Gemini 3.7 Flash, DeepSeek V4-Pro, Grok 4.6 and an Ultrafast tier for GPT-5.6 Sol all landed inside twenty four hours, and what separates them is cost per task, tokens per second and where the model plugs in. The same day's papers point at a less comfortable variable: tool interface shape, skill format and context compaction move agent behavior without touching weights. DeepSeek opened its own agent harness, and a Claude Code maintainer published the daily routines their team runs against shipping apps. The survey data, meanwhile, says the human queue has moved to review, where no release helps.
Gemini 3.7 Flash arrived three weeks after 3.6 Flash at half the introductory price per million tokens, and independent benchmarking puts the four point Intelligence Index gain mostly in agentic and terminal work rather than raw knowledge. It is already selectable inside GitHub Copilot.
A dated brief from the vibe-coding frontier. Today’s Intel.
OpenAI previewed an Ultrafast tier running GPT-5.6 Sol at up to 750 output tokens per second on Cerebras hardware, and LongCat-2.0 is free in Nous Portal for one week on the back of a 70.8 on Terminal-Bench 2.1. Speed and access moved further today than capability did.
DeepSeek opened Harness v0.1 as a pluggable agent runtime in developer preview, a Claude Code maintainer published the daily maintenance routines their team runs against shipping apps, and Cursor cut cloud agent startup to a third of its previous latency. The engineering effort is visibly moving to the layer above the model.
Three papers land on the same claim from different directions: tool interface shape, skill format and harness structure change agent behavior with the weights held fixed. One reports program shaped skills costing less than prose ones over long horizons, another improves an embodied agent by evolving only its skills and harness around a frozen model.
Context compaction can quietly drop a side constraint given earlier in a session. Every long running agent is exposed to this, and the failure leaves no error to catch, so it stays folklore until someone measures it.
A survey of 1,528 developers puts the bottleneck at reviewing AI written code rather than producing it, alongside a separate 1,100 person survey reporting 42 percent of committed code as AI generated and most respondents not fully trusting it. The productive question is not how much to read, it is which cheap checks can fail a change without a human opening the diff.