Laminar's flow-1 Costs $0.0011 a Trace, but Claude Opus 5 Scores Higher on Its Own Benchmark
A.I. / news
Laminar's flow-1 Costs $0.0011 a Trace, but Claude Opus 5 Scores Higher on Its Own Benchmark
The Y Combinator S24 company says its reinforcement-learned error finder matches GPT-6-sol at one-twentieth of the cost. Its launch post and its benchmark table put the multiple at 20 and 23, and show a higher score for Opus 5.
Laminar, a Y Combinator S24 company that sells open-source observability for AI agents, launched flow-1 on Hacker News this week: a small model trained with reinforcement learning to find errors in agent traces. On the company's own 523-trace benchmark it scores a detection F1 of .835, level with Claude Sonnet 5 and below Claude Opus 5 at .890.
The cost claim holds up in Laminar's table. The headline comparison does not name Opus 5, the one model that beat it.
What flow-1 does
Flow-1 runs inside Laminar's Signals harness, which treats each span of an agent trace as a file and investigates with grep and read tools. The Launch HN post, written by the founders under the account skull8888888, calls it "a model trained with RL to find errors in agent traces."
The team did not name its base model. The HN post says only that it finetuned "a small OSS model." The company's flow-1 blog post, dated October 2 and written by Robert Kim, says the base model is not disclosed.
Laminar's platform is Apache-2.0 and self-hostable, according to its GitHub repository, which listed 3.4k stars. The repository does not mention flow-1, and the post does not say whether the model's weights are published.
The benchmark table
The benchmark has no name in the post. It covers 523 difficult traces from coding, legal, customer support, CRM and other domains, and every model ran through the same Signals agent. All scores below are Laminar's, not an outside party's.
| Model | Detection F1 | Precision | Recall |
|---|---|---|---|
| Claude Opus 5 (high) | .890 | .847 | .938 |
| flow-1 | .835 | .817 | .853 |
| Claude Sonnet 5 (high) | .835 | .826 | .844 |
| GPT-6-sol (high) | .816 | .703 | .974 |
| GPT-6-luna (high) | .795 | .727 | .878 |
Opus 5 also leads flow-1 on description F1, .848 against .741. GPT-6-sol has the best recall at .974 but the lowest precision, which Laminar calls noisier. Both GPT-6 models caught a contract redline that omitted mandatory AI-governance protections. Flow-1 missed it.
Why the multiple moves between 20 and 23
The blog headline says "1/20th of the cost" and the homepage says 20 times more traces per dollar. The body of the blog post and the Launch HN post use 23 times, measured against GPT-6-sol on traces under 100,000 tokens. The post does not say how 23 relates to the 20 in its title.
- flow-1888 traces
- GPT-6-luna668 traces
- GPT-6-sol38 traces
- Claude Sonnet 512 traces
- Claude Opus 57 traces
Source: Laminar flow-1 blog post, October 2, 2026
The table gives flow-1 a cost of $0.0011 a trace, GPT-6-sol $0.026 and Opus 5 $0.149. Against Opus 5 the gap is roughly 135 times. Against GPT-6-luna, the nearest rival on cost, flow-1 is 25 percent cheaper and scores .040 higher on detection F1.
Flow-1's listed prices are $0.05 per million input tokens and $0.30 per million output tokens. Anthropic's pricing page lists Opus 5 at $5 and $25 per million tokens and Sonnet 5 at $2 and $10, which accounts for most of the gap to flow-1.
Questions from the thread
Commenters on the Launch HN thread asked whether flags can be checked against human labels and whether the model exposes calibrated probabilities. Others asked how it separates a tool result that looks fine but is wrong from an agent misreading it, and whether it can catch an agent that reports success when tests fail. One commenter asked whether the cost claims hold at scale. The founder said training targets blocked pages and empty results, so they are not flagged as errors.
About 48 percent of the training mix is software engineering, 20 percent customer support and telecom, and 8 percent legal. The post does not give the number of training examples or the reward design.
Other agent coverage on this site includes Trail of Bits' regression hunt and the Claude Duet waitlist.
Laminar said another reinforcement-learning checkpoint is in training to cut costs further. It gave no date.
Sources
- 04Pricing
More in A.I.
- 01WorldTUI Is a 360M-Parameter Model That Fakes a Linux Terminal, and Its Weights Are Not Out YetSimon Edwardsson of V7 trained a trimmed Qwen2.5-Coder on 2 billion tokens of recorded shell sessions. It scores 80 of 106 on his own test and still forgets files it just wrote.
- 02Omnara, a 3,200-Star Apache-2.0 Rival to Claude Managed Agents, Ships With Defaults Its README Calls InsecureThe omnara-ai project runs agents on any model through a Postgres-backed event log and a daemon that dials out. Its README does not name the team behind it.
- 03Meta Paused Muse Chat Sharing After 6,800 Share Hostnames Turned Up in DNS RecordsClaroc found that Muse put each shared conversation's title into its web address, and passive DNS databases kept the addresses. Meta suspended sharing on October 20 and has not said what happens to existing links.
- 04OpenAI Says AI Found the PHP Flaws Before Attackers. Cloudflare Says They Leaked and Drew Attack TrafficThe two Oct. 21 accounts of the php2shell disclosure measure different things. Trail of Bits' Patch the Planet dashboard shows 466 of 1,995 reported issues patched.