Anthropic Targets Figma, and the G7 Summit Remains Aimless on AI

Anthropic: Claude Design Aims to Own the Figma Workflow

When Anthropic quietly released Claude Design in April as a “research preview,” it attracted over a million users in its first week. This came with a problem: a PCWorld tester used up 80 percent of their weekly Claude Pro quota in about 25 minutes for just three variants of a single website prototype. Two months later, the company is delivering a revised version that reduces consumption and transforms the tool from a pretty demo into a service layer for brand compliance. The core is the new design system import: companies pull their own components, typography, and color tokens from a GitHub repository into Claude, which then builds with them, checks against the rulebook, and corrects before the user sees the result. An admin role can approve a standard system and lock changes. Additionally, there's a bidirectional connection to Claude Code via /design-sync and /design, intended to bridge the gap between design and engineering. This is all part of ten weeks of aggressive product expansion: Opus 4.8, the DXC alliance for banks and airlines, Claude for Small Business with QuickBooks and PayPal, plus a study of 400,000 Claude Code sessions suggesting that domain knowledge is more important than coding skill. → venturebeat.com

Synthszr Take: The real leverage isn't in the drag-and-resize editor, but in the lockdown. A company with a 200-page brand manual gains nothing from an AI that builds pretty slides according to its own taste. As soon as Claude knows the real buttons, spacing rules, and tokens and validates every output against them, Anthropic answers the first question of every enterprise buyer: Can we control what comes out? The round-trip integration with Claude Code is the second blow, because the same system designs and codes; the old “the mockup didn't look like that” loop between design and engineering disappears when no one has to interpret someone else's intentions anymore. In April (see “Claude Design attacks Figma”), this was a frontal assault on the tool; now, it's an attack on the entire workflow behind it. The fact that Anthropic is simultaneously defusing the token problem shows compute discipline: a tool you can't afford after 25 minutes doesn't scale to the procurement department. Anyone with a cleanly documented design system today can test the import tomorrow morning, no strategy offsite needed.

G7 Becomes an AI Summit with No Results

Just days after the US government banned foreign citizens from accessing Anthropic's most powerful models, Fable and Mythos, the heads of leading labs met with the leaders of Western democracies at the G7 summit in Évian-les-Bains, France. At a working lunch, Anthropic CEO Dario Amodei advocated for sharing the benefits of AI among democratic nations and coordinating trade in powerful models, excluding China. The export controls were triggered by concerns that Fable and Mythos are so good at finding software vulnerabilities that attackers could misuse them; Anthropic had to shut down access for all users, including its own employees without a US passport. Macron called the decision “in some respects strictly nationalistic,” while Canada's Premier Carney warned against “over-dependence on certain models.” Trump, flanked at dinner by Sam Altman and Demis Hassabis, did not address the ban and instead emphasized the US industry's lead over China and its power plant construction for the AI boom. Hassabis called for an international standards body to regulate AI development safely while fostering innovation. → www.washingtonpost.com

Synthszr Take: We described the Anthropic earthquake as a geopolitical fracture back on June 14, and Évian is now showing the fallout. When Washington can shut down a functioning product overnight without public justification, AI capacity is no longer a commodity you can procure predictably, but a lever in someone else's hands. Carney said the only useful sentence of the summit: No one did anything wrong, but it would be wrong not to learn the lesson and build nothing. For Europe, this means: its own models, its own data centers, its own diffusion power, and as an investment, not as subsidized folklore. Estonia digitized 99% of its administration because someone started, not because a summit decreed it. Dependency cannot be negotiated away in Évian; it can only be reduced by building your own. Anyone waiting for the next standards committee resolution has already relinquished control over the most important infrastructure of the decade.

Microsoft Considers China's DeepSeek for Copilot

According to Axios, Microsoft is exploring using a self-hosted, fine-tuned version of DeepSeek V4 (or another open-source model) under Copilot Cowork, the agentic assistant in Microsoft 365. The reason is money. In parallel, Microsoft is switching Copilot Cowork to usage-based billing, charging for the actual compute consumed instead of a flat fee. Charles Lamanna, EVP for Copilot, says it openly: There are users who process hundreds of tasks per week, very productively, but the costs are skyrocketing. Currently, Cowork runs on Anthropic and OpenAI models, both of which have raised their prices and withdrawn their flat-rate plans. The timing is politically sensitive: Washington has floated a ban on DeepSeek, and Anthropic was just forced to cut off top models for non-US users. Microsoft emphasizes that any DeepSeek option would be optional and fully hosted on Azure, with fine-tuning and guardrails. → Techpresso

Synthszr Take: When even Microsoft can't stomach the bill for its own agents, it's the most honest statement about the economics of the agent economy we've gotten this year. An agent doesn't just call the model once; it calls it dozens of times per task, and that's precisely where any flat rate turns into a loss. Back in late May, we wrote about exploding token costs and how China is filling the gap; now the cheapest candidate is on Redmond's shortlist, and its name is DeepSeek. What's remarkable isn't the geopolitical piquancy, but that Microsoft is buying its way out of tight OpenAI exclusivity and mixing models under its own roof like a buyer who never wants to be dependent on a single supplier again. To run agents productively, you need cost telemetry at the use-case level and a model-swapping lever, not a black-box cloud bill at the end of the month. This can be set up today, not after the CFO slaps the first six-figure token invoice on the table. Compute discipline is currently evolving from a nice-to-have to a core competency.

