Anthropic's Glasswing and OpenAI's $100 Tier Reshape the Vendor Map

Anthropic locked its most capable model behind a security-research gate while OpenAI slotted a new tier right on top of Claude Max. The vendor pecking order just got weirder.

The frontier labs spent the past week making moves that tell you more about their strategies than any benchmark ever could. Anthropic revealed a model it won't sell you. OpenAI introduced a price point aimed squarely at a competitor. Google kept quietly shipping open weights. Read together, the week reframes how each lab wants to be understood.

A model you can't buy

The headline event was Project Glasswing, Anthropic's initiative to harden critical software using a new frontier model called Claude Mythos Preview. According to Anthropic, Mythos has already found thousands of high-severity vulnerabilities across every major operating system and browser. The partner list — AWS, Apple, Google, Microsoft, NVIDIA, JPMorgan, CrowdStrike, the Linux Foundation — reads like a who's-who of infrastructure. And Anthropic is committing up to $100M in usage credits to get it in their hands.

The twist: Anthropic explicitly will not make Mythos generally available. Alex Albert, who's been at Anthropic nearly three years, called it the most consequential AI event he's witnessed up close. Simon Willison, usually skeptical of gated access, called the decision justified given credible alarm bells from the security community. A frontier lab voluntarily withholding its best model is a new posture — one that says capability and deployment are no longer the same question.

OpenAI draws a line through Claude Max

Two days later, OpenAI introduced a $100/month Pro tier offering 5x more Codex usage than Plus. The price is not a coincidence: Anthropic's Claude Max sits at exactly $100. OpenAI is done letting Anthropic own the power-developer tier uncontested. The same week brought ChatGPT Desktop, the Codex App, Codex CLI, and Atlas — a coordinated push to make Codex the default agentic coding surface before Claude Code consolidates its lead.

Meanwhile Google shipped Gemma 4 under Apache 2.0, hit 10M downloads in a week, and kept its open-weights flywheel turning. Mistral announced Forge for enterprise-trained models and a Voxtral TTS that reportedly beats ElevenLabs in blind tests. Meta's Muse Spark launched closed-weights — a reversal of the Llama playbook that drew sharp reactions from the open-source crowd.

What the split means

For builders, the pricing compression at $100 means the Claude-vs-Codex decision is now about workflow fit, not budget. For buyers, Glasswing is a preview of a world where the best model for your problem may simply not be procurable — capability is decoupling from access. For creators and non-technical users, Karpathy's observation is worth sitting with: a huge gap is opening between people who've used last year's free ChatGPT and people using current agentic tools. The labs are now competing on entirely different axes.

The bottom line: If you're evaluating vendors this quarter, price-match the tiers but buy on workflow — the $100 ceiling is now a floor, and the real differentiation is what the model is allowed to do for you.

Sources