Three Frontier Labs, Three Theses, and Why Your AI Stack Now Reflects a Worldview

Two years ago, choosing between Claude, ChatGPT, and Gemini was a benchmark conversation. Which model was best on coding, which had the longer context window, which hallucinated less on the prompt I cared about. That was the whole question.

The spring 2026 cycle made it clear that the question has changed. Anthropic announced a $200 million four-year partnership with the Gates Foundation, published the first results from Project Glasswing, the cybersecurity program it launched in April with twelve enterprise partners and has since expanded by roughly 150 more organizations, and is in talks (per Bloomberg) to close a round at a valuation above $900 billion. Google's I/O keynote framed Gemini as the intelligence layer underneath every Google-owned surface, from phones and watches to cars and the new Googlebook laptops. OpenAI's positioning, articulated by Sam Altman at a New York editors' lunch last December, named enterprise as the top 2026 priority and framed the remaining work as an application problem rather than a training problem.

Three different positions. Three different stories about what AI is actually for. Choosing which one you build on is no longer a benchmark question. It is a brand alignment question, and that means it is now a strategic decision for any business with an AI surface.

What is Anthropic actually betting on?

Anthropic's 2026 moves have a pattern that gets clearer the longer you stare at the announcements. The Gates Foundation partnership commits $200 million in grants, Claude usage credits, and technical support over four years to specific applied problems in global health (starting with vaccine development for polio and HPV, plus new prevention and treatment approaches for cervical cancer and preeclampsia), education (evidence-based K-12 tutoring and career navigation in the US, plus foundational literacy and numeracy programs in sub-Saharan Africa and India), agriculture (smallholder-farmer tools delivered in local languages), and economic mobility. The output deliverables are named. The geographies are named. The problem areas are named.

Project Glasswing is the same shape. At its April 7 launch, Anthropic gave twelve launch partners (Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks, plus Anthropic itself) access to Claude Mythos Preview, an unreleased frontier model, specifically to find and fix critical software vulnerabilities. By the program's initial update on May 22, Anthropic and roughly 50 participating organizations had found more than ten thousand high- or critical-severity vulnerabilities in the world's most systemically important software. On June 2, Anthropic expanded the program to roughly 150 new organizations, including critical infrastructure operators in more than 15 countries. Anthropic's framing for the program is unambiguous about the bet: Glasswing is, in the launch announcement's words, "an important step toward giving defenders a durable advantage in the coming AI-driven era of cybersecurity."

Same shape as the Gates partnership. A named domain. A named application. Specific partners. Measured success criteria.

If you read those two announcements as Anthropic doing applied-AI demonstrations to soften their image, you are reading them as PR. That is not what they are. They are Anthropic publicly committing to a thesis about where frontier AI is most valuable, which is: at the application boundary of hard, specific human problems, deployed in partnership with organizations that already operate in those problem spaces. The Novo Nordisk clinical study report work fits the same shape. The Gates global-health work fits. Glasswing security work fits. The pattern is not "what model is best" but "what hard problem can a frontier model actually move."

That is a different bet than OpenAI is making and a different bet than Google is making.

How is OpenAI's bet different?

OpenAI's 2026 positioning has moved decisively toward the enterprise. The clearest articulation of it is Altman's December editors' lunch, covered by Big Technology, where he framed the gap as application work rather than model work. The implication is that OpenAI believes the frontier capability is largely in hand and that the remaining business depends on whether enterprises can actually deploy it inside their existing systems.

That framing maps to a different product surface than Anthropic's. OpenAI is building horizontal enterprise scaffolding: ChatGPT Enterprise, ChatGPT Agent (which absorbed the old Operator product), Apps in ChatGPT, AgentKit for developers. The bet is that AI value will be created when a sufficiently capable model is wired into every enterprise workflow, and the company that wires it in wins. That is not a domain-by-domain bet. It is a footprint bet.

Both Anthropic and OpenAI have credible roadmaps. They are not the same roadmap. A team building on Anthropic is building toward a specific named application. A team building on OpenAI is building toward an enterprise-wide deployment. The downstream product decisions look meaningfully different.

How is Google's bet different from both?

