DeepSeek Flash may make browser agents far cheaper to run
could change the cost structure of products. The older pattern asks a model to look at a screen, choose the next click, look again, type, and repeat many times. Retriever instead has the model write a JavaScript plan once, then lets a browser harness run the repeated work.
Text page data such as the DOM can be sliced, searched, reused, and cached more easily than screenshots. A task that once needed 40 to 100 can become one planning call plus a few focused extraction steps. Retriever says this cut its cost by more than 100 times while keeping practical close to -class models.
The point is not that is always the smartest model, but that it is cheap and good enough at code to handle the expensive planning path. If are good enough for this role, developers become less locked into expensive APIs from big model labs, and the product’s own harness becomes more valuable.
Key points
- Old often spend too much money by calling a model after every screen change.
- Retriever uses to write a JavaScript plan, then lets a browser harness execute the repeated steps.
- DOM-based text data can be searched, trimmed, reused, and cached more easily than screenshots.
- Retriever claims more than a 100x cost drop while keeping useful -class .
- For builders, harness design and fewer may matter as much as model quality.