Sentry Retires 500 AI Keys After Moving Internal Use to OpenRouter

A lack of visibility drove the migration

Sentry had encouraged employees to experiment with AI, allowing people to create API keys across its enterprise accounts with Anthropic and OpenAI. That approach supported rapid testing, but the company struggled to see who was spending money, which models they were using, and whether the models were suitable for particular tasks.

One engineer’s usage reached $3,000 a day after a custom coding-harness extension broke prompt caching on every request. The issue ran for a couple of days before Sentry identified it. According to Daniel Griesser, Sentry’s Senior Director of Engineering, the incident exposed the limits of the company’s reporting.

“We needed to change something, because it wasn’t sustainable. We had no clue where the money was going or who was spending what,” Griesser said in the OpenRouter case study.

One account replaced hundreds of provider keys

Sentry began its OpenRouter rollout at the start of July 2026. Everyone received access on day one, including teams outside engineering, and users could create their own keys. By default, each key is linked to the person who creates it, while each workspace has a budget cap approved by Sentry’s CFO.

The company cleaned up roughly 500 Anthropic and OpenAI keys within three weeks. Only about 50 had been in daily use, but Sentry tracked down every key owner and gave them a week’s notice to create an OpenRouter key before the old credentials were disabled.

OpenRouter’s Anthropic-compatible and OpenAI-compatible endpoints meant that moving a tool generally required changing three variables while keeping its existing software development kit. In August, Sentry switched off engineers’ direct access to Anthropic and OpenAI, preventing them from creating new keys there.

Sentry enabled zero data retention by default and otherwise routed requests through its existing provider keys using bring-your-own-key. OpenRouter handled attribution and limits, while Sentry’s coding harnesses and most workspaces continued using its Anthropic and OpenAI keys.

Broader model access with spending controls

During its first three months on OpenRouter, Sentry’s account used 147 models, counting each version separately. Engineers can point a key at a newly released model and run an evaluation without seeking a new vendor approval. Griesser described the change as a shift from a lengthy approval process to a configuration change.

Frontier models still accounted for most usage, while open-weight models represented about 12% of tokens, most of it from Kimi K2.6 running evaluation suites for Sentry’s internal tools, including Warden, its agent for code and security review. GPT-5.6 Luna was the default for high-volume tools such as Sentry’s official MCP server, carrying 30.8% of account tokens and 5.2% of model spend.

More than 113 people now have accounts, and 45+ apps send traffic through the shared setup. Those apps include internal agents and tools such as Pi, Claude Code, OpenCode, Codex and Cursor, as well as services built by Sentry’s marketing and support teams.

According to OpenRouter, the move did not change the volume of Sentry’s spending. It gave the company clearer ownership of requests, model-level visibility and budget guardrails, including a $1,500 weekly budget per person in Sentry’s default workspace. For users, the result is broader model choice without requiring a separate provider relationship for every experiment.