Param Kumar
&
October 19, 2026
We believe the most valuable AI in any company will run on open-weight models that the company has fine-tuned on its own data and owns outright. Frontier APIs have done an enormous amount to show businesses what language models can do. But a general-purpose model rented by the token is the wrong long-term foundation for the work that matters most to an enterprise. That work is specific, repetitive and high volume, and it is exactly where a smaller, specialized model trained on a company's own data will outperform a frontier model on both quality and cost.
Most enterprise teams we speak to already sense this. They run a task thousands of times a week, such as categorizing claims, routing support tickets or reviewing documents, and they pay a frontier API to do it. Many are confident that an open model fine-tuned on their own records would do that job better for a fraction of the price. What holds them back is the upfront cost of finding out, along with the concern that a better model will be released next month.
There is also a second group we meet just as often. These teams have modeled what a frontier API would cost at their volume and concluded that it would never pay for itself, so they haven't adopted AI for those processes at all. Some of the most valuable automation in their business is still handled manually or by rules engines written years ago. The models aren't the problem. Renting general-purpose intelligence one token at a time simply doesn't make economic sense for high-volume work.
Today we're launching the Arena Credits Program to remove that barrier. Qualifying companies can apply for $20,000 in Arena credits to fine-tune an open-weight model on their own task. The credits cover training and inference for 30 days from the date they are granted. Each company receives one grant, which can be used across as many teams and use cases as needed.
The gap between open and closed models has narrowed significantly in the last few months, and a capable open base model fine-tuned on a company's own data can already outperform the leading frontier APIs on the task it was trained for. We saw this directly when we partnered with NVIDIA and trained the Nemotron Lightning 3.5 model. We fine-tuned it on Arena using a standard customer support dataset, where the model has to classify each message into one of 27 intents and write the reply. Before training, Nemotron classified 71% of messages correctly, and Claude Sonnet 5 reached 90%. After fine-tuning, Nemotron reached 99%, and its replies matched the reference responses more than twice as closely as Sonnet's. That is a model with 3B active parameters outperforming one of the strongest frontier models available on a real business task.
The reason is straightforward. Every business runs on its own conventions: its intent categories, product names, escalation policies and the tone its customers expect. None of that appears in any model's pretraining data. A prompt can describe those conventions, but fine-tuning is what teaches a model to follow them reliably. A specialized model will continue to outperform a general one on the task it is measured against, because it is optimized for that task and nothing else.
The economics also strengthen the case for fine-tuned open-weight models. Frontier APIs charge per token, so costs rise with every ticket, claim and document processed, which is why many large enterprises are already capping their token spend. A fine-tuned open model runs on infrastructure the company controls, at a cost that does not scale with volume. Smaller models are also faster and cheaper to serve, so the advantage grows with usage. For the companies that ruled out AI on cost grounds, open-weight models are the solution.
Owning the weights brings benefits that go beyond price:
Credits can be used across Arena to fine-tune the latest open-weight models, evaluate them and deploy them, through either the web app or the CLI. Our team works alongside each company at no cost throughout the process. Most teams find the hardest parts are preparing the task properly and measuring the result fairly against their current approach. We run that comparison with each team on its own data, whether the current setup is a frontier API, a self-hosted model, a rules engine or a manual process.
We are selecting participants carefully as we want each engagement to lead somewhere meaningful. The program is designed for ML and engineering teams at established businesses that have:
We review every application, and places are limited.
When the fine-tuned model outperforms the current approach, as we expect it to, we will help take it into production and keep improving it. Applications are open now, and the full program terms are available on our website.