$200 OpenAI Codex Plan Value Plummets

Just weeks ago, a $200 plan on a codec subscription would have provided up to $12,000 a month in inference, a figure that now seems significantly out of reach.

The original intent behind these plans was to offer what felt like unlimited usage, providing substantial value to developers.

However, as AI models have grown more capable and, consequently, more expensive, companies have been compelled to reduce their subsidization efforts.

OpenAI recently halved the usage calculation for its $200 plan, leading to a measured usage drop to about $570 per week, equating to roughly $2,000 a month.

Both the $200 subscription credit and the API costs for AI services are arbitrary figures, according to Theo, and do not represent the actual operational costs to the companies running these large language models.

The host observed that his personal API usage would cost an unsustainable $20,000 in just seven days if priced at the current API rates.

API prices are primarily set by companies to explore the maximum amount enterprises are willing to pay for these advanced capabilities, with historical profit margins estimated to be upwards of 95% relative to the underlying hardware and energy expenditures.

The actual energy and compute costs for consuming $100 worth of tokens are closer to $2 to $5, a stark contrast to the premium price charged for equivalent usage.

This massive disparity is maintained by exceptionally high profit margins, which have been a significant factor in the pricing strategies of AI providers.

Increased competition between major players like Anthropic and OpenAI is expected to exert downward pressure on these margins, potentially leading to a collapse in the current price gap.

Anthropic's Claude subscription is heavily restricted to usage within its own Claude Code environment, limiting its integration with external tools.

Users cannot easily redirect Claude API access to third-party platforms such as Pi, Open Claw, or Hermes, which impacts developer flexibility.

OpenAI, in contrast, has historically allowed broader usage of its Codex subscriptions for developer workflows, provided that the usage does not involve serving external user traffic.

Claude imposes specific usage restrictions on its largest models, such as Fable, to manage operational costs.

These limits are in place because the pricing of the Fable model does not align with Anthropic's originally planned business model, making unrestricted usage financially unfeasible.

This situation highlights a clear division between smaller, more affordable models and larger, more expensive ones, as illustrated by comparing models like Opus 4.5/5.5 to Fable 5/5.1.

To pressure OpenAI, Anthropic launched Fable 5 and Mythos 5 at $50 per million tokens, far below the $125 initially projected for Project Glasswing to maintain a 95% margin.

Project Glasswing (Claude Mythos) was initially projected at $125 per million tokens to maintain a 95% profit margin.

However, Fable 5 and Mythos 5 were released at $50 per million tokens, less than half the initial preview price.

This aggressive pricing caused profit margins to drop from 95% to 87.5%, a strategic move aimed at pressuring OpenAI and signaling a potential end to heavy subscription-based subsidization in the market.

Despite this reduction in revenue efficiency, the company continues to maintain high profit margins, suggesting that the initial margins were exceptionally large.

Fable model usage is strictly limited to 50% of a user's weekly subscription allocation, a restriction that directly correlates with a significant reduction in profit per token compared to previous models.

OpenAI released its Astra model at the same price point as Anthropic's Fable, indicating a direct response to competitive pressures in the market.

Unlike Anthropic, OpenAI chose not to impose usage restrictions on Astra, providing a more flexible option for users.

It is possible that OpenAI had initially planned for higher pricing for Astra but adjusted its strategy to align with market trends and maintain competitiveness.

The 6.1-Soul model, which was released by OpenAI, appears to be an unpredicted offering, potentially intended to replace less-utilized previous models.

OpenAI's internal target price for this model was likely $5 per million tokens for input and $25 per million tokens for output.

However, the final release price for 6.1-Soul was significantly lower, set at $2 per million tokens for input and $10 per million tokens for output.

Cash read costs saw a massive decrease, dropping from $0.20 to $0.10, which indicates a substantial reduction in infrastructure overhead.

The speaker estimates that OpenAI's profit margins on this specific model have consequently fallen from approximately 95% to around 50%.

