Laminar's flow-1 Costs 23 Times Less Than GPT-6-sol, Trails Opus 5
Startups / news
Laminar's flow-1 Costs 23 Times Less Than GPT-6-sol, Trails Opus 5
Laminar, a Y Combinator S24 company, ran the 523-trace benchmark itself, and its own table has Claude Opus 5 ahead on every quality metric.
Laminar, a Y Combinator S24 company that sells observability for AI agents, says its new flow-1 model matches GPT-6-sol at finding errors in agent traces while costing 23 times less. Laminar's own table, published Oct. 2 and run by the company on 523 traces, shows a narrower result: flow-1 is .019 ahead of GPT-6-sol on detection F1 and behind Claude Opus 5 on every quality column. The figures are vendor-supplied, and no independent party is named as having reproduced them.
Laminar introduced flow-1 to Hacker News in a Launch HN post. The core product is Apache 2.0 and self-hostable, and Laminar earns money from a managed cloud.
What the 523-trace table shows
| Model | Detection F1 | Precision / recall | Cost per trace |
|---|---|---|---|
| flow-1 | .835 | .817 / .853 | $0.0011 |
| Claude Opus 5 | .890 | .847 / .938 | $0.149 |
| Claude Sonnet 5 | .835 | .826 / .844 | $0.086 |
| GPT-6-sol | .816 | .703 / .974 | $0.026 |
| GPT-6-luna | .795 | .727 / .878 | $0.0015 |
Every model ran inside the same Signals agent at high reasoning, on traces under 100,000 tokens. Cost includes retrieval steps. $0.026 divided by $0.0011 is 23.6, which is the 23 times, and it is also a tie with Sonnet 5 at .835 rather than a win.
- flow-1888 traces per $1
- GPT-6-luna668 traces per $1
- GPT-6-sol38 traces per $1
- Claude Sonnet 512 traces per $1
- Claude Opus 57 traces per $1
Source: Laminar blog, flow-1 post, Oct. 2, 2026 (vendor-supplied), accessed 2026-10-19
The post says flow-1 "matches" GPT-6-sol in one place and surpasses it in another. It also concedes limits: flow-1 cleared a contract redline that omitted mandatory AI-governance protections, which both GPT-6 models caught, and GPT-6-sol's recall of .974 beats flow-1's .853 while its precision of .703 is lower than flow-1's .817. The post's title promises one-twentieth of the cost, a ratio the table gives against none of the models listed.
How flow-1 is trained and priced
Laminar trained flow-1 with reinforcement learning inside the Signals agent, using the production tools. By its own breakdown, 48.2 percent of training examples are software engineering, 20.3 percent customer support and telecom, and 7.7 percent legal. The post does not give the base model, the number of examples or the compute. The Launch HN post says a small open-source model was fine-tuned on "thousands of synthetic investigations".
flow-1 is priced at $0.05 per million input tokens, $0.01 per million cached input tokens and $0.30 per million output tokens. The Signals quickstart does not mention flow-1 and says model choice applies only on self-hosted deployments. None of the pages read say whether flow-1's weights can be downloaded.
The company behind it
Laminar announced a $3 million seed on March 16, led by Atlantic.vc with Y Combinator and AAL.vc, plus angels Ben Sigelman, co-creator of OpenTelemetry, and Ant Wilson, Supabase's chief technology officer. Its customers include Browser Use, OpenHands, Rye.com, Alai and LegionIntel. The GitHub repository, lmnr-ai/lmnr, lists 3.4k stars, 250 forks and 2,086 commits under Apache-2.0.
The Hacker News thread is a useful check on the pitch. One commenter said they stopped reading at the opening comparison with competitors and asked for bullet points. Another asked whether findings are ranked by frequency or user impact, and the reply was that flow-1 predicts how critical each finding is, without saying how. A third asked about voice agents, and the answer was that Laminar does not yet trace voice SDKs such as OpenAI's realtime API but works with transcripts.
Where it sits among agent tooling
Agent-harness projects are multiplying: this site covered Agent37, a YC F26 company pricing agent sandboxes at $0.37 for 57 hours, and Google Research's Harness2, which has 10,700 stars and three commits. Laminar's difference is that it sells the judge rather than the runtime. Its blog says another reinforcement-learning checkpoint is training to cut flow-1's cost further, with no date given.
Sources
More in Startups
- 01Town Raises $108 Million at $1 Billion Four Months After Public LaunchForerunner Ventures and Menlo Ventures co-led the Series B, and the only revenue disclosure is a seven-figure ARR with no baseline for the growth claims.
- 02Tonbo Imaging's IPO Closed 135.9 Times Covered at 79 Times EarningsThe ₹852 crore offer is entirely a sale of existing shares, so none of the money reaches a company whose FY26 profit fell to ₹50.88 crore from ₹72.76 crore.
- 03Tenstorrent Raises $1.5 Billion at $6.2 Billion, Months After Takeover TalksTrace Capital and Innovation X led a Series E that values the chip company at 2.4 times the $2.6 billion Bloomberg reported for its December 2024 round, and revenue stays undisclosed.
- 04Zetwerk's ₹5,012 Crore IPO Closed 3.06 Times Covered, With Retail Bids at 73% of Their QuotaQualified institutions bid 4.73 times their allocation, about 83% of all shares bid, while retail fell short and the valuation depends on a revenue period the coverage does not pin down.