Developers Find Ways to Use Claude Code Without Anthropic Models

Boris Cherney, the head of Anthropic’s coding tool Claude Code, last month found himself in the middle of an online spat that highlighted how some developers are finding workarounds to using Claude Code with cheaper non-Anthropic models, including those of rival OpenAI, to save costs.

Developer Alex Getman used a proxy made up of lines of code to connect Claude Code to OpenAI’s GPT-5.6 Sol. Fifteen minutes after Getman typed the first coding requests into Claude, Anthropic shut down the account, he said.

Claude Code is popular because it comes with all the tools assembled to read and run files, search codebases and run tasks in parallel. But it doesn’t need to run on Anthropic models. And developers are saving money by using cheaper models with the Claude Code harness. For that reason, Getman’s posts about the ban drew a million views within hours.

Commenters called Anthropic arrogant for pushing developers toward its own models. Thibault Sottiaux, who runs Claude Code rival Codex for OpenAI, chimed in, saying “we should let our users decide on which model is best for them.” (Sottiaux had already stirred the pot by posting a set of instructions for connecting Claude Code to OpenAI models that Getman said he followed.)

Anthropic’s Cherney weighed in, saying that the company does not ban people from using harnesses, or tools such as Claude Code, with other models. Anthropic ultimately reinstated Getman’s account, with a private apology.

Anthropic may not ban connections to such proxies. But it doesn’t support them, arguing that its popular coding tool is designed to work best with its own models. Use of such proxies—while cheaper than using a router to access a model—could also increase risk from hacks known as prompt injections, a person with direct knowledge of the situation said.

“By design, there’s a Claude model for every task, from the hardest problems to the ones where speed or cost matters most,” an Anthropic spokesperson said.

The practice of connecting Claude Code with non-Anthropic models comes at a cost to Anthropic. When developers use Claude Code with a non-Anthropic model, accessed through a proxy or router, Anthropic loses the stream of revenue it would have received in token payments. In an instance when Claude Code connects to OpenAI’s model, OpenAI receives the token payments—not Anthropic.

But the practice is spreading among developers, according to Tomás Hernando Kofman, founder of Not Diamond, a rival to OpenRouter. “That’s where the whole industry is heading because people are realizing a combination of models yields better results.”

After the Getman episode, some developers turned away from using the proxy. They feared losing access to Claude Code. “I tried it once, and I will never do that again,” Getman told me.

Instead, they’ve continued to pair Claude Code with cheaper, non-Anthropic models by using services such as OpenRouter, which allow developers to switch between a variety of models. In one sign of that trend, Claude Code has been used with 366 distinct models on OpenRouter as of today, according to OpenRouter’s public dashboard.

Adam Kirsh, who is developing an AI startup for manufacturers and industrials, said he has been using Claude Code to write industrial and manufacturing code with non-Anthropic models, thanks to features from OpenRouter and another rival, Concentrate.

“Industrial and manufacturing applications are less represented in training data than typical software so I make a mix to get the result I want,” he said. “It's like taking a car and saying, this is a good car. I just need to put a different engine in it.”

He uses a frontier model, such as OpenAI’s GPT-5.6 or Anthropic’s Fable, to plan the work. He then hands execution to cheaper open-source agents running open-source models. “I do it even though I know Anthropic doesn’t really want me to do that,” he said. He estimates the arrangement gets him about 97% of the quality at roughly 60% less cost.

Kirsh said that if Anthropic cut off access to models through Claude Code, “it would be a huge deal in the community,” of developers. “It’s a philosophy of being open.”

Here’s what else is going on…

AI Deep Dive: OpenAI’s Noam Brown on AI Agents

On Monday, we released the first episode of The Information’s new show, AI Deep Dive. The episode is a conversation with Noam Brown, who has been a research scientist at OpenAI for the past three years, working on breakthroughs in reasoning and AI agents.

While OpenAI’s latest AI model, GPT-6 Astra, notched improvements at a variety of professional tasks, Brown told me OpenAI’s main goal when training new AI models is making them better at AI research and development.

“The number one priority is recursive self-improvement, and by a pretty wide margin,” he said.

