Hosts: Lenar Kess, Damra Vol. Friday's news was almost entirely about price. Grok 4.6 posted the same score as Claude Fable 5 on Perplexity's WANDR benchmark at a third of the cost, Gemini 3.7 Flash shipped at half the price of 3.6 Flash, and OpenAI previewed a Sol variant whose pitch is latency rather than intelligence. Underneath that, two Apache-2.0 open-weight releases and a mathematical bound that a model broke and then refused to believe.Perplexity ran Grok 4.6 and Claude Fable 5 on its WANDR agentic-search benchmark: identical 0.496 scores, $7.58 per task versus $20.30. Perplexity has no stake in xAI, which is what makes the number usable.Elon Musk and X Freeze flagged Grok 4.6 at the top of CursorBench 3.2 for real-world coding tasks.Daniel McKinnon reported Grok 4.6 topping RareBench for rare-disease diagnosis ahead of Claude Opus 5 — a research benchmark, not a clinical result.The AI Daily Brief put the list price at $2 per million input tokens and $6 per million output, roughly $0.84 per benchmark task, and noted community reports of occasional truncated outputs.Cerebras published the engineering write-up behind OpenAI's GPT-5.6 Sol Ultrafast preview, up to 14x faster and gated to a small set of API customers.dots studio shipped dots3-note preview — 280 billion parameters, 16 billion active, 512K context, multimodal, Apache 2.0 — the same night Z.ai released GLM 5.3 and Paul Graham noted that tuning open weights has swung back into fashion.Harrison Chase shipped cron scheduling into Managed Deep Agents, Vercel wired nine agent CLIs behind one gateway, and Perplexity turned Sonar into an Agent API.The Rails team published its first agent benchmark: eight models, 21 atomic tasks, three runs each.Two Minute Papers covered an unreleased Claude improving a prime-distribution bound past the human record after roughly 650 prompts — and labeling its own result "too strong to be new."
