A million for maintainers and 71.5% of tokens on open models: why I like free tools

Sentry gave open source maintainers a million dollars over the year, Vercel counted that open models handle 71.5% of the tokens in its gateway, Anthropic handed the Agent Skills standard to a foundation under the Linux Foundation, NVIDIA open-sourced Run:ai, and Red Hat explained in two articles why companies should take part in communities. For a business owner this is about money: the more work you can do on open tools, the fewer subscriptions in the budget and the less dependence on a single vendor.

What came out

Sentry on October 21 wrote that it raised its annual budget for supporting maintainers from $750,000 to $1,000,000. $495,000 went through thanks.dev to hundreds of maintainers of packages Sentry uses, $100,000 through ecosyste.ms Funds to critical packages, and the rest to foundations, from the Python Software Foundation with $60,000 to Alpine Linux with $5,000. The Open Source Pledge, which Sentry started, turned two in October: 37 member companies have paid out $8,421,535 in total. The Python Software Foundation's AWS costs grew about 69% over the year: agents install packages, and CI runs more often.

Vercel on October 15 published the October index of its AI Gateway. Open-weight models handle more than seven tokens out of ten and take almost one dollar in four; their share of spend has grown more than fourfold since June. Spend through the gateway in September was about 13 times December's, with the average token price 47% lower. DeepSeek-V4.1-Flash handled almost half of all September tokens and took 2.6% of spend.

Simon Willison on October 22 noted Anthropic's news: the Agent Skills standard has gone to the Agentic AI Foundation under the Linux Foundation, where the MCP protocol went earlier. Anthropic founded the foundation together with Block and OpenAI.

NVIDIA on October 13 said it's opening the code of the whole Run:ai platform for orchestrating AI workloads and GPUs, as part of the NVIDIA DSX AI Factory Platform. I only have the announcement, and it has no license terms.

Red Hat on October 23 wrote that taking part in a community gives a company a say in the development of the technologies it depends on, and on October 20 went through the Linux Foundation report “Navigating Open Source Software Development in the Generative AI Era” on how AI can grow the same way operating systems and cloud infrastructure did. From both articles I only have the introductions.

Example: call transcription at a service company

A service company records calls and wants a transcript and a short summary for each one. A cloud service charges per minute, and the bill grows with the number of calls. A local speech recognition model on your own server costs what the server costs, and doesn't depend on the number of minutes.

Open tools pay off where the volume grows and the task repeats. In that sense an open model is no different from Nginx or MariaDB.

How it works for me

I like my own tools and free tools, and two projects show it. In late 2023 and early 2024 I built a subtitle system for bloggers for a consulting company in London: in goes a video from a Google Drive folder or a YouTube link, out comes an SRT file with timestamps. Speech was recognized by a local Whisper model, with no paid transcription services, so there was nothing to pay per minute of recording. The server costs money too, but there was no separate charge for volume. Overlaying the subtitles onto the video was in the plans, but the client stopped the project.

The second example is recent. In September and October 2026 I designed and built the “Coffee Together” service: employees of an IT company order coffee together from the wholesale supplier Tasty Coffee, and the service collects the shared order and fills the supplier's cart itself. It runs on the studio's shared server with 2 GB of memory, in Docker behind Nginx, with a SQLite database and an app in React and TypeScript. The service saves employees and administrators more than 100 hours a month combined. I'm proud when something expensive turns out cheap.

Open doesn't mean nobody's: the PDR company and the farm have servers on Debian, MariaDB is in almost all my projects, and kononov.dev runs on Next.js, Node.js and SQLite in Docker behind Nginx. Sentry is right: companies that make money on other people's work should support it.

What I would do

Count which services the company pays for by volume: transcription minutes, OCR pages, tokens. Where the volume grows, I'd look at an open model on your own server.

I wouldn't drop cloud models altogether. In the parsers for the PDR company I use three external services to validate documents, and I wouldn't swap them for a local model to save money without comparing accuracy.

Pin versions: an open model that was “just updated” changes the result the same way an updated library does.

Check each model's license for commercial use before it lands on a client's server. And keep a line in the budget for supporting what you use: at least for one or two projects without which the system would stop.

Related reading

On models in document parsing and checking the result: the article AI in data processing: extraction and verification. Services: AI in a process where the result can be verified and Python for documents, data and repeatable processing. Projects: SRT subtitles from video on a local Whisper and the “Coffee Together” service. Nearby are my reviews on routing between your own model and an API and on speech transcription on your own server, and earlier I wrote about why to keep the model catalog and prices separate from the app release. If you pay for an AI service by volume and want to see what could move to your own server, tell me which tasks and what volume. Discuss your task.