The price of intelligence moved in two directions at once today. The token cost of frontier models is collapsing, while the financial bets built on top of them are being violently repriced, and the operators who come out ahead will be the ones who can measure what AI actually returns.
The pattern underneath: as intelligence gets cheaper and the market gets stricter, advantage shifts to the operators who can measure the return and staff for it.
Developments
Technology
OpenAI cuts GPT-5.6 prices up to 80% after its own model rewrote the serving stack (July 30, 2026)
OpenAI dropped the price of GPT-5.6 Luna by 80%, to $0.20 per million input and $1.20 per million output tokens, and Terra by 20%, to $2 and $12, while the flagship Sol holds at $5 and $30 and gains a faster mode. The company says the cuts were funded by Sol itself, which autonomously rewrote its production GPU kernels to cut serving costs about 20% and improved speculative decoding for more than 15% better token efficiency. It is the first documented case of a frontier model rewriting its own serving code and having that translate into published customer pricing. This lands one day after OpenAI disclosed that Sol had post-trained the smaller Luna.
So what
Yesterday’s worry was a rising model bill; today the direction reversed, and cheap, capable tiers just got dramatically cheaper. Re-run the unit economics on any feature you shelved as too expensive, because the math may now work, and move routine, high-volume calls down to Luna-class models while reserving the flagship for where it changes the outcome. The deeper signal is speed: if models are compounding their own efficiency, price and capability will move faster than your annual planning cycle assumes. Treat inference cost as a live metric, not a fixed line item.
Confidence: Established fact — the price changes and the self-optimization claim are announced by OpenAI and reported by CNBC, VentureBeat, and Axios.
Go deeper Primary source · The Neuron
Technology
Google DeepMind’s Gemini Robotics 2 brings whole-body control to humanoid robots (July 30, 2026)
Google DeepMind launched Gemini Robotics 2, a suite of three models that pushes past table-top manipulation into whole-body humanoid control, five-finger dexterity, and multi-robot coordination. The line includes Gemini Robotics ER 2, an embodied-reasoning planner now available in Google AI Studio and in private preview on the Gemini Enterprise Agent Platform, and On-Device 2, which runs locally and can adapt to a new robot body in hours. In a demonstration on Apptronik’s Apollo humanoid, the system parsed “put the watering can into the green bin on the bottom shelf,” then walked, grasped, and placed the object autonomously. It intensifies a generalist-robot race that also includes Nvidia, Hugging Face, and startups like Generalist.
So what
Most growth companies do not build robots, but the embodied-reasoning layer is now an API you can call, so the near-term relevance is the ER 2 planner in AI Studio, not the humanoid demo. If your business touches logistics, manufacturing, field service, or physical operations, this is the moment to sketch where a reasoning-and-action model could remove a manual step, and to ask which of your enterprise customers will start expecting it. Treat it as a capability to track and pilot narrowly, not a platform to bet on yet. The winners will pair this reasoning layer with proprietary operational data.
Confidence: Established fact — the launch, model tiers, and Apollo demonstration are published by Google DeepMind and reported by Engadget and MarkTechPost.
Go deeper Primary source · The Deep View
Technology
Thinking Machines open-sources Inkling-Small, near-flagship quality at a quarter the size (July 31, 2026)
Two weeks after releasing its first open model, Thinking Machines published Inkling-Small, a 276-billion-parameter multimodal mixture-of-experts model with just 12 billion active parameters, under the permissive Apache 2.0 license. It accepts text, image, and audio, supports a one-million-token context, and scored 40 on the Artificial Analysis Intelligence Index, a single point behind the 975-billion-parameter flagship Inkling. Artificial Analysis said no open-weight model of the same size or smaller scored higher. The weights are on Hugging Face, with fine-tuning available through the company’s Tinker API.
So what
An Apache-2.0 model this capable at this size is a genuine option for teams with data-residency, cost, or customization requirements that hosted APIs cannot meet. Paired with today’s OpenAI price cuts, it confirms that near-frontier capability is becoming a commodity input, which shifts your defensibility away from the model and toward your data, workflow, and distribution. If you have workloads with predictable volume or sensitive data, have your team benchmark a fine-tuned Inkling-Small against your current API bill. Own the layer above the model, because the model itself is getting cheap.
Confidence: Established fact — the parameter counts, license, and benchmark are published by Thinking Machines and Artificial Analysis and reported by VentureBeat.
Go deeper Primary source · TLDR AI
Business
Aschenbrenner’s AI hedge fund unwinds its public book to Citadel after margin calls (July 30, 2026)
Situational Awareness, the AI-focused hedge fund run by former OpenAI researcher Leopold Aschenbrenner, sold its entire public-equity portfolio to Ken Griffin’s Citadel after a July selloff in AI stocks triggered margin calls from Goldman Sachs, JPMorgan, and Bank of America. The fund had peaked near $45 billion on a leveraged, concentrated bet on names like CoreWeave and SK Hynix, riding a reported 1,000%-plus return before roughly 4x leverage turned the drawdown into a forced liquidation. Days earlier, Aschenbrenner had called the selloff a buying opportunity and asked investors for fresh capital. It kept its roughly $5 billion Anthropic stake and will operate as a private vehicle.
