• 6 min čtení
GitHub Copilot moved to token-based billing, and many people were surprised by how much their usage increased. The $10 or $39 Copilot subscriptions that used to be sufficient for many developers now seem to run out within a single week.
Mr. RIGANTI uses a different strategy - instead of per-user subscriptions, we charge for actual token usage. We plan to offer a subscription model as well and add an option to bring your own model, but they will take some time to implement.
At first glance, charging based on actual token usage looks like the most expensive option, but let's look at it more closely.
AI usage between team members differs
Some developers can run out of tokens very quickly, and even the most expensive subscriptions are not enough for them. Whether they use AI heavily to deal with extremely complex tasks, or they just don't give ideal instructions and the model eats more tokens than necessary, doesn't matter - the tokens are burned, and sometimes the developers have to wait for hours or days before their quota renews. This creates an additional inefficiency and can be way more expensive than the tokens.
On the other hand, there are developers who cannot use AI to such an extent, because the nature of their work is different. Agentic AI is great if you need to write hundreds of lines of new code or make changes to the codebase. However, if your job is to debug or trace how data flows through your system, and you happen to only write a few lines of code per day (or even delete some), the added value of AI is not that interesting yet. But you still pay a subscription you may not even fully utilize.
If you have both groups of people in the team, paying for actual token usage may be the most efficient way - AI power-users will not be blocked by limits of their subscriptions, and you will not regret paying subscriptions that are not fully utilized.
Subscriptions will be more expensive
GitHub Copilot was the first to raise prices, but it is well known that all model providers charge less than their actual costs, and current prices are not sustainable in the long term. We have to prepare for either switching to on-premises models or higher subscription prices.
We believe that eventually, the subscription prices will match token prices. As you will see in the last section, even the actual token costs are not that bad, looking at the productivity boost AI gives us.
Precise task specification saves tokens
In our experience, there is a substantial difference in AI performance across tasks with good and bad specifications. The difficulty is that models change very quickly, and there is no way how to define "good specification". However, if the model needs to look up information in the codebase, or even write scripts to extract project knowledge, it will likely consume more tokens than when that information is provided in context.
At RIGANTI, we settled on treating the AI agent as a colleague in a distant time zone. We always try to give them all the information they need to do the job, hint at conventions or practices ("do it as this other feature here"), and try to answer all potential questions about things that are not clear from the code or documentation. The more we constrain the model so it doesn't make important decisions, the better.
However, it takes time to develop the intuition for what works and what does not, and you need to be prepared to reevaluate these "rule of thumb" habits at any time.
Tokens are still cheaper than our time
Even though sometimes we assign a task to find out it consumed $20 worth of tokens, when estimating how long a human would take to do the same thing, it is still cheaper by multiples.
Recently, we introduced Tailwind UI for DotVVM, a new package of modern UI components for our open-source web framework DotVVM. Most of this library was implemented by AI. In the past, we would need over 1000 hours to make such a library. With AI, we got to about 300 hours, including writing actual code, tests, samples, and documentation, plus human verification, which accounted for roughly 60% of that total. The token costs were negligible compared to the time we needed to ensure we were shipping the right thing.
The cost of migration to GitHub
Another argument for trying Mr. RIGANTI is that you can stay on Azure DevOps instead of migrating to GitHub. With 50+ projects developed in Azure DevOps, we would need hundreds of hours to migrate everything to GitHub, and we disliked the idea of having the projects scattered across both services (and paying for both of them).
Our statistics on Mr. RIGANTI's token costs
To give you a better picture of the actual costs, we gathered two months of data from all projects that use Mr. RIGANTI.
It is important to note that we have several projects that use different tooling, as we depend on our customers' infrastructure. For implementing work items, Mr. RIGANTI is not the only option; many developers prefer assigning tasks to agents such as GitHub Copilot on their local machines.
Code review
In May and June, we used Mr. RIGANTI to review 622 pull requests. The average cost of code review was $0.56, with a range of $0.05 to $3.94.
An interesting fact is that code reviews by GPT 5.3 and GPT 5.4 were cheaper ($0.41), and based on internal feedback, the results were better than from Claude Sonnet 4.5 / 4.6 in terms of quality of suggestions and consistency of results (we used the same prompt, but it depends on the orchestration in Codex vs OpenCode and the model itself). However, this is more of an anecdotal evidence than a serious evaluation.
Task implementation & refinement
Mr. RIGANTI was used to implement 194 work items. The cost range for implementing entire work items is larger, as it depends heavily on the complexity and quality of the task description.
From our data, the costs range from $0.05 (changing contact info on a static website) to $23.13 (implementing a complex feature that affects multiple applications and spans all layers, with ~40 files changed), with an average of $1.93.
As you can see from the picture, 90% of the work items cost less than $4; the expensive features were more of an exception.
When the task is finished, we usually add several follow-up comments to make additional changes. There have been 299 comments asking Mr. RIGANTI to make edits to the pull request, with an average cost of $0.79. The range is $0.02 (a few single-line fixes) to $7.10 (rewriting a complex Marten query to raw SQL), with 90% of comments under $1.30.
Summary
During the observed period, we used Mr. RIGANTI on 24 projects, and the total token cost was slightly below $1,000. As mentioned above, it is not the only tool we use - most of our developers have additional $25 - $50 GitHub Copilot, OpenAI, or Anthropic subscriptions, and use them on their local machines.