Leadership & Executive
The Board Room
GPT-6 Astra lists at 2.5x its predecessor, to the penny of Anthropic's Fable 5.1.
Deflation still holds, but only one generation back: Microsoft cut transcription 72% in the same week the frontier moved the other direction. The more consequential line is OpenAI's shift off tokens toward per-task pricing, which means the AI margin plan your finance team approved this year is denominated in a unit being retired mid-increase.
In Play
Frontier Prices Rose While Commodity Inference Hit the Floor
GPT-6 Astra ships at $10 per million input tokens and $50 output — 2.5x the model it replaces and exactly Anthropic's Fable 5.1 price. Meanwhile Microsoft's transcription tier fell 72% to $0.10 per hour. Cost deflation now applies only to capability that is one generation old. The first deep dive works through the pricing split and the placement response.
Ask ClarityMulti-Vendor AI Is Not Diversification
ChatGPT, Claude and Grok went dark together for hours on Thursday, and Morning Brew's account also includes Gemini. Separately, Ramp's spend data puts 1% of OpenAI's and Anthropic's customers behind 80% of enterprise revenue. Vendor plurality is procurement diversity, not architectural redundancy. The correlated-failure and counterparty detail sits in the third deep dive.
Ask ClarityThe Independent AI Scoreboard Goes Dark Until 2027
GPT-6 Astra effectively saturated FrontierMath and ARC-AGI 3, and the successor benchmark ARC-AGI 4 does not ship until Q1 2027, per The Algorithmic Bridge. For roughly two quarters, every superlative — including Microsoft's claim that its transcription model runs 10x faster than OpenAI's — is vendor-reported. Procurement evidence now has to come from a harness you own; the second deep dive covers the verification consequences.
Ask ClarityExploit Windows Closed to a Single Day
The median gap between vulnerability disclosure and exploitation is now one day, with the zero-day rate just under 87% and up roughly 60% in the past year, per a16z's data. The share of vulnerabilities still unexploited reaches 0% at six weeks; in 2022, half were still unused at three months. Attackers separately ran a full ransomware chain in under 10 hours using AI agents, chaining credentials from development environments into cloud. Budget that buys faster patch triage no longer buys protection; blast-radius limits do.
Ask ClarityCapital Is Cheap for Revenue, Expensive for Narrative
Cognition drew roughly $10B of interest at about $47B on more than $900M of annualized revenue, up 80% in three months. Thinking Machines is negotiating a reported $40B target on a few hundred million in annualized revenue, 20% below the $50B that failed to close last fall, with no term sheet reported. Both are reported round targets rather than settled marks. Upstream, debt is doing the work: ByteDance upsized its largest-ever loan from $20B to $30B.
Ask Clarity
Deep Dives
- ●
The Frontier Charges More. Everything Below It Is Nearly Free.
Two suppliers quoting one number is not a market, and the unit you negotiate on is being retired mid-increase, which leaves workload placement as the only cost lever you still control.
The unit of account is being retired mid-increase The frame most procurement teams are using is that a price went up. On the same call where the new price landed, OpenAI president Greg Brockman said "pricing tokens doesn't make any…
3 action items
- ●
Verification Is the Only Thing Nobody Can Sell You
The shared scoreboard, the vendor's own reasoning trace and the human reviewer inside your workflow all stopped working in the same week; regulated buyers will soon ask which of your outputs is provable.
A vendor documented that its own model is harder to audit OpenAI's evaluations found Astra's written reasoning harder to monitor than the model it replaces. Chief scientist Jakub Pachocki declined to say whether a new training technique explains that, then…
3 action items
- ●
Your Multi-Model Strategy Failed Two Tests in One Week
Vendor plurality hedged neither uptime nor demand, and the same narrow buyer base underwrites every supplier on your shortlist, including the compute firms standing behind them.
Concurrent failure points at something shared underneath Products operated by four different companies, on four different balance sheets, went dark inside the same window, and none has published a cause. Morning Brew flagged concurrent multi-provider failure as unusual, and that…
3 action items
The edition continues
Take the signal into the room.
Sign up or log in to read all 3 deep dives in full, plus the final take.
Read the full editionContinue with LinkedIn