▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=Hq2c7pww1Ds
• Original title :: OpenAI is getting desperate

As of October 29th OpenAI doubles usage limits on all Pro tiers ($100, $200, $500) and Business Premium. Theo argues the change mostly restores the $200 plan to where it was before the September cut, while the $100 and $500 tiers get a real 2x. Comparing per-token, per-task and per-subscription value, Theo finds the Codex and Claude subscriptions much closer in real usage than raw dollar figures suggest, with Opus 5.5 still the better pick on capability. OpenAI's missing piece is a model that can compete with Opus 5.5.

The announcement and its history

On September 28th Tibo announced changes to the $200 Pro plan that cut its effective API-dollar value roughly in half. The justification: no more 5-hour limits, a guarantee that each tier always buys more work over time as models improve, and no incentive to inflate API list prices. The cut coincided with GPT-6.1 Soul launching around Dev Day at a much lower API price than expected — keeping the same dollar allowance would have meant OpenAI eating more compute.

That cut only touched the $200 tier. The new doubling therefore:

• restores the $200 tier to its pre-September level;
• genuinely doubles the $100 tier;
• doubles the $500 tier to 50x Pro-tier usage — hard to exhaust without heavy use of fast/ultrafast modes.

Caveats people complain about: ChatGPT chat and Work usage now count against the same limits as Codex. Theo considers this irrelevant for nearly everyone, since chat uses orders of magnitude fewer tokens than coding agents. Dots reportedly still does not consume limits.

Three ways to price a model

1. Per token — input/output price plus caching. Misleading on its own.
2. Per task — tokens needed to solve a problem. OpenAI models use far fewer tokens: GPT-6.1 Soul at xhigh uses fewer tokens than Opus 5.5 medium and scores 52 vs 48 on the Artificial Analysis Intelligence Index, at under half the per-task cost. On per-task API pricing OpenAI dominates the Pareto frontier; Opus wins only at the very top on raw capability.
3. Per subscription — what a $200 plan actually buys.

Dollar value of subscriptions: OpenAI $200 ≈ $4.3K of API usage, $500 ≈ $10.5K. Claude $200 ≈ $12K, or ≈ $6K counting the half limits on Fable (which, Theo notes, no one needs over Opus right now).

Subscription-adjusted comparison

Using Adam Holter's subscription-adjusted charts, the gap narrows sharply. At high effort Opus 5.5 costs about 5 cents per task versus 4 cents for Soul — effectively neck and neck, and these are the settings Theo actually uses (high, occasionally xhigh; never max).

Resets matter a lot: Codex had 52 limit resets between April 16th and October 31st (11 in the last 30 days) versus 14 for Claude (none in the last 30 days). With resets counted, the subscription Pareto line looks much better for OpenAI; without them, mostly Anthropic.

Real-world test: the council skill

Theo's "council" skill sends the same review task to 3–5 model families, giving like-for-like comparisons. Blind judging by both models over 83 reviews:

• Opus preferred Opus reviews 66/83 (80%);
• Soul preferred Opus reviews 54/83 (65%).

Median review cost: Opus $1.65, Soul $0.36. Against expected subscription value (≈ $3.3K Opus on Claude $200, ≈ $1K Soul on ChatGPT Pro $200) that yields about 205 Opus reviews versus 290 Soul reviews per month. A 30–50% volume advantage for Codex, at a clear quality cost — so the Claude subscription is not the poor value it looks like.

Verdict

A good change, overblown complaints, but OpenAI still lacks a model that competes with Opus 5.5. Theo's current setup: OpenAI subs for computer use and verification, Claude subs for day-to-day coding, a Grok sub for Grokbot.