TechLook
Anthropic is about to become the most valuable AI company in the world — and the number that explains why isn't a model benchmark. It's revenue. The Claude maker is finalizing a $30 billion raise at a $900 billion valuation,surpassing OpenAI's $852 billion mark for the first time. That valuation only makes sense if you believe the enterprise AI adoption curve is real — and the revenue data suggests it is. Combined with OpenAI's most practical agent move yet, a small security research team breaking Apple's flagship chip defenses with frontier AI, and warehouse robots now running at 2-4x human rates, today's news is pointing hard at a single conclusion: the gap between AI demos and AI in production just got a lot narrower.
Anthropic Closes on $30B at $900B — Enterprise Revenue Tells the Real Story
Anthropic is finalizing a $30 billion raise that values the company at over $900 billion, surpassing OpenAI's $852 billion valuation for the first time. The number gets attention, but the number underneath it is what makes the valuation credible.
Here's everything you need to know:
- Lead investors include Greenoaks, Sequoia, Altimeter, and Dragoneer, each committing $2 billion or more
- Anthropic's annualized revenue has grown from $9 billion to over $44 billion in the span of this cycle — driven almost entirely by enterprise Claude adoption
- Over 1,000 enterprise customers now spend more than $1 million annually with Anthropic
- Eight of the Fortune 10 are now Claude customers
- Claude leads business AI adoption at 34.4% versus OpenAI's 32.3% — the first time any rival has overtaken ChatGPT in the enterprise
This is the revenue validation the AI industry has been waiting for. Not a benchmark, not a capability demo — actual enterprise contracts at scale. The $900 billion number is aggressive, but the trajectory from $9B to $44B annualized in one fundraise cycle is the kind of data point that makes aggressive valuations defensible. For founders and builders, the signal is straightforward: Claude is no longer a challenger. It's the primary vendor in serious enterprise accounts.
ChatGPT Now Reads Your Bank Account — OpenAI's Most Practical Agent Move Yet
OpenAI launched personal finance tools in preview for U.S. ChatGPT Pro users, connecting the assistant to over 12,000 financial institutions via Plaid. Users can now ask about portfolio performance, upcoming payments, subscriptions, and spending changes directly inside ChatGPT — simply by typing @Finances in any conversation.
Here's everything you need to know:
- Connections cover more than 12,000 financial institutions through Plaid integration
- Available on both web and iOS
- OpenAI says more than 200 million users already ask ChatGPT financial questions each month
- The launch video drew 14 million views in 48 hours
- Next logical step: Intuit integration for tax impact and credit approval questions — pulling ChatGPT into actual financial decisions
This is the most practically useful version of "AI agent" OpenAI has shipped. Financial advice only works with real data, not generic guidance. By linking to actual accounts, OpenAI is moving ChatGPT from suggestion engine to something closer to a personal CFO. The Intuit integration is the real signal — once AI can pull your tax history and credit profile, the advice stops being educational and starts being consequential. For builders, the lesson is clear: agents without real data are toys. Agents with real data are infrastructure.
Apple Spent 5 Years and $100M+ Building M5 Defenses. A Small Team Broke Them in a Week.
A small research team at Calif used Anthropic's unreleased Claude Mythos model to bypass Apple's Memory Integrity Enforcement (MIE) on M5 chips — the hardware security system Apple spent years and hundreds of millions developing to protect device memory. The team published visual evidence of the first public memory corruption exploit bypassing MIE, then hand-delivered their report to Apple HQ.
Here's everything you need to know:
- The exploit was the first public demonstration of a memory corruption attack bypassing MIE on M5 chips
- Calif described Apple's security architecture as being "built in a world before Mythos Preview"
- Human expertise was essential — Mythos didn't automate the research — but frontier AI multiplied what a small team could accomplish
- Both teams are now working to patch the vulnerability
- The broader implication: small teams with frontier AI can now pull off what previously required entire organizational resources
The takeaway isn't that Mythos is magic — it's that frontier AI has crossed a threshold where it meaningfully amplifies expert human researchers. Apple built MIE as a defense against the models that existed when it was designed. Claude Mythos is not those models. For security teams, this is a structural change in the threat landscape. For builders, it means the security assumptions baked into your infrastructure may not account for what AI-augmented research can do today.
Figure's Robots Hit 100+ Hours. 1,240 Packages an Hour. Warehouse Work Is Being Redefined.
Figure turned a humanoid robot demo into a live factory stress test. The Figure 03 fleet ran past 64 hours — and has now passed 100 hours — sorting more than 80,000 packages at roughly 1,240 to 1,250 packages per hour with a 2.9-second rhythm. No stage resets. No camera tricks. Just robots working through the human sleep cycle.
