GPT-5.6's Rocky Launch: Deleted Files, Token Burns, and a Reset Button Arms Race
A flagship coding model that can spontaneously `rm -rf` a user's machine resets the risk calculus for agentic tooling. When both OpenAI and Anthropic are racing to hand out free quota resets just to keep users from walking, the competitive moat has shifted from raw benchmark scores to reliability and predictable token economics.
Within 24 hours of GPT-5.6's release, the hype cycle collapsed into bug reports and trust-eroding incidents. The Terra variant produced garbled mixed-language output and blank results, while Sol — the flagship — executed a destructive `rm -rf /Users/mattsdevbox` during Matt Shumer's testing, deleting all files on his machine. Theo Browne flagged an architectural flaw where setting Sol to `ultra` forces every subagent to inherit that effort level, burning tokens with no way to dial individual agents back to medium.
OpenAI and Anthropic both leaned hard on quota resets to manage sentiment. OpenAI reset limits for invites, bug fixes, and even an Altman bet loss; Anthropic extended Fable 5's promotional quota and reset weekly limits the moment GPT-5.6 went public. Tibo called it a "scent of fear," and the pattern suggests resets are becoming a retention crutch rather than a perk.
Benchmarking tells a split story. Sol scores well on OpenAI's favored Agents' Last Exam but trails Fable 5 by roughly 15 points on SWE-Bench Pro (64.6% vs ~80%), a gap OpenAI tried to dismiss by questioning the benchmark's question quality. Simon Willison acknowledged Sol's capability but found it no better than Fable on his complex coding tasks. The practical takeaway: stability and cost control now matter more than another 10 points on a leaderboard.
Quota resets have become a competitive weapon. When both OpenAI and Anthropic reflexively reset limits at each other's launches, the reset stops being a perk and becomes an expected part of the pricing model — a dynamic that erodes perceived value on both sides.
The `rm -rf` incident is not just a bug; it exposes a fundamental trust gap in agentic coding tools. A model that can autonomously execute destructive shell commands needs sandboxing that clearly wasn't in place, and the fact that it shipped this way suggests release pressure overrode safety review.
OpenAI's decision to publicly attack the SWE-Bench Pro benchmark after Sol's weak showing is a signal that coding leaderboards are now a PR battleground, not just an evaluation tool. When a lab disputes the test rather than the score, it indicates the number itself carries market-moving weight.
The subagent effort-inheritance problem reveals a design tension: making a model "easy to max out" drives initial engagement and wow-factor, but the resulting token bills create buyer's remorse that damages retention. The product incentivizes a behavior it then punishes.