▶ YouTube Summaries

← All summaries
t3dotgg#▶ Watch

Codex $200 plan after the usage cut: cheaper 6.1 Soul, pricier Astra

2026-10-07 Wed ⏱ 33 min t3dotgg

Theo breaks down OpenAI's change that halves the dollar value of the $200 Codex plan (his week gave only about $570 of API-priced usage, roughly $2.5K a month versus $12K before). His argument is that both the subscription number and the API price are made up, and margins are collapsing because Anthropic's Fable pricing and OpenAI's cheap 6.1 Soul forced the gap to close. Per task, 6.1 Soul is so cheap that the cut barely hurts, so the plan is not a rug pull unless you insist on Astra. The real blunder is the $500 "Ultrafast" plan, whose weekly limit burns in about 2 hours. His verdict is that the Claude plan is still the better deal for shipping work, with Codex as a reviewer.

Why the numbers are fake

  • Before the change a $200 Codex plan gave up to $12K a month of inference, and a $200 Claude plan about $8K. The $200 tiers were meant to feel unlimited; the labs kept subsidizing them to avoid the optics of a new model making them feel limited.
  • OpenAI changed the usage calculation right before Dev Day: half the dollars. Theo measured $573.86 in a week and killed a whole account in 29 hours in a loop.
  • The subscription value and the API price are both invented. API prices are what enterprises are charged, historically at about 95% margin: real compute and energy for $100 of tokens is closer to $2-5.
  • With competition the gap between the two should close, and it now seems to be closing.

Claude plan quirks versus Codex

  • The Claude sub works only inside Claude Code (T3 Code works because it calls Claude Code via the agent SDK). Pi, OpenClaw and Hermes don't work, and even a Pi-style system prompt inside Claude Code gets the request rejected.
  • The Codex sub can be used almost anywhere, including CI PR review, as long as you don't resell it to serve user traffic.
  • Fable counts for only half of the Claude allocation. Theo hasn't touched his Fable allowance (600% unused) and instead drains his six Claude accounts with Opus 5.5.

Margin math: Fable, Astra, 6.1 Soul

  • Mythos preview was planned at $125/M output; at a 95% margin that is $6.25 of cost and $118.75 profit. Fable 5 (same model) shipped at $50/M, so margin falls to 87.5% and profit to $43.75. This is "the start of the end" of the subsidization, and Anthropic did it to hurt OpenAI.
  • OpenAI priced Astra identically to Fable ($50/M output), so its pricing follows the competitor rather than its own margins. Anthropic's 50% Fable cap roughly tracks the halved profit.
  • 6.1 Soul looks planned as an Opus 5-class model ($5/$25), but shipped at $2/$10 with cache reads halved to 10 cents. Cache reads were about half of a typical OpenAI bill, so effective cost is nearer 5x lower than planned. Theo guesses margins around 50%.
  • Instead of a special Soul cap, OpenAI simply cut the subsidy across the account.

Effective cost: why it is not a rug pull

  • Theo's own usage: about 1,800 Astra responses cost $534 at API prices, while about 1,500 Soul responses with similar output tokens cost $38.27, more than 10x less.
  • Soul is meant to make the $200 plan feel unlimited again, as in the 5.4/5.5 days, if you avoid Astra and Max.
  • On Terminal Bench (run by Theo in native harnesses), Soul Max had the best score ever and Soul XHigh tied Opus 5.5, at $0.96 per task against $13.11, about 13x cheaper. Slopalytics, his new dashboard, shows Soul and Luna defining the Pareto frontier.
  • So the dollar value halved but tasks per dollar went up several times. He still prefers his Claude sub to his Codex sub, but calls 6.1 Soul the best-value model ever.

The $500 plan and Ultrafast mistake

  • Announcing the 20x plan becoming 10x, with a 25x plan at 2.5x the price, in the same breath as the cut was an optical blunder. Theo says these subscriptions are marketing, not profit, so this one failed.
  • Ultrafast Astra costs $60/M input, $6 cache read, $75 cache write, $300/M output ($450 with long context). Two PRs reviewed cost $600.
  • OpenAI shipped the cheapest model per intelligence and the most expensive one on the same day. Ultrafast apparently isn't on Cerebras after all (reportedly Jane Street bought the chips), so margins are likely very low and capacity over-provisioned.
  • The $500 plan gives about $1,000 of Ultrafast value a week, about 2.1 hours of generation, and about $4K a month versus $12K a month on the old $200 plan: price 2.5x up, usage 3x down.
  • Combined with the cut limits and the new tier, this burned the goodwill Soul had earned. OpenAI's old post-training advantage is also gone now that Anthropic has caught up.

Verdict

  • If the $200 plan is "the Astra plan", it now sucks; get the Claude plan instead. For serious shipping work the Claude plan is "mandatory".
  • Theo uses Codex mostly for emails and computer use, and via T3 Code for Claude-called sub-agent review, where OpenAI models are more thorough.
  • OpenAI should build models as capable and pleasant as Opus 5.5 that are faster at sane prices, and never brand the $500 plan as the Ultrafast plan.
  • Subsidization will keep shrinking as enterprise API prices fall, but a year from now the same plan should still get more work done.