The Codex $200 plan still delivers real value, but the subsidized dollar value it represents dropped sharply after OpenAI recalculated how usage counts against your weekly limit. Theo of t3.gg reports roughly $570 of API-priced usage in a week, down from about $12,000 a month before the change, and this article walks through what that means.
What changed on the Codex $200 plan
The Codex $200 plan from OpenAI buys roughly half the subsidized dollar value it did before early October 2026. Theo of t3.gg, who publishes developer-focused analysis on t3.gg, reported in his video on 2026-10-07 that a week of heavy use produced $573.86 of API-priced usage, which he extrapolates to about $2,000 to $2,500 per month. Before the change he measured about $12,000 per month, and up to $24,000 when resets aligned well. Those figures are the speaker's own measurements from his accounts, not OpenAI-published numbers.
The change landed right before OpenAI's Dev Day. Theo credits the announcement to Tibo, and frames it as OpenAI reducing the implied subsidy rather than raising the sticker price. Nothing about the plan's list price moved; only the usage calculation did.
Two caveats matter. First, all dollar values here are priced at list API rates, which Theo argues bear little relation to actual compute cost. Second, the measured numbers come from one extremely heavy user who deliberately maxes accounts in loops, so a lighter developer will not see the same absolute figures, only the same reduced ratio.
Why OpenAI cut the subsidy
OpenAI cut the subsidy because Anthropic forced a price war it could not fund at the old margins. Theo's argument, built from Anthropic public pricing for its newest model tier, runs like this: Anthropic launched its flagship model at $10 per million input tokens and $50 per million output, well under the $25 per million it had reportedly offered to enterprise preview participants. Same model class, bigger capability, lower price.
Theo calculates that this took Anthropic gross margin on that model from roughly 95% to roughly 87.5%, and cut revenue per unit of compute by more than half. He then argues OpenAI matched the price with its own frontier model the same week, and went further on its mid-tier model, cutting cached-input pricing from $0.20 to $0.10 per million. His estimate is that this model now runs at margins closer to 50% than the industry's historical 90-plus percent.
The mechanism for the subscription cut follows from that math. Anthropic caps its flagship model at half your subscription allocation, because it subsidizes that model at a loss relative to older margins. OpenAI chose instead to shrink the overall dollar value of the subscription rather than single out one model with a custom limit. Both companies, in Theo's framing, are pulling subsidized value back toward sustainable levels.
These margin figures are Theo's derived calculations from public list prices, not audited company disclosures. Treat them as informed inference, labeled as such in the video itself.
Codex $200 plan vs Claude: the numbers side by side
The Claude plan from Anthropic still buys more raw subsidized dollars than the Codex $200 plan, but the Codex plan's cheap mid-tier model narrows the gap in cost per completed task. Theo reports running six Claude accounts to their limits and still considers Claude the better value for shipping serious engineering work, while calling OpenAI's cheaper model the best value per task he has measured.
Note that Claude subscription is contractually tied to Claude Code, Anthropic terminal coding agent. Theo complains that the usage cannot legally power third-party tools, while OpenAI's terms allow Codex usage in CI and external harnesses as long as you are not reselling the subscription itself. That flexibility difference is real and documented in each vendor's own terms, and it is a decision factor independent of price.
Here is the comparison as Theo reports it, with his own measurements and public list prices:
| Dimension | Codex $200 plan | Claude $200 plan |
|---|---|---|
| Reported subsidized value | ~$2,000–$2,500/month after the change | ~$8,000/month before its own tightening |
| Flagship model access | Full weekly limit, no per-model cap | Flagship limited to half your allocation |
| Where you can use it | CI, external tools, personal projects | Effectively Claude Code only |
| Reported cost per task, best mid-tier model | ~$0.96 | ~$13.11 on the comparable flagship run |
The last row is Theo's own Terminal Bench run, an independent evaluation harness maintained by the Claude Institute. He ran the models through both Codex and Claude Code harnesses himself, so these are speaker-reported numbers, not vendor benchmarks.
Why the cheap model changes the value math
A halved subsidy does not halve your output when the model under it also got much cheaper. Theo's week of real usage makes the point with two numbers: about 1,800 responses on the flagship cost $534 at API rates, while roughly 1,500 responses on the new mid-tier model cost $38.27. Same kind of work, an order-of-magnitude difference in list-price value consumed.
