Google's launching a frontier model while its own engineers doubt it can deliver, and the company just killed the feature that let users build their own agents.
The Summary
- Google released Gemini 4 Argon, claiming frontier performance in cybersecurity, coding, and enterprise knowledge work, and positioning it to tie with OpenAI's GPT-6 Astra on cyber defense benchmarks.
- Internal skepticism runs high, particularly around coding performance, the same capability Google is marketing as a strength.
- Days before launch, Google killed Gems, its feature for building task-specific agents, just as competitors lean into specialized AI tooling.
- Initial rollout goes only to vetted governments through the Fairwind Program, not the enterprise customers Google claims to be targeting.
The Signal
Google is calling Gemini 4 Argon its return to the frontier, but the evidence suggests something messier. The company skipped Gemini 3.5 Pro entirely after multiple internal delays due to poor performance. Now it's leapfrogging to version 4, betting that branding and marketing can paper over execution gaps.
Argon's claimed strengths are cybersecurity defense, coding, and high-stakes knowledge work in legal and finance. These are the exact domains where OpenAI and Anthropic have been eating Google's lunch for the past year. Google needs this to work. But Bloomberg reports that Google's own engineers are skeptical about the model's coding ability, one of the headline capabilities.
"Google skipped Gemini 3.5 Pro entirely after multiple internal delays due to poor performance."
Here's where the timing gets interesting. Google ties with OpenAI's Astra on the CWE benchmark for cybersecurity, but the distribution strategy tells a different story. Instead of launching to enterprise customers or even to Google Cloud users broadly, Argon goes first to the Fairwind Program, a set of vetted governments and cyber authorities. This is either extreme caution or a tacit admission that the model isn't ready for the wild.
Meanwhile, two days before the Argon announcement, Google killed Gems, the feature that let users create task-specific agents. The stated reason is a shift toward "skills," but the real reason might be simpler: Gems didn't work well enough, and maintaining a half-baked agent-builder while competitors ship functional agents makes Google look slow. Meta just launched Muse and Instinct as all-in-one agents. OpenAI has custom GPTs. Google had Gems, and now it doesn't.
Key tensions in this launch:
- Marketing says frontier performance, engineers say they're skeptical
- Product positioning targets enterprise knowledge work, but distribution starts with governments
- Google claims to lead in cybersecurity AI while only tying with OpenAI on benchmarks
The Fairwind rollout is worth watching. If Google is using governments as beta testers for a model it plans to sell to law firms and banks, that's a savvy play. If Fairwind is the entire strategy and enterprise access comes later or not at all, that's a retreat dressed up as a launch.
The Implication
If you're an enterprise buying AI for coding or legal work, wait for third-party evals and real-world performance data before committing to Argon. Google's track record on the Gemini line has been aspirational benchmarks followed by delayed or canceled releases. If you're building agents, don't count on Google to provide the platform. The Gems shutdown signals that Google is chasing OpenAI and Anthropic rather than defining its own path.
Watch what governments in the Fairwind Program actually do with Argon. If they adopt it for cyber defense at scale, Google has something. If they test it and walk away, the internal skepticism will prove warranted.
Sources
Google DeepMind | Business Insider Tech | Bloomberg Tech | TechCrunch AI