ABAB News Logo
Back to Flash News

SemiAnalysis: The Value of Claude Package is Approximately Five Times That of OpenAI

SemiAnalysis used its own Tokenomics model to measure the subscription values of various models, stating that Anthropic's $200 Claude package offers an API equivalent value approximately five times that of OpenAI's comparable tier. The authors are Andrew Megalaa, Max Kan, and Dylan Patel.

The comparison is between Opus 5.5 and GPT 6.1 Sol. The report notes that even without considering prices, the difference in token count is still significant, as Sol is cheaper per token, which does not account for the entire gap. At the top tier, GPT-6 Astra and Fable 5.1 have closer limits, but Fable can only utilize up to 50% of its package limit.

According to this rule, after consuming $2,485 equivalent value of Fable 5.1, the $200 Claude still has 50% of its limit available for other models. In contrast, the equivalent OpenAI package, after using $2,897 equivalent value of Astra, has its limit reduced to zero. The extra half is magnified in the comparison of Opus 5.5 to Sol.

Last week, OpenAI halved the API equivalent value of the $200 tier and introduced a new $500 tier. SemiAnalysis's tracking confirmed this reduction. The old $200 tier purchased before October 29 maintains its original limit, while new purchases immediately take effect at the lower limit. The token count for each model tier has been halved. The equivalent value of the Sol tier has decreased by over 50%, as the input price for 6.1 Sol has also been lowered, without a corresponding adjustment in subscription limits.

The new $500 tier only offers 21% more than the old $200 tier for Astra. Due to the price drop of Sol, the API equivalent value for this tier has actually decreased. The $500 tier features Ultrafast at 300 tokens per second, with the report stating that limits are still being tested. After the cuts, the number of tokens per dollar for the $100, $200, and $500 tiers has been equalized. Plus is close on Sol but performs worse on Astra.

Anthropic's tiers were originally designed to have equal value per dollar, and they remain consistently higher than OpenAI. After reducing input and output prices by 20% and cache reads by 60%, Anthropic increased the Max tier Opus limit by about 20% and the Pro tier by about 50%. Neither company has fully passed the API price cuts onto subscription users.

The buyers are heavy users paying monthly, while the sellers are two labs exchanging inference power for retention. This is a repricing of limits, not a new model release. OpenAI has retracted subscription subsidies, while Anthropic continues to keep more computing power in its packages. The report estimates that subscriptions account for about 10% of Anthropic's revenue but may consume over 40% of inference power, lowering the blended revenue per megawatt by about $36 million. Users fully utilizing mid-tier models benefit from Claude, while new users purchasing OpenAI's high tiers face pressure.

Source: Public Information

ABAB AI Insight

SemiAnalysis's method does not look at official prices but instead empties the package and then back-calculates based on API prices. In June 2026, they found that the equivalent value of the $200 tier was much higher than market speculation. This time, they captured OpenAI's price change last week: the old $200 tier is retained until October 29, while new purchases are immediately halved, effectively separating existing users from new users at the expiration date.

Anthropic's subsidies are reflected in the computing power accounts. Subscriptions contribute only about 10% of revenue but may consume over 40% of inference power, lowering revenue per megawatt by about $36 million. OpenAI has equalized the token count for $100, $200, and $500 tiers, effectively eliminating the premium for high tiers. After the price drop for Opus, Anthropic still increased the Max limit by about 20% and the Pro limit by about 50%, continuing to use limits rather than tier differences to retain users.

The approach is similar to telecom operators reverting unlimited data back to fair use, and also resembles cloud vendors locking usage with reserved instances and using on-demand pricing to protect margins. The difference is that the "traffic" for model subscriptions is tokens, with the top tier still able to set a 50% cap separately. The industry is transitioning from customer acquisition subsidies to control phases; whoever first recoups package value will restore inference margins.

Essentially, this is a transfer of pricing power. After the API price drop, subscriptions did not proportionally increase, so the labs kept the price difference for themselves. OpenAI has redefined tiers with the halving and the new $500 tier, while Anthropic maintains a fivefold gap using remaining limits. The value for users seen in cost-effectiveness is actually how much inference power both companies are willing to offer for free in monthly fees.

ABAB News · Cognitive Law

  1. The monthly fee remains unchanged; the limit is the true price.

  2. The top configurations are close, while the everyday tiers widen the gap.

  3. The cheaper part of the subscription is subsidized by computing power.