Here's everything you need to know:
- A human warehouse picker or sorter typically handles 300 to 600 items per hour
- Figure's fleet is running at 2x to 4x that range, continuously, for days
- The robots don't fatigue, don't need shift changes, and don't have attention decay
- Figure has now crossed the threshold where the economic case for robotic replacement is becoming a procurement decision, not a prediction
The sorting category is now the clearest early target for humanoid robotics at commercial scale. The math is no longer speculative: 1,240 packages per hour at $0 variable cost per unit versus human rates with all the associated costs and constraints. For founders building logistics, fulfillment, or warehouse-adjacent products, this timeline just moved up. For everyone else, it's a reminder that the jobs AI takes aren't just knowledge work — they're the physical, repetitive work that powers the economy.
OpenClaw's $1.3M Monthly Token Bill: What 100 Autonomous Agents Actually Do
OpenClaw founder Peter Steinberger published a bill that looks almost fictional: $1.3 million in 30 days for roughly 100 Codex agents running through OpenClaw to review PRs, reproduce bugs, test fixes, open patches, verify benchmarks, and keep engineering work moving. That's about $13,000 per agent per month.
Here's everything you need to know:
- The work wasn't pure code generation — it was the operating layer around software: bug triage, issue routing, regression checks, community handling, support noise
- The agents absorbed the invisible work that typically keeps human engineers occupied between feature builds
- Critics joked this might be why Steinberger joined OpenAI: access to unlimited tokens
- The OpenClaw team frames it as a proof of concept for what software development looks like when compute cost stops mattering
This is the most concrete public record of what autonomous AI dev teams actually cost at scale — and what they actually do. It's not replacing programmers. It's absorbing the administrative, repetitive, and attention-dispersing work that slows human engineers down. At $13,000 per agent per month, it's expensive today. If costs fall toward $1,300 per month in two years — a reasonable trajectory — the economics change entirely for how software teams are built. Watch this space.
OpenAI Is Building the Super App — Merging ChatGPT, Codex, and Atlas
OpenAI is consolidating ChatGPT, Codex, and the Atlas browser into one unified product. The three products that currently represent separate surfaces — conversation, code execution, and web browsing — are being merged into a single experience.
Here's everything you need to know:
- ChatGPT, Codex, and Atlas are being unified into one product
- OpenAI is moving from a multi-product strategy to a single integrated surface
- The consolidation signals a shift from "best model for each task" to "one interface for everything"
A unified AI interface from OpenAI changes the competitive dynamics for every AI tool that's currently sitting on top of a different model or surface. If OpenAI is building the front door, the value of apps that live behind it depreciates. For builders evaluating their product architecture, this is a relevant data point: the platforms you build on top of may not stay distinct forever.
Mira Murati's Thinking Machines Lab: Real-Time Audio, Video, and Text at Once
Former OpenAI CTO Mira Murati's new company, Thinking Machines Lab, unveiled its first product: real-time interaction models that process audio, video, and text simultaneously. The lab is positioning the work as a new category — interaction models designed for live, multi-modal exchange rather than asynchronous prompting.
Here's everything you need to know:
- Thinking Machines Lab is Murati's new venture after departing OpenAI
- The first product processes audio, video, and text simultaneously in real time
- The focus is on live, interactive use cases — not document-based workflows
Murati is betting that the next frontier is live, multi-modal interaction — not just better static models. If she's right, the competitive landscape shifts again. Real-time audio and video processing at quality opens use cases that text-and-image models can't touch: live customer service, real-time collaboration, interactive tutoring, physical presence replacement. It's early, but the pedigree is worth watching.
⚡ Quick Hits
-
Kling AI: Hit #5 on the App Store free charts following OpenAI's Sora consumer shutdown. AI video apps are moving fast to fill the gap.
-
HUBX AI Video: Hit #6 on the App Store free charts on the same Sora shutdown tailwind. Two apps picking up where OpenAI stepped back.
-
NVIDIA SANA-WM: Open-sourced a world model that turns a single image and camera trajectory into 60-second, 720p video you can control frame-by-frame. The open-source video generation space just got more competitive.
-
Anthropic + Gates Foundation: $200 million partnership to deploy Claude in vaccine screening, disease forecasting, and K-12 tutoring across developing nations. AI health impact at scale.
-
Microsoft AI CEO Mustafa Suleyman: Said AI could automate most computer-based professional work within 12 to 18 months — legal, accounting, marketing, coding. The internal timeline from one of the largest AI buyers in the world.
-
CorelDRAW: Now embeds Stable Diffusion 3.5, Flux, and Nano Banana directly in its vector design environment. AI image generation moving into professional creative tooling.
-
YouTube: Opening its AI likeness-detection tool to creators 18 and older. Deepfake detection becoming a platform-level feature.
-
Vercel Zero: An experimental systems language designed so AI agents can read compiler feedback and repair native programs more easily. Infrastructure for AI writing AI code.
Techlook — AI & tech signal for founders and builders.