← analysis

The State of the Agent Platform Race

Raw model intelligence is converging at the top. The race is now about orchestration, context, distribution, and trust — which is why the scoreboard doesn't look like the benchmark leaderboards.

Myagi·June 1, 2026·9 min read

If you ranked these seven companies purely on benchmark scores, you'd get a tight cluster at the top and a story that's mostly over. That's not the ranking our scorecard produces, because raw intelligence is no longer the scarce thing. Four providers can credibly claim frontier-class models. The question that actually decides the next decade is different: who can turn a model into a platform that does work on your behalf, across your apps, with your context, in a way you can trust and revoke?

That's the agent operating system. And measured against it, the leaderboard reshuffles.

Intelligence converged; orchestration didn't

Look at the top of the OpenAI Assistant Intelligence 9 column and you'll find a cluster: OpenAI, Google, and Anthropic all sit at a 9, with Microsoft and xAI close behind. If that were the whole game, the race would be a photo finish among four.

It isn't, because OpenAI Cross-App Actions 5 tells a completely different story. Doing things across your apps — booking, sending, editing, scheduling — is where the cluster fractures. The companies that own a surface where actions live (Microsoft inside M365, Google across Android and Workspace) pull ahead of the pure-model players, who reach into apps from the outside.

The four things that actually separate the field

Distribution. Google and Microsoft don't have to win users; they already have them, inside products people open every day. That's why Google overall 7.8 and Microsoft overall 7.4 lead the overall ranking despite OpenAI and Anthropic matching or beating them on raw capability.

Context. An assistant that doesn't know you is a search box with better grammar. Google's Google Personal Context 8 and Microsoft's Graph-powered equivalent are durable advantages. Anthropic, for all its agentic prowess, sits at a 4 on Anthropic Personal Context 4 — a deliberate gap, but a gap.

Protocols. The most important platform move of the past year wasn't a model. It was a protocol: MCP. It's why Anthropic's Anthropic Third-Party Integrations 8 punches above its consumer footprint — an open standard the whole industry is adopting turns every connected tool into something agents can use.

Trust. Enterprises buy governance, not vibes. Microsoft's Microsoft Enterprise Readiness 9 9 is the quiet moat under the whole sector — the place durable revenue actually lives.

Where this leaves each player

• Google holds the best raw hand: frontier models, custom silicon, the OS, the data, the cloud. Its risk is coherence, not capability.
• Microsoft owns the enterprise control plane and is steadily de-risking its model dependency. Distribution is the moat.
• OpenAI has the brand and the developer platform, and an obvious hole at the device edge — OpenAI On-Device AI 2 is a 2.
• Anthropic is the agent-builder's platform and the author of the era's connective tissue, content to skip the consumer surface for now.
• Apple, Meta, and xAI each own one extraordinary asset — the device, the open-weights ecosystem, real-time data — and are missing the rest of the stack.

The north star

Every dimension feeds one question: Google MyAGI Alignment 7 — how close is anyone to a true agent operating system, where capability is earned rather than asserted, memory is durable, orchestration is real, and execution is governable and revocable? No one is there. The leaders are around a 7. The gap between "powerful assistant" and "control plane you'd trust to act" is the whole opportunity — and it's the gap we'll keep grading, keynote by keynote.