The decision of big players to cancel unlimited achievements prompted Ukrainian software company DevForth to conduct an analysis of which AI agents are the best in terms of quality and limits. The results of the study were unexpected: some subscriptions provide tens of times more value than they cost, while the market leader has become frankly unprofitable.
The decision of big players to cancel unlimited achievements prompted Ukrainian software company DevForth to conduct an analysis of which AI agents are the best in terms of quality and limits. The results of the study were unexpected: some subscriptions provide tens of times more value than they cost, while the market leader has become frankly unprofitable.
In practice, most tariff plans function differently than they seem at first glance, according to the DevForth team.
When providers like OpenAI or Anthropic charge a fixed monthly fee (e.g. $20), on the backend they actually set a consumption limit in percentages for a certain time window. The most interesting thing is the mechanism of how each request increases these percentages.
"Our hypothesis was the most obvious: a certain amount in virtual credits or dollars is set in the backend window (of course, different for different plans). And the actually consumed tokens and their official prices "inflate" the percentage after each request in proportion to the provider's official prices for the API of this model," DevForth founder Ivan Borshchov said on the Reddit platform.
For the analysis, the company took a 7-day time window, where it is very easy to track the exact start and finish times in UTC. The team used one or two models and gradually “inflated” the window limit to an adequate percentage without unnecessary noise — say, to 43%. Then it calculated the net value of tokens in this window, summing up the inputs, outputs, and input cache at the official price of each model. Next, the company applied the usual proportion: if $N worth of tokens were burned for 43%, then this is the maximum amount of code a user can generate or feed neural networks for all 100% of the limit.
"And the most delicious thing is that if you do the experiments carefully (there are certain peculiarities), this proportion works for different people at different percentages and gives a stable value in the end, which we indicate in the table," Borshchov added.
For the DevForth team, it was a real discovery that the tariff plans of different providers offer radically different benefits.
In particular, experts found that a $20 per month Personal Codex Plus subscription provides approximately $130 in monthly limits, while the more expensive $25 per month Business Codex plan limits the user to only the same $25.
At the same time, the Claude Code tool demonstrates equally high limits — the equivalent of $1,200 in tokens on both the Team Standard seat ($25/month) and the personal plan for $20 per month.
On the other hand, the popular Copilot Pro disappointed the team after the June update, because for a subscription cost of $10 per month, it provides a volume of tokens for only about $15, which is why the tool actually lost its economic meaning.
Against this background, the $20 AI Pro plan turned out to be quite balanced, providing approximately $171 worth of tokens on Gemini models and additionally offering a bonus $40 for working with several third-party models.
“Claude still gives the most, but Codex still feels higher quality, and its coding benchmarks are a little better,” concluded the DevForth founder.
You can find more detailed information about the analysis results at the link .
Нажаль пару місяців назад Cursor прибрав безліміт у режимі авто. І за 20$ ліміт вичерпався дуже швидко. І таке відчуття що токени дійсно стали дорожчати для усіх та й одразу …. Повністю перевести роботу через віртуальних асистентів буде складно це буде оверпрайс. На зараз людина для бізнесу буде вигідніша ніж суцільна його АІ версія.
Після відправки коменту я би вам радив передивитись куди редиректить.