The gap in subsidization between subscription accounts and API usage costs drove OpenAI to avoid arbitrary limits on Astra, an announcement that was nonetheless poorly received.

The company aimed to reduce the disparity in subsidization between its subscription accounts and direct API usage costs, seeking a more balanced pricing structure.

While OpenAI's commitment to transparency is acknowledged, the announcement regarding these significant shifts in pricing and value was largely met with negative feedback from the user base.

The monthly usage value of the $200 plan dropped from approximately $24,000 to $2,500 due to new limits, representing a significant reduction in effective token budget.

The host experienced a roughly tenfold decrease in effective token budget, observing only $573.86 of usage and noting the ability to deplete an entire account in 29 hours through continuous looping.

OpenAI implemented these changes to prevent financial instability caused by heavy users who maximized their token usage, a practice that was becoming unsustainable for the company.

Despite these changes, the $200 plan still offers practically unlimited usage for real-world scenarios, provided users avoid resource-intensive 'Max' models.

Developers creating production-ready code find the $200 Claude plan's Opus 3.5 offering to be of unparalleled value, according to Theo.

Despite the implementation of usage limits, the host continues to favor Claude for coding tasks intended for deployment, citing its superior performance in this domain.

Users are currently obtaining more value than anticipated from their Claude subscriptions by consistently pushing the usage limits up to their weekly reset points, effectively maximizing their investment.

Theo emphasizes that for those whose primary objective is to leverage their subscription for shipping code, the benefits derived from Opus 3.5 are exceptionally high and difficult to surpass by other offerings in the market.

To maximize the utility of their subscription, users should strategically avoid the 'Astra' model, which incurs significantly higher costs per task.

The apparent decrease in dollar value is mitigated by the substantially lower token cost per task of newer models, meaning users can still achieve a similar amount of work for their investment.

Localization is a notoriously difficult engineering problem, involving complex challenges such as translating content across various interfaces and ensuring cultural relevance.

General Translation addresses these issues by providing React helpers for JSX translation, number formatting, and pluralization, alongside centralized dictionaries that maintain consistency across mobile, web, and desktop applications.

Theo successfully utilized General Translation for the T3 Code marketing site, leveraging agent-assisted prompts to streamline the localization process.

The service aims to make it easier for developers to translate real-world applications across different platforms and reach a broader global audience.

Theo criticized OpenAI's decision to reduce the '20x' plan to '10x' while simultaneously introducing a '25x' plan for 2.5 times the cost, calling it a 'self-inflicted L' and a significant marketing blunder.

He argued that this move alienated users, particularly given that subscription plans constitute only a small fraction of OpenAI's total revenue, which primarily comes from consumer plans and API access.

Operating on distinct, expensive infrastructure via specialized Cerebras chips, Ultra Fast debuted alongside the highly efficient Pareto Frontier models.

The Ultra Fast mode reportedly relies on specialized chips from Cerebras, leading to high operational costs for the company.

Theo argues that combining these releases with a simultaneous reduction in subscription value points to a disorganized and potentially contradictory product strategy from OpenAI.

These constraints likely forced the company to engage in massive overprovisioning of resources to achieve the advertised speeds and performance for the Ultra Fast mode.

Lower profit margins on Ultrafast operations, combined with the expensive overprovisioning, make the current service model unsustainable for OpenAI.

Rumors circulating in the industry suggest that Jane Street may have acquired the available supply of Cerebras chips, further complicating OpenAI's hardware strategy.

Theo's $500 plan now provides only about $4,000 per month in value, a stark contrast to his $200 plan which previously offered $12,000 per month just a few weeks prior.

This change represents a 2.5 times increase in price alongside a three times decrease in total usage capacity, significantly diminishing the return on investment for subscribers.

OpenAI has lost its competitive edge in post-training and reinforcement learning, as Anthropic's Sonnet and Opus models have effectively closed the capability gap, offering comparable or superior performance.

The current subscription models no longer provide the 'limitless' feeling that attracted many users, leading to a perception of poor value among the user base.