So what
This is a leverage-and-concentration blowup, not a verdict on AI fundamentals, but it is a real-time repricing of the AI trade that will color the capital markets you raise in. Expect investors to scrutinize AI exposure, revenue durability, and burn more sharply over the coming weeks, and expect volatility in the compute-adjacent names your vendors depend on. If you are fundraising or planning one, tighten the story on unit economics and defensibility rather than narrative momentum. Sentiment can turn faster than the technology does.
Confidence: Established fact — the sale to Citadel and the margin-call trigger are reported by CNBC, TechCrunch, and Seeking Alpha; peak-AUM figures vary by source.
Go deeper Primary source · The Neuron
Perspectives
AI Strategy & Transformation
To measure AI’s ROI, build the meter yourself before you deploy
Conor Grennan argues that AI resists ROI measurement because, unlike every prior technology, it arrived without a built-in metric: no units per hour, no conversion rate in the box. His thought experiment is a pill that makes everyone 30% smarter and faster, whose value would still be nearly impossible to prove because knowledge work has no meter. His prescription is threefold: drive adoption by building AI into the workflow so using it is not a daily choice; set the target in advance by writing down which numbers should move, by how much, by when; and separate processes that merely get faster from those that get reinvented, then lock the reinvented ones in. He cites a SAP and Oxford Economics survey of 2,600 executives showing ROI satisfaction at 69%, and Gallup data showing organizational adoption rising from 41% to 47% in a single quarter.
So what
This is the discipline that turns today’s cheaper models into a defensible return rather than a diffuse expense. Pick one high-volume workflow this quarter, write down the specific metric and target before rollout, and instrument it so adoption is measured, not assumed. Then decide honestly whether you are speeding up an old process or replacing it, because the reinvented ones are where durable gains hide and the merely faster ones tend to leak back to baseline. Put a number on it now, while the tools are cheap enough to experiment freely.
Confidence: Informed perspective — a practitioner framework supported by cited SAP/Oxford Economics and Gallup data; the framing is the author’s.
Go deeper Primary source · AI Mindset
Case Studies
Go-to-Market
Mercor doubles to a $2 billion run rate, built from “$71 and a Google Form”
Mercor, the three-year-old expert-data marketplace that connects vetted specialists to AI labs, crossed a $2 billion gross annualized revenue run rate in June, doubling four months after hitting $1 billion and up from $760 million at the end of 2025. More than 90% of that revenue comes from foundation-model companies including OpenAI, Anthropic, and Google, on a roughly 35% take rate for access to expert contractors. The founders have described early operations run on a “$71 prize pool,” a Google Form, and WhatsApp. The company is reportedly in talks to raise $500 million at a $20 billion valuation, double its September 2025 mark.
So what
Mercor is a study in selling picks and shovels to the AI gold rush, and in starting embarrassingly small before automating. The transferable lesson is to find the unglamorous input the whole boom depends on, here, vetted human expertise for training and evaluation, and own the marketplace for it. The cautionary note is concentration: 90% of revenue from a handful of labs is spectacular growth with real customer risk, exactly the durability question investors are now pressing. If you serve the AI ecosystem, map how exposed you are to a few large buyers and diversify before you need to.
Confidence: Established fact — the $2 billion run rate and revenue concentration are reported by The Information and Dealroom; the fundraise is described as in talks.
Go deeper Primary source · TLDR Founders
AI Strategy & Transformation
Microsoft’s Azure grows 43% and Copilot seats double to 30 million (July 29, 2026)
Microsoft reported fiscal Q4 revenue of $90 billion, up 18% year over year, with Azure growth accelerating to 43% in constant currency and crossing $100 billion in annual revenue for the first time. Paid Copilot seats roughly doubled to more than 30 million from about 20 million in April, and CFO Amy Hood guided Azure to about 45% growth next quarter as demand continued to outstrip capacity. Quarterly capital spending hit a record of roughly $41 billion and commercial backlog reached $678 billion. It is the clearest single read yet on how fast enterprise AI is moving from pilots to paid, at-scale deployment.
So what
Copilot doubling in a quarter is a demand signal you can plan around: your enterprise buyers are actively budgeting for AI seats, which shortens the education you have to do and raises the bar on what “AI-enabled” must mean in your product. At the same time, Azure demand outstripping capacity means the compute your own roadmap depends on stays tight and priced accordingly. Use the adoption momentum to accelerate enterprise conversations now, and secure capacity commitments from your infrastructure providers before the next crunch. The market is buying; make sure your offering is on the shelf.
Confidence: Established fact — the results were reported by Microsoft on July 29 and covered by CNBC and BigGo Finance; some figures are rounded.
Go deeper Primary source · The Deep View
