promptfoo 0.122.1 adds multi-turn red-team token accounting and new model providers

Promptfoo released v0.122.1 (Aug 26, 2026), adding features to distinguish eval scan footprint from incurred token usage, account for multi-turn attacker and grading tokens in red-team runs, and add Gemini 3.7 Flash/Flash-Lite provider support.

Improves cost transparency and accuracy of multi-turn adversarial red-teaming runs in a widely-used open-source LLM security testing tool, incrementally strengthening AI red-team tooling accessible to any team.

AI red-team/eval practitioners using promptfoo for adversarial testing should upgrade for accurate multi-turn token accounting.