Shawn Wang Says Ripple-65B Cost Just $32M To Train

Shawn Wang said Ripple-65B, a new tool for writing code, was trained for $32 million, which he claimed was one-tenth the cost of earlier top open coding models.

Key Takeaways

• Shawn Wang said Ripple-65B was trained for $32 million, claiming one-tenth the cost of prior open-weight agentic coding SOTA
• Agentic coding systems inspect codebases, plan steps, write files, test changes and loop through fixes, which require longer context and repeated tool use

Writer and investor Shawn Wang, known online as swyx, posted the figure on X.

Still, the number itself is the news because training cost has become one of the sharpest battlegrounds in open AI.

Ripple-65B Puts Cost In The Front Seat

Open models are increasingly judged on two curves at once. One is benchmark quality.

The other is how much money and compute it took to get there.

That is why a $32 million claim matters. If true, it suggests teams chasing coding performance may not need frontier-lab budgets to stay competitive, at least for specialized model classes.

Agentic coding refers to AI systems that do more than autocomplete a line.

They can inspect a codebase, plan steps, write files, test changes and loop through fixes. Those tasks are expensive because they need longer context windows, heavier verification and repeated tool use.

Also Read: AI Lab Nanbeige 4.3 Drops 33B Model As Test Lead Shakes Rivals

The Open Model Fight Has Shifted Since Early 2026

Earlier in the year, many open model launches stressed parameter counts, context length or benchmark wins.

More recent launches are leaning harder on cost efficiency, because enterprises care about what it takes to train and serve a model, not just how it scores in a lab table.

That shift has also changed what counts as a meaningful release. A model that matches a rival while using less capital can influence funding, cloud pricing and open-source adoption even before outside evaluations are complete.

The same logic helped drive attention to other open coding and reasoning releases over the past day.

Readers are no longer only asking which model wins. They are asking who can afford to keep improving it.

What Ripple-65B Must Show Next

The obvious gap is documentation.

Without a full paper, training recipe or benchmark appendix, outside developers cannot yet test whether the one-tenth cost claim compares like for like.

The next meaningful proof would be a model release, evaluation pack or public leaderboard showing how Ripple-65B stacks up on coding agents, not just static code completion. If those artifacts appear, the cost number may carry more weight with developers choosing between open and closed stacks.

Training Budget What The $32 Million Claim Means

A training budget is the total compute and related expense used to turn raw data and base infrastructure into a model.

For a large model like Ripple-65B, much of that bill would normally come from GPU time, storage, networking and engineering effort. When a team says it trained an open model for $32 million, it is making an economics claim as much as a research claim.

That matters to two groups.

Developers want cheaper models because lower training cost can flow through to lower inference prices and more frequent updates.

Investors want it because it changes who can compete. A company that can build a strong coding model without spending frontier-lab money widens the field.

For Ripple-65B, the next step is simple.

Publish enough evidence for others to test whether the cheaper build also holds up in real coding workflows.

Read Next: European Commission Opens 525M Plan So Small Firms Can Make Chips