|
Coinbase and some other tech-savvy firms such as Shopify and Ramp have built their own AI coding agents for internal use, giving their employees an alternative to pricey options from firms like Anthropic and OpenAI. Coinbase, the cryptocurrency exchange, built an AI coding agent called Forge, which launched to all of its engineers in April of this year, senior director of AI engineering for Coinbase Wallet Chintan Turakhia told me. Engineers can access the coding agent through Slack, Github, and its own web user interface. Usage of Forge is growing fast internally and has resulted in cost savings, said Turakhia, although he declined to be specific. “There was a time where we didn't have, really, sort of any limits on AI spend, and yes, the spend went up,” he said. But as Forge has become more popular among engineers, “now our spend has gone down while our token usage continues to go up.” That’s partly because Forge connects to the company’s internal model router to tap different models from Anthropic, OpenAI, Google, and open source model providers to save on cost. Coinbase CEO Brian Armstrong in a late June post on X said the company has experimented with using Chinese open source models such as GLM 5.2 and Kimi K2.7. He said the company has also used other methods to cut its AI spending “nearly in half,” such as starting new sessions with AI tools to avoid giving the AI too much information to process and thus raising costs. Forge can be tagged in Slack channels, which are essentially large group chats within the messaging app. It can perform tasks like brainstorming ideas for new software features or products and building them. Forge can work inside Coinbase’s code repositories and enterprise software systems such as Datadog and Sentry. Still, Coinbase hasn’t stopped using Anthropic’s Claude Code. In fact, Claude Code is still the most-used AI coding tool used among the 2,500 Coinbase engineers who use such tools, including Cursor and Forge, Turakhia said. “There is no crisp answer on when someone uses Claude Code versus Forge. It depends on the engineer's preferences,” he added. Other companies have figured out how to control their spending on Anthropic’s coding tool by developing internal tools. In this story from last week, I talked about IT firm Globant using an internal coding agent that can swap between models from Anthropic and OpenAI to open source ones for coding work. Then there’s Shopify, whose engineers use its agent River through Slack in addition to Claude Code as well as Codex. Shopify executives said in a May presentation that 75% of its employees use the tool and that it is “very efficient at using tokens,” or doing more coding work at a lower cost. To be sure, sometimes internal AI developments at companies still result in high costs. For example, Walmart had to cap internal usage of its in-house coding agent called Code Puppy after soaring usage, Bloomberg reported in June. Palantir CEO Feeds the AI Fear Fire Palantir CEO Alex Karp has lately been beating the drum about the potential competitive threat AI labs pose to their own customers because they have access to those customers’ proprietary business data. Though OpenAI and Anthropic say they don’t train their AI models on their enterprise customers’ data, there are ways they could use some of those customers’ chat log data to improve their products—an issue my colleagues previously covered in this column. And Anthropic has a history of introducing products that compete with its own partners such as Cursor and Figma. Karp played into the fear factor in his quarterly letter to shareholders Monday: “Every organization in the world is awakening to the risks of handing the creators of the language models the keys to their institutions, of letting the models loose within their homes.” While his warnings have some merit, it’s an all-too-convenient message for Palantir and other software firms, which have seen their stocks fall this year over worries that AI tools could diminish their competitive positions. The companies—Palantir, Microsoft, Salesforce, and others—want to capitalize on recent criticisms of AI labs and position themselves as necessary intermediaries for enterprises. “Our customers have declined to become vassal states of the language labs,” Karp wrote with his characteristic dramatic flair. “And the market is now shifting dramatically underfoot.” Whether the latter will prove true long term remains to be seen, but Palantir certainly appears to be benefiting from popularity among enterprises in the near term, as seen in the record value of contracts it signed in the June quarter with U.S. enterprises.
| 
