William Angel, who apparently writes about energy technology and has opinions on utility mergers, tried to use Claude's deep research mode to investigate the Dominion Energy/Next Era merger announcement. His conclusion after burning through most of his $20/month Pro plan quota in a single query: "at most one successful research mode question per session."

The Quota Reality Check

Angel walked through his experience step by step on his blog. He fired up Claude Sonnet 4.6 to research the mergerβ€”a topic he has domain expertise in, so he'd know if the AI was blowing smoke. The deep research mode launched, started citing sources, and then... kaput. Hit the usage limits before finishing. But here's the kicker: it failed at "80% of my limits," which Angel rightfully points out is weird because 80% β‰  100%. That's either a rounding bug or something in the allocation logic that doesn't add up. He tried again hours later with an identical promptβ€”no extra connectors, no MCP servers, just a clean session. This time it worked, but consumed 56% of his monthly quota for one research query. The deep research cited 236 sources and did cross-referencing, which is thorough but brutal on token budgets.

Claude Code Does It Better

The comparison that should make Anthropic's product team wince: Angel had a similar experience last week with Opus 4.7 doing spreadsheet modeling work in a resumed session. Burned through an entire quota session and still failed to finish. His suspicion is that "Claude Code would have used significantly less quota for the same amount of actual work." The irony here is thickβ€”Anthropic's own CLI tool might be more efficient at certain tasks than their flagship consumer product.

What's Actually Broken

Angel offers three concrete suggestions if anyone at Anthropic is listening: First, add some kind of hint to Claude as usage starts running low so it can wrap things up gracefully instead of mid-task. Second, build UI hints that suggest using simpler web search for queries that don't need full deep research mode. Third, consider soft overagesβ€”let users burn a little past their limit when a task is clearly in progress rather than killing it cold.

Agentic Workflows Need Budget Awareness

"I'm increasingly coming to believe that features failing outright is a very bad user experience." Angel's take on agentic workflows more broadly is the real insight here: these systems should either reject tasks upfront when projected budget won't cover them, or degrade gracefully within their actual budget. "Otherwise people won't trust the feature itself, let alone trust it to use as many tokens as it can get."

Key Takeaways

  • Deep research mode on Pro plan consumes roughly half monthly quota per query
  • The 80% cutoff before hitting limits suggests a potential bug in usage tracking
  • Resumed sessions + heavy context = quota death spiral
  • Claude Code may be more token-efficient for equivalent workloads

The Bottom Line

If Anthropic wants deep research mode to actually work as a selling point for Pro, they need to solve the budget estimation problem. Users shouldn't feel like they're gambling their monthly allocation every time they ask a complex question. Either give us better visibility into projected usage upfront or build in graceful degradation that completes tasks without mystery failures.