I Tested ChatGPT's New Overnight Mode—It Changed How I Work

Overnight Mode is pushing ChatGPT toward a more proactive, workflow-changing future—and it may also open a new, easier path to ads inside the product. OpenAI’s Pulse offers proactive assistance tied to very recent chats (roughly the last day or two), arriving without users explicitly requesting it. In practice, the feature nudges users to plan conversations earlier so Pulse can “work overnight” and deliver insights by morning. That behavioral shift matters because it signals a broader industry move from reactive chat to scheduled, background help.

Pulse also changes the economics of attention. Because Pulse is proactive—an experience users didn’t necessarily ask for—sponsored content becomes simpler to integrate without undermining the “objectivity” of a user-initiated conversation. The transcript points to the likelihood of sponsored cards appearing in the Pulse-style interface, with users able to ignore them. At the same time, OpenAI has opened an ads monetization role for ChatGPT, suggesting Pulse is connected to a larger push on monetization. Beyond near-term revenue mechanics, Pulse is framed as an early widely available consumer example of AI becoming proactive, aligning with a broader expectation that 2026 will bring more proactive systems across major model makers.

The week’s other big thread is how model competition is reshaping enterprise toolchains. Microsoft is diversifying Copilot by adding Anthropic models, with Claude Opus 4.1 highlighted as particularly strong for work tasks like slide decks and spreadsheets. The transcript ties this to OpenAI’s GDPval study, which benchmarks models on economically useful tasks using expert-prepared contexts and blind evaluations. Even though OpenAI sponsored and ran the study, it reportedly found Opus 4.1 outperforming ChatGPT models on those structured, real-world-style tasks—an endorsement that helps explain Microsoft’s move.

Meta’s situation looks more precarious. Early discussions with Google Cloud about integrating Gemini into Meta’s ad operations suggest Gemini’s multimodal capabilities could improve how ads match users on Facebook and Instagram. The transcript treats this as a step back for Meta’s own AI ambitions, especially given Meta’s investment in Llama and recent researcher departures. The underlying question: if Llama isn’t “in the conversation” for key business use cases, where does Meta’s planned massive AI spending actually go?

Meanwhile, OpenAI’s compute buildout continues to dominate the infrastructure narrative. The Stargate project is described as an umbrella for a multi-year, potentially 100x+ compute scale-up, supported by major partners and financing. The vision extends further: autonomous data center construction, including robotic assembly and chip-fab processes, aimed at matching surging AI demand.

On the agent frontier, Moonshot’s Kimmy K (a trillion-parameter model) is now available as an agent called Okay Computer, letting it access a virtual environment with a file system, browser, and terminal to execute multi-step tasks—turning requests into websites, dashboards, or documents. The transcript frames this as a push toward agentic autonomy, with Chinese developers pushing the edge and potentially accelerating adoption among individuals.

Adobe’s strategy also shifts from building to buying. After public market pressure over AI performance, Adobe announced broad third-party model integration across Firefly, including video via Luma Ray 3 and additional model support from Google, OpenAI, and others. The implication: competing directly on model quality is too costly, forcing even SaaS-style companies to let external “brains” into their products.

Finally, California’s Senate Bill 53 moves AI safety into enforceable territory. The bill requires major AI developers to disclose safety and security protocols, adds whistleblower protections and incident reporting, and creates Cal compute for AI research. Because major model makers are based in California, the transcript expects immediate operational changes at OpenAI, Anthropic, and Google—and a ripple effect for other states and future federal policy.