AI labs like Anthropic and OpenAI ship coding tools such as Claude Code and Codex largely to keep developers comfortably locked into their own, usually most expensive, models. But a growing number of users want to run those same tools on third-party or cheaper models instead, to save money. As a recent report from industry outlet The Information shows, that puts the vendors in a genuine bind—and one very public suspension at Anthropic made the friction concrete.
Key takeaways
- According to The Information, Anthropic and OpenAI face a dilemma: let users freely choose which model powers their coding tool and lose revenue, or restrict it and risk their standing with developers.
- In early August 2026, developer Alex Getman was briefly suspended by Anthropic after running Claude Code through a local proxy pointed at OpenAI’s GPT-5.6 Sol—the exact cause was never confirmed.
- Claude Code lead Boris Cherny publicly clarified that Anthropic doesn’t ban accounts simply for using other models: the suspension was likely a different classifier, and the account was restored.
- Anthropic tolerates third-party models through documented gateways but advises against it on security grounds—extra proxy layers can create risks such as prompt injection.
- More developers are reportedly turning to model-router services like OpenRouter and Concentrate instead of running their own proxy setups.
The Getman case: one suspension, many open questions
How deep the conflict runs became clear in early August 2026 through one concrete case. Developer Alex Getman published a small tool on GitHub called claude-proxy: it leaves the official, unmodified Claude Code CLI untouched but routes model requests through a localhost-only CLIProxyAPI server to a model of the user’s choice—in Getman’s case, OpenAI’s GPT-5.6 Sol, billed through his own OpenAI account. He was reproducing a setup OpenAI Codex lead Thibault “Tibo” Sottiaux had publicly shared himself. Shortly after testing it, Getman’s Anthropic account was suspended for “suspicious signals.”
Getman made the case public and tagged both Anthropic and OpenAI, prompting Sottiaux and Claude Code lead Boris Cherny to weigh in directly. Cherny clarified that Anthropic doesn’t ban accounts simply for running their harness against another model; the suspension was likely triggered by an unrelated fraud or abuse classifier. The account was restored a few hours later. Whether the proxy setup actually caused the suspension remains unconfirmed to this day—Getman himself stresses in his repository that he only knows the sequence of events, not the cause.
Why Anthropic and OpenAI are torn
The conflict of interest is out in the open: coding tools like Claude Code and Codex exist largely to steer developers toward the vendor’s own, often priciest, models. Developers counter that many tasks work just fine with cheaper or open models. Anthropic officially justifies its caution mainly on security grounds: extra layers in proxy setups can open new attack surfaces, such as prompt injection—an attack where inserted text tricks a model into unwanted actions, a risk that also played a role in the AI misbehavior cases OpenAI recently disclosed. Anthropic now officially documents the use of such gateways and even names specific tools like LiteLLM—while explicitly stating that routing to non-Claude models itself isn’t supported. OpenAI positions itself more openly: Codex lead Sottiaux said publicly that users should be free to decide which model they use in Codex.
Developers look for workarounds
After the Getman incident, many developers now see homemade proxy setups as too risky for their own accounts. Instead, according to The Information, more users are turning to specialized model-router services like OpenRouter and Concentrate, which let them pick commercial or open models centrally without running their own proxy infrastructure. For developers, that means one more intermediary to depend on—but also more distance from the individual AI vendors’ classifiers, which in Getman’s case apparently fired too quickly.
Context
For developers who rely heavily on AI-assisted coding tools, this dispute is more than a footnote. It lands on top of an already strained calculation: AI labs still aren’t earning back what they spend on compute for their cloud models—an imbalance also visible in the current capacity crunch around new flagship models. If cloud model prices rise noticeably, or if vendors tighten free model choice inside their tools, it will hit exactly those users already hunting for cheaper alternatives. Many developers’ hope for the kind of openness they’re used to from operating systems or free development environments runs up against a reality where vendors’ financial incentives point the other way. Cherny and Sottiaux have both publicly said users should be free to choose—whether that promise holds once revenue pressure keeps rising remains an open question.