Posit AI is priced for the long run

In March 2026, we released Posit AI, a subscription service to power Posit Assistant and Next Edit Suggestions that starts at $20/month.

In the last few years, many model providers have kept token prices low by subsidizing them. They are accepting a trade-off: losing money in the short-term in exchange for growing their user base. We do not subsidize tokens with Posit AI. Instead, our intention has always been to build a sustainable business from the start. This has meant that, in early 2026, it was impossible for us to compete on price with major model providers.

However, this is starting to change. The inexpensive token party was always going to end, as eventually model providers need to make a profit. We've already started to see this transition take effect, as Microsoft's Copilot moves to usage-based billing and OpenAI and Anthropic clamp down on rate limits in their subscription plans.

Posit Assistant and Next Edit Suggestions reflect our learnings from twenty years of building for data scientists and researchers. As the token subsidies end, the differentiator between products becomes more about the value those products provide, and we think we’re ahead in value for the data community.

As competitors raise prices, we believe we’ll be at parity on price, but with a superior product for data science tasks. 

Where a Posit AI subscription goes

Posit AI costs $20 a month. $15 of that subscription charge is allotted to API credits for Posit Assistant. Right now, the sole model provider in Posit Assistant is Anthropic's Claude, and we pay the publicly advertised rates per token. For a majority of our current users, this $15 of credits covers their full month of usage.

Posit Assistant makes efficient use of those $15 in API credits. We've built Posit Assistant with careful context management in mind from the get-go. We monitor our cache hit rate closely, maintaining >90% cache efficiency.¹

Independent of costs, context management is also important from a quality perspective; we want to surface only the context that we believe the agent should be focusing on.

We are also exploring models that are smaller and cheaper than our current cheapest offering, Claude Haiku 4.5. We have been impressed with some of the recent open-weights models and are taking the prospect of using them for Posit Assistant seriously. These models would be even more inexpensive per-token than Haiku 4.5, allowing users to get more mileage out of their Posit AI subscription.

The remaining $5 of the $20 charge, beyond supporting our work on the service, goes to a number of fixed costs. The most significant one is GPU hours for the compute that serves Next Edit Suggestions (NES). For NES to feel usable, the latency needs snappy enough for suggestions to feel instant—we loosely target 150ms-200ms. To meet that latency target, we need north of 1,000 tokens per second.² There is no model provider that will sell us tokens delivered that quickly. Instead, we rent H100s and serve our own customized deployment on the GPU. Whether we support 100 NES users or 10,000, we still have to pay for the GPU, making this a fixed cost. 

Posit Assistant is very good

Some subset of our users are transitioning to Posit Assistant from AI chat interfaces, where they've mostly relied on copying and pasting code and results back and forth to code. Another subset has spent time with coding agents like Claude Code and Codex. Posit Assistant is a qualitatively different experience than either of these tool kits.

Most importantly, the agent can see (and run code in) your active R/Python session. This immediately cuts out the copy-and-paste exchange many of our users have grown accustomed to, allowing the agent to retrieve the information it needs much more quickly. This also means that the user and agent share a computational session as ground truth, allowing both to see eye-to-eye.

Posit Assistant's UI is also designed to be auditable. When the agent runs code and sees its output, the same code and output is visible to you. This is especially important when results include plots; our benchmarking has shown that even the most advanced models sometimes disregard evidence in plots that contradicts their expectations. While other coding agents may be able to run R code through the Rscript executable (or an MCP server), those interfaces do not provide an easy entry point for the user to see the same plots that the model is seeing.

Finally, Posit Assistant has access to comprehensive prompting, skills, and tools that promote fundamentally sound data analysis. LLMs can sometimes enact a performance of "excelling" at data analysis, a behavior in tension with the realities of real-world data science, an experience dominated by sitting with ambiguity. For example, in one piece of prompting:

While there's no single "Posit Assistant prompt", you can check out the prompting your agent has access to as you please; we tell the agent that it's free to share its prompting.

Taken together, these features make Posit Assistant a very capable data analysis assistant.

Up to this point, it's been difficult for Posit AI to compete on cost with subscription offerings from large model providers. The subsidization that drove those cost differences is quickly coming to an end, though. Once everyone is paying close to the real cost of tokens, the choice between tools is really a choice between products. Posit Assistant and Next Edit Suggestions reflect our extensive experience building tools for data scientists and researchers, and we think that experience shows in how they handle real data work.
 

¹Cache reads are ~90% cheaper than "normal" input tokens. Cache efficiency measures the share of cacheable tokens that actually hit the cache. 

² For comparison, at the time of writing, Anthropic is serving their smallest and fastest model, Haiku 4.5, at 66 tokens per second.