The Agentic Squeeze: Inside OpenAI’s API Shift and Airwallex’s $11B AI Wallet

Yesterday in AI  |  June 27, 2026

 The Agentic Squeeze: Inside OpenAI’s API Shift and Airwallex’s $11B AI Wallet 

Federal oversight is shifting from formal mandates to quiet, downstream gatekeeping. Today's episode breaks down the launch of OpenAI's GPT-5.6 Sol, Terra, and Luna, and the White House's direct request to restrict initial access to a hand-picked list of trusted partners. We examine OpenAI's public pushback against this new pattern of government intervention. 

We dive into the product updates from Figma’s Config conference, including Figma Motion and bidirectional GitHub sync tools designed to change how code and design assets interact. We cover Adobe's strategic acquisition of Topaz Labs to run high-end enhancement models locally on consumer hardware. Plus, we analyze Airwallex's massive $320 million funding round for autonomous corporate finance, and look at internal OpenAI data showing that agentic workflows have officially taken over the majority of API infrastructure, creating a massive pricing crunch.

Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, or Bluesky. If you like the show, please take a minute to rate and review it so others can find it!

Transcript

SPEAKER_00 0:03

Yesterday in AI. Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of AI in 10 minutes or less. I'm Mike Robinson. It's Saturday, June 27th, and the U.S. government just put its hand on the release lever for Frontier AI for the first time. Figma completely changed what a design canvas can do, Adobe made an acquisition to move past raw generation, FinTech is building the money rails for autonomous software, and we finally have hard data on how fast the enterprise shifted to agentic AI. Let's get into it. OpenAI launched three new models on Friday, GPT-5.6 Sol, Terra, and Luna. Sol is the flagship engine, showing massive capability gains in coding, biology, and defensive cybersecurity. OpenAI specifically pointed out that Seoul excels at helping software teams find and patch security vulnerabilities rather than executing offensive cyber attacks. That specific engineering framing is intentional. The White House directly asked OpenAI to restrict initial distribution to a list of 20 government-approved trusted partners before opening Seoul to the general public. OpenAI agreed to the request. They haven't named the 20 companies involved, but they plan to add more partners next week and target a broad public launch in the coming weeks. OpenAI simultaneously published a statement making it clear they disliked this setup. They explicitly stated a government access process should not become the standard long-term default for software releases. They agreed to this isolated request while fighting the broader policy principle in the exact same breath. We tracked the hard export controls that forced Anthropic to pull its models offline two weeks ago. That was a mandatory federal order. This OpenAI situation has soft power, an informal request, voluntary corporate compliance, and a public protest. This exact mechanism establishes a pattern where future administrations can request soft blocks before public launches. The cybersecurity footprint of Seoul is what alarmed Washington, a model that isolates code vulnerabilities as a dual-use asset by definition, capable of patching systems or shattering them depending on who controls the prompt. The White House executive order from June 2nd created the covered frontier model category, specifically to gate software with these exact capabilities. This leaves enterprise software buyers in an awkward position. Corporate infrastructure teams now have to wait on an opaque government review process before deploying the best tools on the market. It creates a secondary layer of procurement friction that tech teams have to map out in their vendor risk logs. And while OpenAI was negotiating with federal regulators in Washington, Figma was in San Francisco showing how AI can build production ready software layouts. Figma hosted its annual config conference from June 23rd through June 25th, dropping a series of massive product updates. The standout edition is Figma Motion, which embeds a full animation timeline directly into the design canvas. Designers can build and refine animations organically, or instruct the native Figma agent to generate UI transitions from simple text descriptions. The platform then compiles these animations into production ready code for engineering teams. They also introduced code layers, establishing a bidirectional GitHub sync that anchors design assets directly to live code bases. Product teams can initiate builds from visual layouts or raw code, and the underlying AI agent fixes the drift between them. They added AI-generated shaders, allowing designers to dictate complex visual textures in plain language and immediately receive functional code blocks. They also updated Weave, their generative plugin architecture, packing 20 separate AI image tasks directly into the core canvas. CEO Dylan Fields stated at Config that AI lowers the floor, but designers will raise the ceiling. Lowering the floor means anyone can generate a low fidelity asset in seconds. Figma is betting that professional designers using native animation, automated code sync, and procedural shaders will build things that were physically impossible before. It shifts the goal from raw speed to creative complexity. Regulators blocked Adobe from buying Figma for $20 billion on antitrust grounds back in 2022. Figma is now shipping native code synchronization in integrated animation environments while