Another point he made that interested me was that the multi-agent capabilities on display in the Hugging Face incident (setting up secret message boards, sacrificing themselves for the benefit of the collective) were the result of the multi-agent training that OpenAI had given them. That’s the same level of collaboration that users should expect from future models, Brown told me (though presumably without the penchant for rogue hacking they displayed in the case of Hugging Face).—Rocket Drew

Overheard

President Donald Trump on Monday shot down calls by AI leaders to slow down the pace of development. In a series of social media posts, Trump called claims about AI’s risk to humanity a “hoax,” and equated it with other views he’s sharply criticized, such as climate change.

Meanwhile, during an event on Monday, Nvidia CEO Jensen Huang picked up a call from Trump mid-interview and put him on speaker. “They’re just playing right in the hands of a lot of people that don’t want to see it happen,” Trump said. “That could be political people. It could also be China. And we’re not going to let that happen. It’s a hoax.”

Microsoft on Monday published a new AI “code of conduct” promising to avoid building products that could evade human oversight, following CEO Satya Nadella’s call on Sunday for the tech industry to “pace” its progress in developing new AI. The measures follow similar calls for pacing from leaders of Anthropic, OpenAI, xAI, Meta and other tech companies over the weekend. Microsoft said in its AI code of conduct that it would refrain from building AI models whose behind-the-scenes thinking isn’t clear to humans, a growing concern inside other labs like OpenAI.

Deals and Debuts

See The Information’s Generative AI Database for an exclusive list of private companies and their investors.

SoftBank closed a two-year loan facility with about $11.87 billion in commitments from roughly 20 banks, above its original $10 billion target, to fund its investment in OpenAI, according to Bloomberg.

Firmus, a company that develops and operates AI data centers, is seeking to raise up to $5 billion in an initial public offering in Australia, targeting a listing as soon as the end of October, in what would be one of the largest floats in the country's history, according to Bloomberg.

Z.AI, the company behind the GLM family of models, is raising about $5 billion, including roughly $2 billion from placing 22 million Hong Kong-listed shares, according to the WSJ.

Temporal, whose open-source software keeps long-running computing jobs from breaking, raised $550 million in a Series E funding round at a $12.55 billion valuation led by Lightspeed, Wellington Management, Growth Equity at Goldman Sachs Alternatives and Tiger Global.

OpenAI has bought Glass Imaging, a camera-software company, in a deal valuing it at more than $300 million, according to the WSJ.

Cornelis, a company creating networking technology to help AI chips communicate more effectively, raised $205 million in funding led by IAG Capital Partners.

Buildots, a Tel Aviv-based startup that offers AI software for construction projects, raised $130 million in funding led by O.G. Venture Partners.

Tandem Health, a company whose software listens during a patient appointment and writes up the clinical notes, raised $100 million in a Series B funding round led by the Scaleup Europe Fund.

Nuance Labs, a company developing a single model that lets an AI avatar see and hear a person and respond with face and voice at the same time, raised $50 million in a Series A funding round led by Lightspeed Venture Partners.

Fortaegis, a company that develops security into the silicon of a chip itself, raised an oversubscribed $50 million in a Series A funding round led by Serendipity Capital.

Artificial Intelligence Underwriting Company, which aims to bring AI safety to enterprises and model developers, raised $40 million in Series A funding led by Ribbit Capital.

Jack & Jill, a company that runs recruiting agents, raised $40 million in a Series A funding round, according to Axios.

Flam, a company whose software generates interactive video, real-time 3D scenes and conversational avatars for brands, raised $40 million in a Series B funding round led by QED Investors.

Aron, a company whose AI agents run the purchasing process for corporate procurement teams, launched with $8 million in funding, including a $6 million seed funding round led by Storm Ventures and a $2 million pre-seed funding round led by Menlo Ventures.

Salesforce announced a new AI model called Koa, built on Nvidia’s open-weight Nemotron model. The two companies worked together to post-train Koa to excel at sales, marketing and customer-support-related tasks, they said.

Thank you for reading the AI Agenda Newsletter! I’d love your feedback, ideas and tips: [email protected].

If you think someone else might enjoy this newsletter, please pass it forward or they can sign up here.

Alix Coutures is a reporter covering startups and VC based in New York City. She graduated from the M.A Business concentration at Columbia Journalism School in May 2026. She previously covered tech and telecom in Paris, at the French business publication Challenges. She can be reached at [email protected] or on Signal at Alix_coutures.49