X Debate: Theo Says Open-Weight Models Are Under 0.5% of T3 Code Traffic and Only 6% of Users Tried One in the Last 7 Days, Reigniting the Open vs Frontier Argument

Background and Industry Pain Point

T3 Code is an open-source "control plane for coding agents" from Theo's team. Per its own description it controls Claude Code, Codex, Cursor, Grok Build, OpenCode, Antigravity and other agents on your machine through desktop, web and mobile apps. On the night of November 11 Theo posted "Remember open weight models? Those were fun", which reached about 240K views. Around midday November 12 he quoted it with data: open-weight models make up less than 0.5% of T3 Code traffic. How much open models are actually used in real coding work has been a recurring argument this year.

Architecture and Mechanism

The definitions matter. According to Theo's replies, 0.5% is the share of all prompts sent to open-weight models; the other figure is that 6% of users used an open-weight model at least once in the last 7 days, down from about 15% a month earlier. T3 Code does not sell models; it mostly drives subscriptions users already have, and open models usually come in through channels such as OpenCode. So the numbers reflect the choices of developers who already use subscription agents.

Benchmarks and Measured Results

This is not a benchmark and no independent statistics have been published. Everything above is Theo's self-reported product telemetry, which we cannot verify. On engagement, the main post has about 83K views, 1,432 likes and 173 replies. In replies Theo also offered an analogy: buying a Honda for $30,000 versus renting a Ferrari for $200 a month, adding that anyone saying open weight is cheaper is being dishonest.

Developer Guide

The replies split three ways. One camp agrees: in a coding agent a bad turn costs far more than tokens, so people pay for frontier models. Another disagrees: T3 Code was built around subscriptions, and open models shine in local inference, narrow high-volume jobs such as classification and extraction, and security work that needs data control. Others separate "local" from "open weight", arguing the former is fading for now while the latter still has real uses. Practical advice: do not rely on someone else's traffic share; compare success rate, cost per task and data compliance needs on your own task set before setting a routing policy. To see how T3 Code connects to each agent, read its GitHub README.