Google's bet, made public at I/O 2026, is the OS-layer bet. Google's own framing for the keynote was "Welcome to the agentic Gemini era," and the parallel announcement at the Android Show one week earlier introduced Gemini Intelligence as a proactive agentic layer across Google-owned surfaces: Android phones, watches, cars, and the new Googlebook laptop category, with glasses on Android XR coming later this year. Gemini Spark, the 24/7 personal agent, is the surface product. The OS-layer positioning is the strategic thesis.

That is a third distinct bet. Google is not betting that AI value lives at the application boundary (Anthropic's bet). Google is not betting that AI value lives at the enterprise deployment layer (OpenAI's bet). Google is betting that AI value lives one layer below the apps, at the OS surface that mediates between them. The reason Google can make this bet is that Google owns the OS surfaces in question. The reason the bet is risky is that the apps it mediates may not cooperate, which is the same risk every OS-mediation play has faced since the iPhone.

I wrote about the OS-layer thesis in more depth in my post on Google's OS-layer bet. The relevant point for this post is just: it is a different bet from the other two.

What does the choice of frontier lab signal about your business?

Here is where the framing starts to matter for any business building on top of these models. If two years ago the choice was a benchmark conversation, and now the choice is a strategic conversation, then the lab you choose is a public signal about what kind of business you are.

Anthropic-built businesses look like specialty applied AI: legal, medical, scientific, public-sector, security-sensitive. The choice of model signals that the business cares about deep domain integration, measured outcomes in a named problem space, and the kind of safety posture Anthropic has been articulating since the company started. The Gates partnership reads as proof of concept for that posture at scale.

OpenAI-built businesses look like horizontal enterprise tooling, productivity surfaces, and developer platforms. The choice signals scale, breadth, and the implicit bet that capability parity at the model layer is settled enough that distribution and integration are the only remaining levers. ChatGPT-anchored consumer products fit here too.

Google-built businesses look like OS-mediated workflow surfaces, especially anything that benefits from cross-app context: scheduling, communication, multi-tool research, content production that draws across Workspace. Spark's beta, currently limited to Google AI Ultra subscribers, makes that surface real for the first time.

These are not exclusive choices. Many teams have all three model providers in their stack and route by job. But the primary model provider, the one your product is built around, is becoming a brand signal whether you wanted it to or not. The choice will read to your customers the way "we run on AWS" read in 2015 versus "we run on Heroku" versus "we run on bare metal." It says something about you.

What about the safety positioning, given recent history?

Anthropic's safety story is the most complicated of the three labs. Last winter Anthropic publicly walked back parts of its safety pledge, which I covered when it happened. The honest summary is that Anthropic shifted from a posture of "we will commit to specific external constraints" to "we will commit to specific applied outcomes." The Gates partnership and Project Glasswing are downstream of that shift, not departures from it.

Whether that shift makes Anthropic more credible or less credible depends on which audience you are. For a defense-and-security buyer or a regulated-industry partner, the applied-outcomes posture is easier to evaluate than the abstract-pledge posture, because there are measurable deliverables to point at. For a safety-research audience, the shift looks like Anthropic moving closer to OpenAI's and Google's commercial framing.

For most readers of this blog, who run a business and need to choose a model provider, the safety story matters at the same level as any other vendor due-diligence question. What can the vendor commit to in writing for your specific use case, what is the SLA on that commitment, and what is the remediation path if it fails. Frontier lab choice is brand-level. Frontier lab procurement is contract-level. Both matter, and they are different questions.

What do I bet happens to frontier lab branding by end of 2027?

Here is the prediction I will own. By the end of 2027, the brand alignment between which frontier lab you build on and what kind of business you are building will be as load-bearing in B2B conversations as the cloud provider choice was in 2015. Specifically, expect to see RFPs that explicitly ask which model you build on, and expect at least a third of those RFPs to weight the answer.

If by December 2027 frontier lab choice still reads as a technical implementation detail in B2B sales conversations, the brand-thesis layer did not crystallize the way I am betting it will. If RFPs start asking about it explicitly, the labs' 2026 positioning work successfully shifted the conversation up the stack, and the choice becomes a real strategic decision.

The choice is not Claude versus GPT versus Gemini. The choice is between three different views on where AI value actually lives. The teams that pick deliberately will be the teams whose product positioning aligns with their model positioning. The teams that pick on benchmark and price will keep paying for the same horizontal productivity gains everyone else is buying, and will have nothing distinctive to say about it.