October 8, 2026
Cheaper AI agent oversight: Goodfire test cuts reviewer model costs nearly 49-fold
In Goodfire's Oct 8, 2026 evaluation, checking a million turns cost $65 instead of $3,166 with reasoning disabled.
Goodfire launched monitors that read a model's internal signals. They call in a second LLM only to check suspicious actions.
Checking every turn is expensive. With full screening, a second model reads everything an AI agent does. The agent plans its own steps and calls tools, so the reviewing model receives an entire sequence of actions. Goodfire deployed selective screening for Kimi K3 and GLM 5.3.
In Goodfire's Oct 8 evaluation, reviewing model costs and average added latency changed as follows:
| Reviewing model mode | Previous cost per million turns | With monitor | Previous latency per turn | With monitor | | --- | --- | --- | --- | --- | | Without reasoning | $3,166 | $65 | 160 ms | 2 ms | | With reasoning | $5,420 | $185 | 2,750 ms | 42 ms |
Goodfire evaluated the cascade on 2,400 sessions and more than 60,000 turns. It detected around 93% of malicious sessions while sending 5.5% of safe sessions for additional screening.
With the monitor, Kimi K3 blocked 96% of malicious sessions, up from 29%. The share of safe sessions interrupted rose from 5% to 9%.
In FAR.AI's published report, the number of successful universal jailbreaks in a static test fell from 66 to 0. These prompts try to make a model bypass its safeguards. The number of successful individual interactions fell from 700 to 18.
Integration through Goodfire. The company directs teams using open models to the Contact Us form on its website. This involves integrating a monitor into the model's operation, rather than clicking a button in a regular chat.
Sources: [TechCrunch, Oct 8, 2026](https://techcrunch.com/2026/10/08/goodfire-says-its-new-inside-out-monitors-catch-rogue-ai-agents-at-a-fraction-of-the-cost), [Goodfire integration form](https://www.goodfire.com/contact-us).
Goodfire monitors are already available to Baseten customers.