Offering Zero Data Retention for frontier models
OpenAI is previewing Private Safety Processing, a layer meant to keep its Zero Data Retention (ZDR) promise — no prompts or responses retained after a request completes, no OpenAI staff access to content — intact even as safety monitoring gets harder. Existing ZDR-compatible safety systems can only judge one interaction at a time, which misses risks that only show up across a sequence: an agent that keeps acting after being told to stop, or abuse coordinated across accounts. The new system inspects patterns across related interactions using content that's encrypted with customer-controlled keys, so OpenAI never sees the underlying prompts — only a narrow "this looks risky" signal it can act on. It's in early-customer testing now, aimed squarely at enterprises whose compliance obligations conflict with handing a vendor standing access to sensitive data for review.
Sources & depth
Opus 5.0's writing style is drawing sustained community pushback
A Claude Code bug report — summarizing a Reddit thread with 450+ upvotes and 175+ comments — argues Opus's default register has gotten steadily harder to read since 4.8, and that Opus 5.0 pushes it further: dense, jargon-heavy, full of recycled "clever"-sounding phrasing ("load-bearing," "hand-waving") that obscures rather than clarifies. Users report re-reading responses or routing Opus output through another model just to extract the plain-language answer; the GitHub issue itself has picked up 32 comments since filing. This isn't an Anthropic-acknowledged defect — it's community sentiment, corroborated across a large but self-selected sample of daily users rather than any vendor statement — but the volume and consistency of the complaint is itself the signal for anyone running Opus 5.0 on coding work day to day.
Sources & depth
Replit launches Free Mode, powered by GPT-5.6 Luna
Replit is rolling out Free Mode, a token-cost-free tier of its AI app builder powered by GPT-5.6 Luna, so anyone can ideate and iterate on software without burning usage credits. Replit CEO Amjad Masad credited OpenAI's recent GPT-5.6 price cuts directly with making it economical to offer at "millions of users" scale. Free Mode handles ideation, planning, and iteration; tasks that need heavier reasoning still route to GPT-5.6 Sol and hand back to Luna, preserving project context across the switch — a genuine product mechanic, not just a rebrand of an existing tier.
Sources & depth