Adyen and Stripe Fight for Checkout While OpenAI Fails

Two of the world's largest payment processors have increased their bets on agentic commerce within days of each other. Adyen has launched a complete API suite for transactions by AI agents, while Stripe is expanding its own infrastructure via Amazon's AWS cloud. Both moves came three months after OpenAI's Instant Checkout collapsed: Walmart saw conversion rates three times worse, and fewer than 30 Shopify merchants ever went live. A few weeks prior, Stripe joined Google's Universal Commerce Protocol, a quiet admission that the standard built with OpenAI had lost the race. The most revealing detail: none of the providers can show real transaction volume from the channel they are building for. → Linas from Linas's Newsletter

Synthszr Take: The entire industry is building a highway where no cars are driving yet. OpenAI's Instant Checkout showed what happens when you automate the funnel before users even want it: three times worse conversion and only 30 merchants participating (out of millions). Adyen and Stripe are smart enough not to bet on a proprietary standard, but on the delivery logic behind it—the API layer that any agent can connect to. That's the right bet, because the tool layer will become a commodity, while the pipeline logic remains the moat. What makes me skeptical: no one can show transaction volume, and a payment system without payments is a PowerPoint illusion with quarterly figures. Anyone building here should invest money in the infrastructure while also being brutally honest in measuring when real people start delegating real purchases to agents. The speed of building is impressive; the proof of demand is still missing.

CFOs are Panicking About Tokens and Bringing Out the Lawnmowers

At Databricks' Data + AI Summit, enterprise managers told the reporter from The Deep View the same story: inference costs have run so far over budget that it's becoming a crisis. “Last year, boards let every flower bloom; now they're coming with the lawnmower,” was one quote. Databricks CEO Ali Ghodsi called the situation “completely unsustainable” and said his customers' number one question is: How do we cut costs without killing AI investments? Two answers are emerging: model selection (not every simple query needs a frontier model) and hybrid compute (open models locally on your own hardware). Databricks promptly delivers the right tool, the Unity AI Gateway, which makes token consumption visible and routes requests to the appropriate model. For OpenAI and especially Anthropic, whose revenues have exploded due to employees' unlimited token access, this is bad news. Databricks} → The Deep View

Synthszr Take: A year ago, CEOs were begging their people to finally use AI; now, coding agents are eating up the budget. A market could hardly show it's maturing more beautifully. In March, we wrote about “tokenmaxxing,” when tech companies were still competing for the highest token usage; in May, the budget shock came, and China delivered models with up to a 99 percent discount. Now, architectural discipline is paying off: those who keep the model layer interchangeable (via adapters) and route requests cleanly, instead of using a chainsaw for every daisy, save money immediately without burying their AI strategy. Jevons paradox is in full effect here: the more efficient and cheaper the token becomes, the more we consume, which is precisely why compute discipline isn't a cost-cutting measure but a prerequisite for scaling. The list of providers wanting to help tame agents is getting long, and Databricks is just one of them. Observability of what agents do and cost should be part of every pilot architecture from now on. This can be decided tomorrow morning, not after the next strategy offsite.

Vercel Deleted 80% of an Agent's Tools, and the Agent Got Better

In his Substack, Nate describes a case that brings the whole agent euphoria down to earth: Vercel trained a sales agent on one of its best salespeople, had it filter inbound inquiries, qualify leads, research companies, and draft responses, and reduced the inbound team from ten people to one person who supervises the agent in Slack. Business Insider told this as a staffing story. The more useful story lies elsewhere: the agent works because a workbench was built around it, with sources, defined tasks, handover points, and a human who reads along. Nate names two ways agents break: the world around them drifts away, or, strangely, the model underneath gets better, and the structure that was supposed to compensate for its old weaknesses becomes dead weight. This is exactly what happened at Vercel: 80% of the tools were removed, performance went up. His core point: “More” feels like care, but it's usually what makes the agent rot from the inside. → Nate from Nate's Substack

Synthszr Take: This aligns brutally well with what we wrote in Code Crash about orphaned agents: left to their own devices, they become outdated and lose their users' trust. The reflex to pile on more tools, more context, more memory is not an expression of care, but of laziness in thinking about the actual intent. Maintenance is the unsexy discipline that never appears in any pitch deck, and it will become one of the most valuable AI skills of all, because an agent keeps producing long after it has stopped being right. Vercel's lesson can be applied tomorrow morning, not after the next AI offsite: go through your agents, ask for each tool whether it still solves a problem that the model doesn't already solve on its own today, and delete mercilessly. The control plane everyone is talking about is worthless without a human support layer that curates and corrects. Anyone who thinks the work is done after deployment still has the most expensive part ahead of them.

