# How to Get Free Tokens? Almost. You Just Need… URL: https://danylovorvul.com/blog/how-to-get-free-tokens-almost/ Author: Danylo Vorvul Date: 2026-08-12 Tags: agents, tooling Description: A $20 subscription for the brain, free models for the hands. What my agent's 48-hour run actually cost. I'm feeling clickbaity today. Sorry πŸ˜„ For the past 48 hours, my personal agent Hermes has been grinding through my financial data without a break. Two days of continuous agent work. Total damage: 12% of the weekly quota on a $20 subscription. The title promised free tokens. This is the "almost." The brain costs $20. The hands cost nothing. The $20 is ChatGPT Plus. Inside it: Codex, which carries the GPT-5.6 family (Sol, Terra, Luna), with enough Sol messages in every five-hour window to run a serious agent, if the agent isn't doing the work itself. That's the whole trick. Sol never touches the bulk work. It plans, splits, delegates, and checks. The work happens somewhere tokens cost zero. I love watching Sol orchestrate Terra and Luna. That's the in-house version, and it's already good. But the version this post is about reaches outside the subscription entirely: Sol orchestrating free models from other people's clouds. ## Where the free tokens live Three catalogs, one shopping rule: pick one high-intelligence model, one medium, one small coding model. NVIDIA (build.nvidia.com) hosts 100+ models behind an OpenAI-compatible endpoint, free to call at roughly 40 requests per minute. Yes, per minute. The prize there right now is GLM-5.2: a 753B open-weight MoE that lands within about 1% of Opus 4.8 on FrontierSWE. On most other benchmarks Opus is still ahead. A near-frontier agentic-coding model, then, not a free Opus. For delegated bulk work, the gap doesn't matter. GLM is what spent those 48 hours on my finances, turning the transaction feed from my vibecoded bank integration (read-only, so we're good) into an actual budget. That's my high tier. OpenRouter tags free models wy fifteen exist as I write this: wall-to-wall Nemotron, plus strays from Google, Poolside, and Cohere. Limits: 20 requests per minute, 50 per dcredits once, ever, whichunlocks 1,000 per day permanently. Best $10 in this whole setup. The small coding models there β€” the kind that drafts a ten-line script faster than I can describe it β€” are my low tier. Nous Portal runs a rotating free catalog at 50 requests per minute, typically their own Hermes series plus a handful current list lives in thedashboard; it changes, which is the point of the next section. ## The wiring The portable version needs none script that curls the endpoint, registered as a toohe brain's prompt: bulk analysis goes to GLM, script drafts go to the mini, only conclusions come back into your context. If your agent supporr worker; the curl version works everywhere. Then the part nobody does: the catalogs churn. Last year's lists promised free DeepSeek and Qwen on OpenRouter; todayxt build is a weekly cron job:Hermes pulls each catalog's /models endpoint, filters price = 0, runs one probe task per tier, and files a report best low, what changed. Rightnow I do this scan by hand every week. The manual version already works; the cron just removes me from it. ## The cons, honestly It's slow. 40 requests a minute is a queue, not a firehose. Free endpoints may train on what you send. Read the data policy before piping anything through that you wouldn't hand a stranger. Delegation itself has a tax: the brain spends tokens writing the handoff and verifying the result, and on small tasks the tax exceeds the savings. And errors compound: a five-step chain of 95%-reliable steps completes 77% of the time. The literature on multi-agent failures is genuinely ugly. The loudest counterargument β€” that mixing weaker models drags quality down β€” is published and fair: when Princeton ran the numbers, dropping the weak models beat keeping them. So this is not for the task you need at 3 pm. It's for the task you need eventually: the backfill, the audit, the 48-hour grind. Slow. Rate-limited. Occasionally wrong on the first pass. But running while you sleep, on a subscription you already had. --- Notice what the $20 actually bought, though. Not the tokens β€” those were free. It bought the one model in the stack whose job is judgment: which task deserves which worker, and what "done" looks like. Which is the same thing a company pays a manager for. Not doing the work β€” knowing whose work it should be.