Laminar launches flow-1, a reinforcement-learning-trained model for finding errors in AI agent traces

• Model: "flow-1." Laminar does not disclose the base model, parameter count or whether weights are open (blog).
• What it does: you give it a trace and a prompt, which can be as broad as "identify any logical errors." It returns structured JSON in the format you specify.
• Price: $0.05 per million input tokens, $0.01 per million cached input tokens and $0.30 per million output tokens (blog).
• Access: through Laminar Signals, which can send alerts to Slack or email. The Laminar platform is open source and can be self-hosted (docs). Whether self-hosted installs can use flow-1 is not documented. The docs change itself says flow-1's use in Cloud Signals was inferred from the blog and pricing calculator (PR #235).

• How it works: Laminar says flow-1 behaves like a coding agent. It treats the trace as a repo and each span (one recorded step) as a file. It starts from a preview of the whole trace, searches across spans with grep, then reads selected spans in full (blog).
• Training: reinforcement learning inside the Signals agent. Laminar says it learned how to investigate a trace, not just to classify it as good or bad (blog).
• Evaluation: every model ran in the same Signals harness. That makes the comparison controlled, but it also favours a model trained inside that harness.
• Reproducibility: the reviewed sources do not say how the 523 traces were labelled, do not define detection F1 versus description F1, and do not release the dataset. About half the set (48.2%) is software engineering.

All figures below are reported by Laminar (blog). No independent evaluation was found.

• 23x cheaper: 888 ÷ 38 ≈ 23, measured on traces under 100k tokens.
• More accurate than GPT-6-sol: true by 0.019 in detection F1. GPT-6-sol's gap comes mostly from precision: its recall is very high (0.974), but it flags more traces that turn out to be fine.
• The video's framing: it presents flow-1 as ahead of "GPT-6" 02:19 and leaves out the Claude results entirely.

Flow-1 belongs to a category of tools that use one LLM to judge another LLM's work, often called "LLM-as-judge." Its difference is that it actively investigates the trace with tools instead of reading it in a single pass. Laminar itself is an open-source tracing platform with integrations for frameworks such as the OpenAI Agents SDK (Laminar article), in the same space as Langfuse (Langfuse blog). The bet is that a small, specialised model can replace frontier models as the evaluator, so checking every trace becomes affordable.

• A known miss: Laminar says flow-1 failed to notice that mandatory AI-governance protections were left out of a contract redline. Both GPT-6 models caught the omission (blog; video 06:06).
• Not the most accurate option: Claude Opus 5 outperforms flow-1 on every quality metric in the table. Flow-1 also trails Sonnet 5 on description F1 (0.741 vs 0.773).
• Narrow cost claim: the cost figures cover only traces under 100k tokens. Results may differ with other trace formats or longer workflows (AI Post Hub).
• Biased amplification: the commentary sources are promotional and sell coaching around the tool (video 03:35).

• Full coverage becomes affordable. At about $0.0011 per trace, checking every run of a high-volume agent costs little. Opus 5 costs about $0.149 per trace on the same benchmark (blog).
• Tiered setup: a reasonable pattern is to run flow-1 on every trace and send flagged or high-stakes runs, such as legal work, to a stronger model or a human reviewer.
• Compare against GPT-6-luna before switching. It is close in price, and the main argument for flow-1 over it is higher precision.
• Validate first: before adopting it, run flow-1 on your own traces with known failures. The benchmark labels are not public and it is heavy on software engineering.