An Adapted 27-Billion-Parameter Open-Weight Model Outscored Claude Opus 4.8 on Legal Work at a Tenth of the Cost

The AI-memory startup Engram and the legal AI company Harvey built a fictional law firm, Calderwood and Harkness, with more than 100 million tokens across roughly 10,000 documents and 266 matters. They then adapted the open-weight Qwen3.8-27B to that firm using parametric memory, compressed notes, and a retrieval tool. Across 250 legal tasks, the adapted model averaged 67 percent, ahead of every model tested, and on strict all-or-nothing correctness it led Claude Opus 4.8 by 30 percent to 25 percent, according to Engram's writeup. The Russian-language channel AI ML Big Data summarized the result for its readers.

The cost difference is the number engineers should note. Engram reports roughly 0.13 dollars per query for the adapted 27-billion-parameter agent, against 1.32 dollars for Opus 4.8, a tenfold difference on the same task set. Alibaba separately promotes Qwen3.8-27B as the leading open-weight model on Harvey's Legal Agent Benchmark, which scores end-to-end completion of complex legal work under an all-pass standard where absolute scores stay low for every model.

Treat the headline claim with the caution any vendor-run evaluation deserves. Engram sells agent memory, Harvey sells legal AI, the firm tested is synthetic rather than a real client corpus, and no independent group has reproduced the comparison. Harvey also runs Opus 4.8 in its own product, so this is not a competitor attacking an incumbent from outside. What the numbers do support is narrower, and still important: on a domain the test was built around in advance, an adapted small open-weight model narrowed the difference with a frontier model that Anthropic prices as premium work, alongside its newer Opus 5 tier.

The mechanism is the adaptation process, not the underlying weights. A model that has already absorbed a workspace does not have to reread it on every call, which converts context tokens into a one-time adaptation cost.