His own Terminal Bench runs put the mid-tier model's high-effort setting roughly level with the flagship on score, at about a thirteenth of the cost per task. Extrapolating, he argues the halved dollar subsidy still buys roughly five times more tasks than before for workloads that fit the cheaper model.
His benchmark-agnostic evidence comes from the Artificial Analysis Intelligence Index, which he visualizes in a project he built for exactly this comparison. On that chart the cheap model sits so deep on the cost-capability Pareto line that the flagship's low and medium settings are not worth using by comparison.
The honest summary: the flagship on Codex is now the wrong default, the cheap model is the best value per task Theo has measured, and Claude flagship remains the better raw deal if you insist on running a frontier model all day.
Ultrafast mode and the $500 tier problem
Ultrafast mode on the $500 plan is where Theo stops giving OpenAI benefit of the doubt. The mode runs on different inference infrastructure, which he associates with Cerebras hardware, and it carries sharply higher internal pricing: he cites $60 per million input and $300 per million output at standard context, rising to $120 and $450 with extended context. Those are effective API-equivalent rates he says the mode consumes.
His measurement: the $500 plan funds roughly $1,000 of Ultrafast API value per week, which his usage patterns exhaust in about 2.1 hours of generation. One PR review session on Ultrafast reportedly consumed $600 of API-equivalent value. On those numbers, Ultrafast is a fast way to burn a weekly limit, and branding the $500 tier around it is the decision he calls OpenAI's biggest self-inflicted mistake.
Theo also argues the $500 plan should not have launched simultaneously with the subsidy cut, the cheap model, and Ultrafast. His claim is that these subscriptions are marketing rather than profit centers, so fumbling the optics damages the one thing they exist to do. That is his interpretation of OpenAI's strategy, labeled as such, not a documented company position.
Which plan should you actually buy
Buy the Claude plan if your goal is shipping production code with a frontier model all day; buy the Codex plan if you want cheap, flexible capacity across tools and CI. Theo's own setup uses both: Claude drives the work, and Codex runs review sub-agents because he finds OpenAI's analysis catches issues Anthropic models miss in his workflow. He also uses Codex for email and computer-use tasks.
His concrete guidance from the video:
- If you will only use the flagship model, get the Claude plan. The Codex flagship no longer justifies its subsidy level.
- If you work through external tools or CI, get Codex. Its terms allow it; Claude effectively do not.
- If you pay raw API prices, the cheap OpenAI mid-tier model is the strongest value per task he has measured.
- Skip the $500 tier unless you specifically need Ultrafast latency and accept a weekly limit measured in hours.
One forward-looking caveat from Theo: as enterprise API prices fall under competition, subscription subsidies will likely shrink further. He predicts the graphs will look worse every year while the actual work completed per plan keeps rising, because model efficiency has so far outrun the subsidy cuts.
FAQ
- Does the Codex $200 plan still get you $12,000 of usage? No. Theo's published measurements after the early October 2026 change show roughly $570 of API-priced value per heavy week, extrapolating to $2,000–$2,500 per month. The $12,000 figure dates from a few weeks before the change.
- Is Claude still the better value? For raw subsidized dollars and flagship-model access, yes, per Theo's usage. For cost per completed task on work that fits OpenAI's cheap mid-tier model, his Terminal Bench runs favor Codex by a wide margin.
- What is Ultrafast mode? A latency-optimized mode on the $500 Codex tier that runs on separate infrastructure. Theo reports it consumes about $1,000 of API-equivalent value per week, which his usage exhausts in roughly two hours.
- Can you use a Claude subscription outside Claude Code? Effectively no. Theo reports that third-party tools cannot use the subscription's usage, while Codex permits CI and external harness use as long as you are not reselling access.
- Are the dollar figures official? No. Every usage figure in this article is the video author's own measurement or derived calculation from public API list prices, not numbers published by OpenAI or Anthropic.
From a video transcript to an article like this one
Everything above came out of a single 34-minute video: the usage measurements, the margin reasoning, the comparison table, the caveats. That is exactly the kind of knowledge that stays locked inside a video while the people who need it are searching for text.
If you publish analysis like this on YouTube, Skalablog turns the video into a structured, source-linked article: paste the YouTube URL, it transcribes the talk, and you review a drafted article before it goes live. A transcript shared by Gustavo Dev Doido works the same way as any other, and the result reads like the piece you just finished.
And if you want to sharpen the engineering skills behind videos like this one, CrazyStack Typescript is a practical place to build them.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits