a surprising bill to a specific model or user. m:…5916 m:…5880

Cheap decisions become the product

Baseten, the company says, has released joey-1, a decision model with 2 billion parameters, 20 millisecond median end-to-end latency, and pricing of one cent per million tokens. m:…7269

Decision models, the company says, return a label and a confidence score in one forward pass, which suits high-volume tasks where a full text-generating model would be wasteful, and joey-1 supports a 16,000-token context window on the company's model APIs. m:…7269

joey-1 itself is post-trained from a 2-billion-parameter open model, the company says. m:…7269

Canva, the company says, launched Canva Brush on October 29, a vision-language model for design edits that is 69 percent faster and 46 percent cheaper per edit task than its predecessor, using an agentic loop that writes and runs code to adjust design elements. m:…3515

Canva says the model was trained using large-scale reinforcement learning, and its pitch is deliberately narrow: faster and cheaper edits for general design work, not a claim to replace the designer. m:…3515

Taken together, the three launches reframe the buying question: not which model is most capable in isolation, but which vendor lets a buyer cap, attribute and audit AI usage before the bill arrives. m:…5916 m:…7269 m:…3515