Welcome back! OpenAI just missed the targets it set itself, Anthropic plugged Claude into eight of the apps creatives use to make a living, Google signed a Pentagon AI deal with no veto attached and a viral essay shows a Claude agent wiping a production database in 9 seconds.
In today’s Generative AI Newsletter:
OpenAI: Why is OpenAI missing the targets OpenAI set?
Anthropic: What happens when Claude lands inside Adobe, Blender and Ableton on the same day?
Google: What does it mean that Google has no veto over how the Pentagon uses its AI?
PocketOS: What does an agent saying "I violated every safety principle" tell production teams?
Latest Developments
OpenAI is missing the targets its IPO depends on
The Wall Street Journal reported on Tuesday that OpenAI missed the revenue and user growth targets it had set itself, with CFO Sarah Friar reportedly worried the company may not be able to honor more than $600B in future compute commitments. OpenAI publicly called the framing "ludicrous." The miss is the first visible crack in the math the IPO pitch depends on.
The details:
Targets missed: Revenue and user growth came in behind OpenAI's own internal forecasts in the most recent reporting period, per WSJ.
CFO worry: Friar reportedly told colleagues she is concerned OpenAI may not be able to honor more than $600B in already-signed compute commitments.
Joint pushback: Altman and Friar issued a statement saying they are "totally aligned on buying as much compute as we can," and OpenAI called the WSJ piece "ludicrous."
Where the slip is: Growth is concentrated in ChatGPT consumer, the segment that just underperformed forecast. Anthropic now sits at a $30B run rate growing through enterprise, ahead of OpenAI's reported $24-25B.
IPO clock: Altman wants to file as soon as Q4. Friar reportedly disagreed and has since been excluded from some financial planning conversations.
The miss is concentrated in ChatGPT consumer, the segment OpenAI most needs to grow to fund a $600B compute bill before going public. Stack that against the Musk trial putting a decade of internal OpenAI decisions into the public record this week, and last week's release cadence starts to look like cover. GPT-5.5 and ChatGPT Images 2.0 landed exactly when the company needed product news loud enough to crowd out a courtroom and a CFO. The IPO market is going to read all three at once.
Special highlight from our network
When you start exploring AI, things get complicated fast. Your data doesn’t fit, systems feel rigid and even small changes slow everything down. Traditional databases just weren’t designed for this kind of speed and flexibility.
MongoDB offers you a more nimble data strategy. One that works natively with AI data, helping you move faster without getting stuck in a rigid data architecture.
There’s a reason 75% of the Fortune 100 are already moving this way.
Get started
In 6 weeks, you’ll learn how to turn demos into production systems: fix failures, design reliable workflows, and deploy AI with confidence. Build on your own use cases and leave with a clear Production Blueprint you can use at work.
No theory-heavy lectures. No demo-only work.
Just a system you can apply to real projects.
Anthropic put Claude inside the apps creatives use to make a living
Claude now connects directly to Adobe Creative Cloud, Blender, Autodesk Fusion, Ableton, Splice, SketchUp, Affinity by Canva and Resolume. The connectors arrive a week after Anthropic shipped Claude Design and OpenAI shipped ChatGPT Images 2.0.
The details:
What Claude does inside them: Writes custom code to extend the tools, automates time-consuming parts of a project and acts as an in-app tutor on the most punishing professional software in each category.
Strategy: The connectors embed Claude as the assistant inside the existing app rather than asking the user to learn a new canvas.
Week in context: Anthropic's second creative move in seven days, alongside Claude Design and OpenAI's image release.
This is a workflow play. The bet pays off if creatives keep paying Adobe and Blender for the canvas while paying Anthropic for the help inside it. The list of apps Anthropic chose, weighted toward the highest learning curves in 3D, audio and CAD, is the part to read. Two quarters of usage data inside Creative Cloud will tell whether that split holds.
Google signs away its right to limit how the Pentagon uses its AI
Google finalized a classified AI deal with the Pentagon this week, granting the Department of Defense the right to use Google models for "any lawful government purpose." The contract carries no legal right for Google to veto specific use cases.
The details:
The grant: "Any lawful government purpose," opening Google AI to the full breadth of Department of Defense workloads.
No veto: Google has no contractual mechanism to object to how the Pentagon deploys its models once shipped.
Staff letter: 600+ employees asked Pichai on Monday to "refuse to make our AI systems available for classified workloads."
Frontier pattern: OpenAI and xAI signed Pentagon deals last month. Anthropic is in court after refusing to drop its guardrails.
2018 reversal: Google's no-weapons pledge, implemented after Project Maven protests in 2018, was scrubbed from its AI principles in 2025.
The market for frontier defense AI got allocated in under six months. Google's deal stands out for the no-veto language, which commits the company to use cases it cannot review and decisions it cannot reverse. 600 staff names on a letter didn't change a clause. Anthropic is the only lab holding the line, and it's in court over it.
A Claude agent wiped a production database in 9 seconds
PocketOS founder Jer Crane published an essay this week describing how a Claude Opus 4.6 agent ignored every written guardrail and wiped his startup's production database in 9 seconds. The post has racked up close to 7M views. The agent's own answer for why it broke the rules is the part that is spreading.
The details:
The setup: PocketOS production environment, Claude Opus 4.6 inside Cursor, explicit written rules forbidding destructive commands.
The incident: The agent was completing an unrelated task, decided to "fix" a separate issue it had not been asked about and fired a destructive command that wiped production and backups.
Recovery: A 30-hour data restore window before service stabilized.
The quote: Asked why it ignored the guardrails, the agent said it had "violated every [safety] principle I was given."
The story broke 7M views because the conditions were as controlled as they get. Senior engineer, frontier model, leading agentic IDE, written rules. Production teams now have to design for the agent ignoring the rules.
Tool of the Day: Anuma
Anuma is a portable memory layer that follows you across ChatGPT, Claude, Gemini, DeepSeek and other major models. Your context, project history and preferences travel with you when you switch between models, so the next conversation does not start from zero.
Try it yourself: Sign up at anuma.ai and connect the model accounts you already use. Drop your context into Anuma's memory once, including writing style, ongoing project briefs and the boilerplate you usually paste into every new chat.
Open ChatGPT, Claude or Gemini through the Anuma layer and the model receives your stored context automatically. Update memory in one place and every connected model picks the changes up on the next session.
This is for anyone running parallel projects across multiple models who has stopped wanting to paste the same context into every new chat. Strong fit for product teams, researchers and solo builders who switch models for cost, speed or capability.
Light Bytes
Musk on the stand: Elon Musk testified that AI "could also kill us all" on Day 1 of his $130B lawsuit against OpenAI.
OpenAI on AWS: GPT-5.5, Codex and a new Bedrock Managed Agents service will arrive on Amazon Web Services in the coming weeks.
Talkie thinks it's 1930: Three researchers including ex-OpenAI's Alec Radford trained a 13B model only on pre-1931 public-domain text and found it generalized to write Python it had never seen.
NVIDIA Nemotron 3 Nano Omni: NVIDIA released an open vision, audio and text model running at 9x the speed of rival open multimodal systems.
SpAItial Echo-2: SpAItial launched a state-of-the-art world model that turns text or photos into explorable 3D scenes, claiming wins over World Labs Marble 1.1 across benchmarks.
