Overview
AI products are acquiring the business models and extension surfaces of mature platforms. ChatGPT's ad pilot has found early demand, Codex now packages third-party capabilities, OpenAI has explained how its public behavior specification is used, and fresh research shows that speculative decoding improves when draft models match the workload.
Developments
1. ChatGPT's ad pilot reaches a $100 million annualized run rate
OpenAI told Reuters that its US advertising pilot exceeded a $100 million annualized revenue run rate within six weeks. It had more than 600 advertisers, planned self-service tools in April, and said fewer than 20 percent of eligible users were shown an ad on a typical day.
Annualized run rate extrapolates current sales; it is not $100 million already collected or a forecast guaranteed for the next year. OpenAI also says ads remain separate from answers and conversations are not shared with marketers. Independent trust and relevance measures will matter as the pilot expands.
Sources: Reuters on ChatGPT's advertising pilot
2. Codex plugins package skills, integrations, and MCP servers
OpenAI added a plugin browser to Codex. Plugins can bundle reusable instructions, app integrations, and MCP servers, turning configurations that power users could already assemble into installable packages; initial examples connect services including GitHub, Gmail, Box, Cloudflare, and Vercel.
Convenience changes the security threshold because installation can introduce tools, authentication flows, external data sharing, and hooks. Enterprise distribution requires component and permission review, approved marketplace sources, and tested revocation rather than treating one-click installation as sufficient assurance.
Sources: Ars Technica on Codex plugins · OpenAI's Codex plugin documentation
3. OpenAI explains what its Model Spec can and cannot promise
OpenAI described the Model Spec as a public framework for intended model behavior, organized around an instruction chain of command, hard safety rules, steerable defaults, and worked examples. It also says the document is partly aspirational, typically targeting behavior up to three months ahead of deployed models.
That caveat makes the spec an interface and evaluation target rather than a description of every response available today. Applications can test expected behavior against it while retaining controls for model deviations, product-layer policies, and capabilities outside the document.
Sources: OpenAI's approach to the Model Spec
4. Google releases Gemini 3.1 Flash Live for real-time audio
Google made Gemini 3.1 Flash Live available in preview through the Gemini Live API, in an enterprise customer-experience product, and through Gemini Live and Search Live. Google reports lower latency, improved tonal understanding, and stronger function calling, and says generated audio carries a SynthID watermark.
The published benchmark scores and customer endorsements are Google's evidence, not independent validation. Voice-agent tests should include interruptions, accents, noise, tool failures, language switching, and watermark persistence; average latency and clean-audio accuracy will miss the failures that determine whether a live interaction is usable.
Sources: Google's Gemini 3.1 Flash Live announcement
5. OpenAI opens a bounty for agentic safety failures
OpenAI launched a public Safety Bug Bounty for reproducible abuse risks that fall outside conventional security vulnerabilities. Covered agentic cases include third-party prompt injection, data exfiltration, disallowed actions at scale, and other actions with plausible material harm; ordinary jailbreaks without such impact remain out of scope.
The scope gives researchers a defined reporting channel for failures that policy testing often leaves between safety and security teams. Requiring a prompt-injection exploit to reproduce at least 50% of the time favors actionable reports, although lower-frequency failures can still matter in high-volume products.
Sources: OpenAI Safety Bug Bounty announcement
6. Task-matched draft models improve LLM speculative decoding
A March 27 preprint trained lightweight HASS and EAGLE-2 draft models on math instructions, ShareGPT conversations, or mixtures, then measured how many proposed tokens larger LLMs accepted on MT-Bench and three math benchmarks. Math-trained drafters specialized on reasoning tasks, while ShareGPT-trained versions led on MT-Bench.
Naive checkpoint averaging weakened the specialized models, while confidence-based routing improved over individual drafters and merged-tree verification produced the highest acceptance length for both tested backbones. Acceptance length is only a serving proxy; wall-clock latency, routing overhead, memory use, and the stability of workload classification determine whether specialization pays off in production.
Sources: TAPS preprint, version 1 · TAPS code repository