I Paid for Claude's Gmail 'Superpower'—and Anthropic's Compute Crunch Made it Useless

Anthropic’s Gmail/calendar “superpower” for Claude underdelivers because the system is compute constrained—leading to hard rate limits, incomplete tool calls, and shallow outputs even on high-priced plans. The result is a frustrating customer experience: Claude pulls only a fraction of meetings and emails, struggles to “try again” to recover missing items, and fails to generate the kind of merged, cross-source insights a user expects from connecting calendar and inbox.

The breakdown starts with tool usage limits. Even on the max plan—described as costing around $100 per month—Claude appears to be rate-limiting backend calls to calendar/docs/email to save costs. In practical terms, that means only about 50 total calls to those sources are available, which burns quickly for everyday workflows like checking calendar twice a day, reviewing emails, and drafting responses. Anthropic signals the limits may be lifted later, but the fact that higher payment doesn’t meaningfully increase the quota is treated as evidence of deeper compute constraints.

That constraint shows up in behavior. When asked to generate a React artifact for a daily briefing using both email and calendar inputs, Claude performs poorly: it makes only one call, drops lists, and captures only part of the available data (e.g., roughly the first seven or eight meetings and the first five emails). Follow-up attempts don’t fix the problem—re-running the task still fails to retrieve the missing calendar entries and the later emails. One plausible explanation offered is that the system ingests sources separately rather than merging them into a single working view, which would prevent the model from synthesizing across the full dataset.

The disappointment also ties to output limitations. Even with a “gigantic” context window in theory, the effective output budget per turn feels much smaller—on the order of ~8K tokens—making Claude’s responses feel chunky and cut corners. The transcript contrasts this with how OpenAI’s architecture can mask token limits through techniques like streaming tokens and storing content server-side, so users experience fewer visible constraints.

More broadly, the transcript argues that the industry’s push toward agentic tool use is colliding with compute realities. OpenAI’s recent rollout of “o3” is framed as having access to hundreds of tools under the hood (around 600), chosen and used dynamically—something that requires substantial compute headroom. By comparison, Claude’s tool use appears throttled, so it can’t reliably execute large multi-step workflows.

The takeaway isn’t that Claude is a weak model—coding performance is described as generally strong, and Claude 3.5 is called a “sweet spot.” The core issue is capital and GPU constraints shaping product behavior: limited tool calls, incomplete retrieval, and gated experiences that don’t unlock even at premium pricing. The transcript concludes that compute constraints are likely to keep influencing rollout quality and “agentic” capabilities across the competitive race, and that capital constraints deserve more attention than token-limit headlines.