Harvey post-trained a 27B open-weight model into the frontier band on its own legal benchmark

A LAB task is shaped like the week an associate dreads: a partner-style instruction, a closed universe of matter documents, and a work product someone will read line by line against a rubric an expert wrote ^[5]. Grading is all-pass, so a task counts only when every criterion passes ^[5]. That is why the under-10% end-to-end figure ^[2] and the 63.0% criterion pass rate from the training run ^[10] are not in tension. The first measures whether the work would have shipped unedited; the second is closer to how much of it a reviewing lawyer would keep ^[17].

The cost line is what should shape a pilot. Roughly $50 a task and over 20 minutes of latency at the top of the closed-source leaderboard ^[3] means running LAB's own taxonomy of more than 1,200 tasks across 24 practice areas ^[4] once would cost about $60,000 ^[14] and take 400 hours if the tasks ran one after another ^[15]. No firm runs the benchmark, but the unit economics carry over to any workflow where the agent needs several attempts before a human keeps one.

Before any training, the hold-out baselines already complicate the buy decision ^[18]. Sonnet 4.6, Opus 4.7 and GPT 5.5 lead, with GLM 5.1 and DeepSeek v4 mixed in among them and the other popular open-weight models below ^[8]. Two open-weight models sitting inside the leading group before Harvey touched anything is the strongest part of its case, and it is a case about the price floor rather than the 10% ceiling ^[2].

Harvey's account of where frontier quality comes from is about reading habits: Opus, Sonnet and GPT-5.5 opened most of the data room and read documents in full, often near 90% coverage, while weaker open-weight models used grep to surface passages and rarely read anything end to end ^[9]. The GRPO run on Qwen3.5-9B tested whether that habit can be learned from outcome reward alone, with reward coming only from the rubric and no retrieval-specific signal ^[10]. Criterion pass rate moved 20.5 points ^[16], though Harvey does not say in this post whether document coverage moved with it ^[10].

The conclusion Harvey draws is that post-training and harness optimization have to be developed together ^[11]. For a buyer that is the fine print: the gain belongs to a model and a harness as a pair, not to a 27B checkpoint anyone can lift into their own stack. The next-step list asks for richer, less lossy compaction and for compaction artifacts to anchor partial credit in long-horizon RL ^[12], which is a plain way of saying the long runs currently drop information they need and the reward signal is too coarse to see where.

Two questions sort the candidate workflows: whether a human reviews the output before it leaves the firm, and whether the cost of one attempt fits inside what the matter can absorb. Reviewed and cheap is where legal agents already earn their keep. Reviewed and expensive is the $50 quadrant ^[3], and it is the only one where post-training changes the answer, since price and hosting the weights in your own environment are what it is meant to buy ^[7]. Unreviewed and cheap runs into the under-10% number ^[2], and unreviewed and expensive has no case.

Harvey has not published the post-trained model's cost per task or its latency on LAB, so the frontier-band result is doing the work a cost number should do, and the score comes from the company that wrote the benchmark ^[19]. Until that price appears, the quadrant boundaries sit where they already sat.