Jason Says
The biggest signal today is GPT-5.6 topping LiveBench while shipping as the default in Microsoft 365—model competition has evolved from benchmarks to distribution wars; Mercor doubling its valuation in five months also shows AI vertical plays have far higher ceilings than anyone expected.
OpenAI launches GPT-5.6, topping LiveBench benchmarks with better performance per dollar than Opus-class models. It becomes the default model in Microsoft 365 Copilot across Word, Excel, and PowerPoint. Abacus AI predicts 80% of Opus API workloads will migrate to GPT-5.6.
OpenAI's ChatGPT Work is a new agent that takes action across apps and files, sustaining focus on a project for hours. It marks ChatGPT's transformation from chatbot to autonomous productivity agent, competing directly with tools like Cursor and Devin.
Meta enters the AI coding race with Muse Spark 1.1, targeting enterprise-scale agentic workloads, bug fixing, and large code migrations. It competes directly with GitHub Copilot and Cursor, leveraging Meta's open-source model ecosystem and enterprise reach.
The New York Times filed a motion for sanctions, alleging OpenAI concealed tools and datasets that could identify copyrighted journalism in ChatGPT outputs. The escalating lawsuit could set landmark precedents for AI training data legality across the industry.
AI hiring platform Mercor is in talks for a $20B valuation, doubling from $10B reached just five months ago. It automates talent matching and skills assessment via AI, making it one of the fastest-growing companies in the AI-powered HR space.
Modal CTO Akshat Bubna shares lessons from two years building agent cloud infrastructure, arguing that agent workloads differ fundamentally from inference tasks in latency, state management, and concurrency. Existing AI infrastructure is becoming a bottleneck, demanding ground-up rethinking.
AgentLens introduces a new benchmark that evaluates code agents on their full execution trajectory—instruction following, tool use, self-verification, error recovery, and communication—not just pass/fail. For developers building coding agents, this shifts evaluation focus from task success to overall execution quality.
SARA proposes a single-rollout asynchronous RL pipeline that updates LLMs as each rollout completes, solving inefficiency in long-horizon agent training. For developers, this means faster, more stable post-training for agents, lowering the barrier to fine-tuning agent models with RL.
Subscribe for daily AI updates + free playbook
📘 Subscribe Free