Hacker News front pageJul 19, 2026, 20:01 UTC+8

I burned all my tokens researching how to save tokens

Bartosz Kotrys at Quesma burned his entire Claude Max 5x plan in 30 minutes running /deep-research with zero results. He turned Claude Code into a harness, added Codex (GPT-5.5) and Antigravity (Gemini 3.1 Pro) as headless subagents, and pooled tokens across three subscriptions—no extra cost, roughly 10x longer runs. Models are role-pinned: Sonnet 5 searches, Opus 4.8 verifies, Fable 5 only plans and judges, Haiku 4.5 handles extraction. Every claim must be verified by a different model and backed by a URL plus a direct quote. /deep-research now runs last, at end of day, not first.

Why it matters: First-person experiment with concrete numbers on Claude Max deep-research quota limits and a reproducible workaround using existing subscriptions. Capped at 72 because it's a personal blog post, not a product launch or research release.

Read the original ↗Export Markdown

Take

This one's worth clicking because it solves a real pain point: a Claude Max 5x plan burned through in 30 minutes of /deep-research with zero results. Bartosz didn't spend more money—he pooled tokens from three subscriptions he already had (Claude, Codex/GPT-5.5, Antigravity/Gemini 3.1 Pro), used Claude Code as the harness, and ran the other two as headless subagents. Models are role-pinned: Sonnet 5 searches, Opus 4.8 verifies, Fable 5 only plans and judges, Haiku 4.5 handles extraction, Codex runs terminal tools, Antigravity provides a second opinion from a different model family. Every claim must be verified by a different model with a URL and direct quote. /deep-research now runs last, not first.

I'd discount this a bit: it's a one-person engineering setup, not a product. The core trick is a small Bash script that wraps different CLIs as subagents with shared local memory and automatic fallback on rate limits. If you don't have multiple subscriptions or don't want to mess with the command line, this path isn't for you. But the thinking is sharp—don't fight Anthropic's pricing head-on, orchestrate the resources you already have. The post doesn't give a quality comparison of the research output, just says "roughly 10x longer runs." You'll have to test the actual results yourself.

Source: Hacker News front page · quesma.com
Published: Jul 19, 2026, 20:01 UTC+8 · Jul 19, 2026