The Moat is Now Called Megawatts

The AI race has reached its absurd phase: it's no longer about who trains the best model, but who buys turbines the fastest, who sues their way through environmental permits most elegantly, and who relabels a data center as national infrastructure most convincingly. The new competitive advantage, according to AI Secret, lies in megawatts, gas, lawyers, and a DOJ memo. Frontier AI competes on logistics and legal cover, and xAI has just demonstrated that the US federal government provides this cover: Elon Musk's company is apparently getting backing to classify its gas turbine plant in Memphis as defense-relevant. This shifts the entire bottleneck away from the algorithm to the physical world of electricity and permits. If you don't have access to energy and political cover, you don't build models anymore, no matter how good your team is. → AI Secret

Synthszr Take: The ability to build was never a moat, and now the ability to compute is no longer one on its own either. The real bottleneck is in the power line and the environmental agency's permit stack. This is a Cornered Resource in the purest sense of Hamilton Helmer: whoever secures turbines, gas supply contracts, and a favorable DOJ classification holds something that a clever competitor can't replicate in six months. In June, I wrote here about Meta's Mad Max-style tent data centers; the sequel is that physical infrastructure is now becoming legal infrastructure. What's interesting is the flip side: this very dependence on permits and politics makes the big players vulnerable to anyone who can get by with less power—that is, to efficient open-source models and leaner architectures. Anyone building an AI strategy today should take the question of power as seriously as the question of the model. And not be convinced that the most expensive stack is automatically the superior one.

The AI Product Operating Model: How AI-Native Companies Win

Marty Cagan made the Product Operating Model famous, but today's leading AI companies work with a different recipe. At Anthropic, engineers prototype hundreds of solutions and ship without ever asking a product manager or designer. OpenAI's new product Codex started with two PMs, one designer, and about 40 engineers, jointly responsible for 10 to 12 product surfaces, each of which would have its own team of 15 to 20 people in a classic company. Cursor operates with 40 engineers and exactly one PM and has cracked $4 billion ARR faster than almost any company in history. Aakash Gupta calls this pattern the “AI product operating model” and predicts that most companies will migrate to it in the coming decade. He developed the analysis with Rohan Varma, a PM at Codex and formerly the first PM at Cursor. → Aakash Gupta from Product Growth

Synthszr Take: The exciting number isn't the revenue, but the divisor. One PM for 40 engineers at Cursor, that's a tenfold increase in product velocity per capita, and velocity is the solution to almost all software problems. What's disappearing here is the coordination layer: the stakeholder meetings, the handoffs, the squad logic that sends a single feature through five hands until it belongs to no one. Engineers who prototype and ship themselves are reclaiming the product instinct that classic organizations have externalized into processes for years. The problem for the German Mittelstand is located right here: 20 percent AI adoption doesn't do much if the underlying operating model remains byzantine and every decision has to go through six committees. Changing the divisor can be decided tomorrow morning, not after the next strategy offsite. Whoever equips small, deep teams with real skin in the game and lets them ship will win the decade.

60 Percent of US Consumers Find 'AI' in Brand Messaging Off-Putting

A new study by WordPress VIP, the enterprise arm of Automattic, reveals a gap between machine visibility and human trust. 60 percent of surveyed US consumers find brands that advertise with “AI” to be a turnoff, and 86 percent distrust AI answers so much that they still want to go to the original source. Particularly damning: 42 percent trust AI answers without clear sourcing less than airline fees, convoluted privacy policies, or medical bills. Nearly three out of four respondents feel that the internet feels “less human” than it did ten years ago. At the same time, pressure is growing from the other side: 60 percent of companies report increasing traffic from AI search engines, and 74 percent of decision-makers call AI discoverability a key priority. The survey includes 2,000 people, including 800 decision-makers and CMOs, as well as 1,200 adult US consumers, surveyed in April. 80 percent want web information to remain openly accessible rather than being controlled by a handful of large corporations. → Techpresso

Synthszr Take: The brand now lives in two places, and both are pulling in opposite directions. For the model, you have to be machine-readable, otherwise you won't show up at all (in B2B, only one in eight SaaS brands appears in AI answers). For the human sitting behind the agent, that same AI prominence turns negative as soon as it's presented as an advertising promise. “AI” in the claim was a valuation driver in 2024; in 2026, it's a 60 percent turnoff. To resolve this, you build content that is citable for the machine and credible for the human, with real sourcing instead of cotton-candy AI marketing. That can be decided tomorrow morning: strike the word “AI” from the headline and make the substance behind it visible. Trust is the scarce resource in the agent economy, not visibility.