We’re extending our Legal Agent Benchmark (LAB) to in-house contracting. This adds 500 new tasks focused on the contract work in-house legal teams do every day, spanning tasks like drafting, redlining, and escalation, and 90 contract types across 10 areas such as financing, IP, M&A, and employment. 1) Why contracting matters Contracts are the operating layer of the enterprise. They define how companies sell, buy, hire, partner, finance, license IP, share data, and manage risk. In-house legal teams handle the high-stakes work behind those contracts. This includes drafting agreements, negotiating redlines, managing fallback positions, and escalating risks. All of these activities, taken together, are contracting. 2) How LAB models contracting Each task has four parts: Environment: all context available at a point in the negotiation, including a data room, playbooks, guidance, emails, and prior negotiation turns. Instruction: a short next-step prompt, like draft the agreement, respond to a redline, or prepare an issues list. Output: the artifact that moves the negotiation forward. Verification: expert rubrics grade the agent's output against the appropriate next step, including what the agent edited, accepted, rejected, negotiated, or escalated. See @gabepereyra’s article for a walkthrough of an example LAB task: responding to a counterparty redline of an MSA that has already been heavily negotiated. 3) What comes next Today, LAB measures how agents handle each stage of a contract negotiation as a discrete step. Next, we’ll expand across more contract types, more negotiation states, and interactive benchmarks where agents must decide whether to negotiate, escalate, or walk away. Read more from our cofounder @gabepereyra below.

We partnered with @trajectorylabs to post-train NVIDIA Nemotron 3 Ultra for legal. Here’s what we found: 1) Open-weight models can reach frontier legal performance. On our Legal Agent Benchmark (LAB), Nemotron 3 Ultra started at a 0% all-pass rate. After post-training, it reached 5.8%, placing it between Sonnet 4.6 at 4.2% and Opus 4.6 at 6.6%. 2) Post-training dramatically improves reliability. Before training, many held-out tasks missed enough rubric dimensions to land around ~70% pass rates. After training, those tasks shifted toward ~95% pass rates. 3) Open-weight performance comes at much lower cost. Post-trained Nemotron 3 Ultra reached a similar quality band to leading closed models while running at roughly 1/8th to 1/50th the per-token price of Sonnet 4.6 and Opus 4.6. Most importantly: we post-trained this model on the @trajectorylabs platform less than 24 hours after Nemotron 3 Ultra launched, using the same harness, data, and recipe we used for Nemotron 3 Super. More to come as we continue to experiment with open-weight legal agents. Read more on post-training with Trajectory below:

The first prompt I always run on a new model to test legal capabilities is “Draft an S-1” Immediately gives a good sense of the model's general capabilities. Can tell from the length of the S-1, formatting, structure and writing how big a jump in general and legal performance the new model is likely to be. We’ve found that just length correlates extremely well to how well the model will work in our legal agents. Screenshots show SpaceX S-1 drafted by Fable 5, Opus 4.8 and the actual one filed. Very clear that Fable 5 is a big step up from Opus 4.8 which is already significantly stronger than most other models at this task. The formatting and structure is significantly better and is also reflected in our LAB benchmark (13% vs. 10%). Super impressed in early testing of this model both on benchmark and in product. Huge congrats to the @anthropic team because this is very clearly a big step forward in model capabilities.

Last week we held Harvey Hacks, our internal hackathon. 27 projects total across our 200-person eng team. Wanted to highlight a few hackathon projects:

At @harvey, the engineering team integrated Spectre — their internal background agent — into Devin Desktop. Now Spectre's organizational context can live on every engineer's laptop and flow across their favorite agents.