▶ YouTube Summaries

• https://www.youtube.com/watch?v=3A4y_Iw_gpU
• Original title: Spending $40,000 in Copilot for $40

Theo breaks down why GitHub Copilot's pricing change isn't a rug pull but a long-overdue fix to a billing model that became absurd in the agent era. He demonstrates by spinning up 50 parallel Copilot sessions running a deliberately unsolvable cryptography puzzle on his $40/month plan to torch thousands of dollars of Microsoft compute as revenge for unreturned Azure credits.

Four ways inference is billed

• Subscriptions with rate limits — Claude Code, Codex. Vague 5-hour and weekly windows. No documented token counts, limits change silently.
• Subscriptions with message limits — old T3 Chat, current Copilot. Fixed N messages/month. Broken when one message can fan out into hundreds of tool calls.
• Subscriptions with spend limits — Cursor, Open Code Zen. You see dollar usage in dashboard, can overflow.
• API per-token — what providers actually charge. Input cheap-ish, output expensive, cached input ~10x cheaper.
• Dedicated compute — labs and enterprises rent GPUs directly.

Why message-based billing dies with agents

• Pre-agent: one message = one API call. Cost variance narrow.
• Agent era: one message can spawn dozens to thousands of tool-call cycles. Per-message cost ranges from cents to $60+.
• T3 Chat hit this in Feb: tiny minority of users used Repo Mix to paste whole codebases, costing $200+/mo on an $8 plan. Forced their billing change.
• Selling "messages" is like selling "cars" at a fixed price — Ferrari vs. abandoned Subaru.

The Copilot exploit

• Theo got a comped $40/mo Plus plan (1,500 premium messages).
• Used 5% of allowance, already cost Microsoft $550+ in inference.
• Method: feed a cryptography puzzle (changed one letter to make it unsolvable) via prompt.md to GPT-5.4-extra-high. Model grinds 16 hours per message, ~111M input + 1.6M output tokens. With caching, ~$62/message. Without caching, $163.
• Ran 50 staggered sessions on a Mac Mini to dodge rate limits. Average $10/message across runs, theoretical max $93,600/month on the $40 plan.
• Reasoning level (low/high) doesn't affect the message multiplier — high is free uplift.
• Model multipliers: GPT-5.4 = 1x, GPT-5.5 = 7.5x, Opus 47 = 15x. Annual plans see multipliers rise June 1 (Opus 15x→27x, GPT-5.4 1x→6x).

The cryptography puzzles (digression that's actually the exploit's engine)

• Puzzle 1: encoded with fake "base 23" (T at pos 20 hints at T3), decoded to JSON pointing at a Git commit hash on his Dogecoin sim repo. Solved by community in 9 minutes after gist link given.
• Puzzle 2: inverted base 64 alphabet (0=`/`, 1=`+`, etc.). Took community 10 hours. Agents couldn't solve without heavy hints.
• The puzzles run for 80-180 minutes on GPT-5.4-Pro. Theo's modified unsolvable variant runs 16h+.

Why the change isn't a rug pull

• New Copilot model (June 1): traditional token-based credit allotment, additional usage purchasable.
• Calling it a rug pull = not understanding cost math. Microsoft has been subsidizing $1,000s of inference per $40 user.
• Real reason for the change isn't greed — it's compute scarcity. Copilot signups disabled because no GPUs. Anthropic partnering with Musk for compute.
• Every other agentic-coding vendor (Anthropic, OpenAI, Cursor) moved off message billing months ago. GitHub was just slow.
• "You should not be able to get more than $40 of inference for 40 bucks. Get good."

Aside: Repo Mix

• Theo blames Repo Mix (pastes entire codebase as XML) for costing T3 Chat ~$500K. Calls users abusing small-business chat apps with it "cringe."