[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks’ +60% Astra cost)

Steve Yegge has shut down Gas Town and said that despite spending many thousands of dollars a month on coding agent subscriptions, he only ever built Gas Town with it. The disclosure was noted by Dan Luu, who said he had not found ultra "vibed" orchestrators useful because of reliability in completing tasks, and observed that the author of the most famous one had the same issue.

Databricks is reporting a 60% increase in overall spend when its AI engineers switch to Astra. The increase stands alongside reports that Astra is often cheaper than Sol in terms of cost per task by many benchmarks, attributed to token efficiency, though it is not universally cheaper everywhere.

Elsewhere in the 9/15/2026-9/16/2026 news roundup, OpenAI published a formal framework for tracking, investigating and disclosing model misalignment incidents, plus six case reports from the last six months. Xiaomi's @_LuoFuli announced the MiMo-V2.6 RL run with live training stats, harness mix, reward details and cost telemetry; @eliebakouch estimated roughly $493k/day for the 1T-class Pro run and $247k/day for Flash. @pwendell reported that Databricks rolled out GPT-6 Astra to about 3,500 engineers after a pilot with about 200 users, with Astra outperforming prior top-end models on complex long-horizon tasks. Demis Hassabis and Shane Legg launched the DeepMind Institute. @cline made Union Alpha free in Cline, claiming near GPT-6 Astra / Opus 5-class coding performance at far lower cost. In local-model discussion, a 30-day Qwen3.8-27B deployment reported a mean 73.8 tok/s generation, and Swift-Qwen3.8-27B cut completion tokens from 12,547 to 7,301 in one Aider coding eval.