Adobe is forced to buy external plugins to keep pace. It shows how fast competitive defensive modes evaporate in software. Figma is now shipping native code synchronization in integrated animation environments while Adobe is forced to buy external plugins to keep pace. It shows how fast competitive defensive modes evaporate in software. The entire creative sector is moving toward extreme quality and immediate deployment readiness, completely abandoning the early wave of simple image generation. Adobe clearly recognizes this shift, which explains why they just bought the premier upscaling lab in the industry. Adobe announced Thursday that it is acquiring Topaz Labs. Topaz is known among professional creators for its specialized image and video enhancement tools, specifically Astra for video upscaling and Wonder for photographic retouching. These are the utility tools editors use when an AI generator gets an asset 80% of the way there, and they need the final 20% to look flawless on a broadcast screen. Adobe plans to bake Topaz's models directly into Firefly and the broader Creative Cloud Suite, though Topaz will continue operating as a standalone application. The transaction is scheduled to close in the second half of 2026. This acquisition targets a massive engineering problem. The first wave of creative tools focused entirely on raw generation, generating new pixels from scratch. This second wave focuses on production quality. Topaz spent years figuring out how to run massive video upscaling models locally on consumer computers. Most of Adobe's 30 million creative cloud users work on local laptops, not remote GPU servers. Running high-end upscaling code inside Premiere or Photoshop in near real time without a latency-heavy cloud round trip requires intense mathematical optimization. Adobe didn't just buy a feature, they bought the local execution architecture. But while creative professionals are optimizing pixels locally, fintech giants are raising hundreds of millions to build the compliance infrastructure for autonomous corporate wallets. Airwallox raised $320 million at an $11 billion valuation, marking a 38% increase in its private market cap compared to six months ago. Addition led the funding round with participation from T-Row Price, Bailey Gifford, and Amex Ventures. The company's financial records show its annualized revenue hit $1.3 billion as of March, a 74% spike year over year. The funding numbers are large, but the specific product launches explain the premium valuation. AirWALEX launched TZero, an autonomous corporate finance platform that processes bookkeeping, tax filings, legal compliance, and financial reporting without human intervention. The system is currently in a closed private beta. They also launched AirI, an agentic wallet engineered for autonomous AI software to execute corporate purchases. The wallet includes hard programmable spending caps and granular permission controls. We covered visa embedding consumer payment protocols into ChatGPT on June 12th. That was a legacy card network building consumer spending paths. AirWalls is attacking the corporate side, building the B2B verification layer. Trusting an AI agent with a corporate line of credit is an operational nightmare for a CFO. You need ironclad audit trails, real-time fraud monitoring, and regulatory compliance across 85 distinct international jurisdictions. Airwalix holds those 85 regulatory licenses natively, giving them a massive head start in becoming the primary financial rail for autonomous software. Enterprise software needs these automated wallets because internal corporate usage data shows that software is completely abandoning human-in-the-loop operation. New internal data from OpenAI reported by the Deepview reveals that over 50% of all OpenAI API traffic is now completely agentic. Autonomous multi-step tool using workflows are now the literal majority of their business. A year ago, API calls were dominated by single-turn completions. A user asked a question, the model returned a paragraph of text, and the connection closed. Agentic workflows are different. The model spends minutes reasoning across multiple separate loops, calling external databases, verifying information, and holding memory across long operational chains. Those calls now command the majority of the network. This structural shift carries massive economic consequences for tech budgets. Agentic calls ingest massive amounts of context, run longer, and strain cloud infrastructure. They cost significantly more to process. A basic chat completion takes one second and consumes 500 tokens. An autonomous agent booking corporate travel, auditing calendars, and drafting legal alerts can easily run for 40 seconds and burn through 15,000 tokens. When you multiply that compute density across millions of active enterprise users, the flat rate subscription models collapse. This data explains the financial strain behind the GitHub copilot billing numbers and the sudden token overage fees hitting corporate tech invoices. The major platforms built their pricing models for simple chat conversations, but corporate usage rapidly morphed into automated agents. That friction is creating a massive financial crunch, and no one has dropped a sustainable solution yet. And that's it. If you have any feedback about this show, you can email Mike at yesterdaynai.news, or you can find me on LinkedIn, X, or Blue Sky. And if you like this podcast and want to see it continue, please take a minute to write and review it so others can find it. Thanks. That's all for this edition of Yesterday and AI. Stay curious, have a great weekend, and I'll see you on Monday.