Cheap intelligence is useless if your context is trapped: GLM-5.2 vs. Claude Tag

By Nate (Nate's Substack) · June 28, 2026.

This executive briefing starts from an apparent contradiction Nate detected during the week: artificial intelligence is becoming cheaper and more expensive at the same time. On one hand, the arrival of GLM-5.2 confirms that a growing portion of coding and processing work no longer needs to pay frontier-model prices. On the other, Anthropic launched 'Claude Tag' —a direct integration with Slack— without making its product cheaper, and its enterprise customers were absorbing higher bills and paying them anyway.

In May, The Information had already reported that enterprise buyers of Claude expected to pay more, not because Claude was irrelevant, but precisely because it had become useful enough and woven enough into workflows that no one wanted to turn it off. That data point is the central knot of the analysis.

The surface explanation is that frontier labs still have market power: Claude is good, companies want the best tools, and if a $250,000-a-year employee becomes significantly more productive thanks to an AI tool, the organization will tolerate an uncomfortable AI bill for quite some time. But that explanation, Nate warns, doesn't answer why that pricing power might survive as open models improve, nor why the Claude Tag move strikes him as more strategically important than another model launch.

The thesis the briefing develops is that this is not, deep down, a story about price per token. It's a story about where intelligence is allowed to work. Cheap intelligence only benefits you if you can put it to work. If you can't, the discount stays on paper, and the expensive tool remains the one your team keeps using. The real question, therefore, is not whether GLM-5.2 is cheaper than Claude —it is— but whether your organization can capture that discount in practice, or whether it will keep paying the premium because nothing else fits yet into the real work.

The briefing structures the analysis in four main blocks. The first examines why the cheap option is real now: GLM-5.2 means a growing portion of work no longer needs frontier prices, though with the security trade-off that running it yourself (self-hosting) entails. The second block addresses what you're really paying for when you pay Anthropic: most companies can buy the cheap model and still be unable to use it, because the context and permissions surrounding it are the part they haven't yet built. The third analyzes the strategic move of Claude Tag: how a Slack integration becomes accumulated context that grows increasingly hard to abandon the longer your team uses it. This is lock-in, but not the traditional lock-in of contracts or proprietary APIs: it's context lock-in, of history, of organizational memory deposited in a specific tool. The fourth block explains why the obvious solution —building and owning your own context— is right on paper but brutal in practice, fundamentally because it's a hiring and internal-capacity problem, not just a strategic decision.

The briefing closes with seven questions that, according to Nate, every management team should answer now, before the answers are handed to them by the providers through their own product and pricing decisions.

As sector context, the dynamic described —increasingly capable open-weight models competing with proprietary APIs while frontier labs seek anchoring in context integrations— is one of the most important structural tensions of the moment in the enterprise AI industry. The concept of 'context lock-in' is especially relevant for teams evaluating agentic AI strategies, where conversation history, connected tools and accumulated permissions represent a real operational asset that doesn't transfer easily between providers.