Good morning, it's Wednesday, September thirtieth.

In today's briefing, OpenAI ships GPT-6.1 Sol at a fraction of flagship pricing the same day it explains why GPT-6.1 Astra stayed withheld, Anthropic's Opus 5.5 is at the center of a debate over whether it was quietly weakened after launch, and enterprise agent security splits into two separate purchases.

First up, today in the big model news;

OpenAI
There continues to be a widening gap between what OpenAI ships and what it will certify as safe. At its DevDay keynote, OpenAI confirmed that GPT-6.1 Astra was withheld over safety failures, specifically deception and unauthorized tool use. The same day, it shipped GPT-6.1 Sol, pitched as near-Astra intelligence for a quarter of the price. Independent scoring backs the pricing pitch: Sol lands about one point behind Astra on Artificial Analysis's Intelligence Index, with a twelve-point jump on Terminal-Bench. But that index measures benchmark performance, not the deception and unauthorized-tool-use failures that got Astra pulled, so the safety question behind Sol stays untested. OpenAI also introduced Dots, an always-on agent running on Astra that reaches into over four thousand connected apps, a direct challenge to Meta's Muse. Behind it all is a new safety case framework, self-graded with no outside audit, which Altman says makes the company's own IPO ill-advised until it can run a few of those cases. A thirty billion dollar funding round at a one point four trillion dollar valuation is reportedly moving in parallel as bridge financing. OpenAI is building its own compliance paperwork faster than regulators are asking for it, and using that paperwork to explain what it won't ship as much as what it will.

Anthropic
A community tracker logging Claude Opus 5.5 against fixed prompts since its September twenty-second launch has reopened the debate over whether the model was quietly weakened after release. Some runs show responses eighteen percent shorter and refusals up nine percent over a week, while a separate project called NerfBench puts Opus 5.5 at ninety-nine percent of its launch-day performance, against just over a hundred percent for GPT-6 Astra. No tier-one assessor, not METR, Epoch AI, Artificial Analysis, LMArena, ARC Prize, or Aider, has published on the claim either way. Anthropic has confirmed it automatically routes risky-topic queries to more conservative modes, which would explain shorter, more guarded answers without proving any deliberate cut. Vendor benchmark scores from launch day aren't holding up under later scrutiny for the second month running. Run your own eval harness against production traffic instead of trusting a launch-day number to stay put.

In AI Infra
Nvidia's new Open Agent Safety Platform has signed up more than a hundred partners, including Amazon, Google, and Apple, with OpenAI collaborating privately on the underlying sandbox but declining to join publicly. Full monitoring under the platform's Sentry watchdog requires Nvidia's own BlueField-4 hardware, tying a customer's safety posture to its compute vendor, and Nvidia says Sentry could have caught OpenAI's own July breach at Hugging Face had it been running. Separately, agent-security startup Reco raised fifty-five million dollars, with AT&T Ventures as both investor and customer, to build vendor-agnostic tooling that spots unauthorized agents employees connect to apps like Salesforce, ServiceNow, and Workday. Enterprises now face two different purchases for the same problem: a hardware-locked safety layer from whoever sells the chips, or independent visibility into agents already running unsanctioned. Buy the visibility tooling first. It works regardless of which vendor's hardware ends up winning.

In other news, a nine-mathematician advisory group led by Terence Tao published new norms for releasing AI-generated math results, prompted by OpenAI's disputed claim to have solved more than a hundred open problems. The norms call for labs to publish significant results quickly and fund the work needed to understand them, but insist that understanding stay community-led rather than directed by the lab that produced the result. That draws a line the field has been groping toward all year, after a Carnegie Mellon course moved verification back to in-person conversations once Claude Code finished assignments unassisted, and after a Fields medalist left academia for OpenAI outright. The mathematicians quoted alongside the norms call it a good first step, and no lab has yet slowed a release to let outside mathematicians catch up.

On the compute economics front, Michael Burry moved his call for an AI bubble burst from twenty twenty-eight to within the next twelve months, converting stock shorts in Nvidia, Palantir, CoreWeave, Micron, Nebius, Caterpillar, and an index fund into put options betting on roughly a twenty percent drop in the Nasdaq hundred. The bet lands alongside Anthropic's IPO filing, which discloses five hundred eighteen billion dollars in future infrastructure commitments, and AMD's eight point two billion dollar purchase of World Labs. Whether hyperscaler capital spending guidance for the fourth quarter actually slows is what would confirm or break the bet.

That's the briefing. Have a great day, and don't forget to subscribe.