Harvey released its Legal Agent Benchmark (LAB) as open source on May 6, 2026. It tests whether agents can finish long, multi-step legal work across 24 practice areas. Grading is all-or-nothing: a task passes only if every rubric criterion passes. Vals AI runs a leaderboard on Harvey's private 120-task holdout set. On October 26, 2026 it replaced its v1 board with "HLAB-Vals-1.1", and Gemini 4 Argon is now first at 27.08%, ahead of Muse Spark Open at 25.00%. On the archived v1 board, Meta's Muse Spark 1.2 had been first.
This matters because legal-AI buyers and builders use these boards to choose models, and the headline "Argon over Muse Spark" reads like one model overtaking another. The primary pages tell a different story. The benchmark data, harness, scoring and roster all changed at once, and Vals itself says v1.1 results "are not comparable with the archived v1 board."
What we checked
- How much of the ranking reversal is caused by the v1.1 holdout and harness changes rather than changes in the evaluated models? partly confirmed. There was no head-to-head reversal between these two models. On v1, Gemini 4 Argon (19.58%) already scored above Muse Spark Open (16.67%). On v1.1 they score 27.08% and 25.00%. Most models that appear on both boards gained 1.7 to 8.3 points after Vals fixed the holdout, re-synced the harness and added its own harness tweaks, which points to benchmark and harness effects. Vals publishes no ablation, so the size of each effect can't be measured. [2][1]
- Did the Muse Spark models that led v1 lose to Gemini 4 Argon when rerun on v1.1? unresolved. Muse Spark 1.2 (v1 #1, 25.42%), 1.3 Max and 1.3 don't appear among the 29 models on the v1.1 board. The leadership change comes from who is listed, not from a measured loss. The page doesn't say why they were left out. [2][1]
- Do independent runs of Harvey LAB agree that Gemini 4 Argon leads? partly confirmed. Mercor's APEX Harvey LAB board also ranks Gemini 4 Argon (High) first at 27.4% ±5.4%. Artificial Analysis's hallucination-gated LAB-AA v1.1 puts it 10th at 3.9%, because it averages 3.82 material hallucinations per task. [4][5]
What we found
- On the archived v1 board (results as of Oct 18, 2026), Muse Spark 1.2 led at 25.42%, followed by Muse Spark 1.3 Max at 23.75% and Muse Spark 1.3 at 22.92%. Gemini 4 Argon was 6th (19.58%) and Muse Spark Open was 9th (16.67%). Argon already led Open by about 3 points before v1.1. [1]
- The v1.1 board (Oct 26, 2026) lists 29 models. Gemini 4 Argon is first at 27.08% ($10.55/test, 29m15s), then Muse Spark Open at 25.00% ($13.63/test, 3h59m, 22.26M tokens), Wonder at 23.75%, Wonder Full at 23.33% and Grok 4.7 at 18.33%. Muse Spark 1.1, 1.2, 1.3 and 1.3 Max aren't listed, and the page gives no reason. [2]
- Vals lists several changes made together. Harvey fixed the private holdout, removing 17 unsolvable criteria and replacing one mismatched source document. The harness was re-synced to Harvey's upstream code, which adds an explicit finish tool, a reasoning-first judge schema, judge-error retries, visible Word comments and pandoc 3.11. Vals added its own tweaks: context compaction, a nudge when a model stops without calling finish, a cap on tool-result size, image display and path expansion. The criteria pass rate was redefined so it reads higher, while the headline task pass rate and the judges (GPT 5.5 medium and Claude Sonnet 4.6) stay the same. Every model was rerun from scratch. [2]
- Scores rose across the board, which suggests the benchmark changes drove much of the shift. Comparing v1 to v1.1: Argon went up 7.5 points, Muse Spark Open 8.3, Wonder 6.7, Grok 4.7 5.8, GLM 5.3 6.25, Mistral Large 4 1.7 and Wonder Full 1.7. Muse Spark Open's cost per test roughly doubled ($6.83 to $13.63). Several other models show exactly the same cost on both boards (Argon $10.55, Wonder Full $5.41, Wonder $7.08), which the pages don't explain. [2][1]
- Harvey's PR #76, merged June 11, 2026, lets individual criteria keep DOCX tracked changes visible to the judge. Before it, pandoc accepted all redlines and hid them. This is the redline grading fix Vals mentions. It came before both boards, so it doesn't explain the gap between v1 and v1.1. [3][1]
- Independent runs disagree. Mercor's APEX board ranks Gemini 4 Argon (High) first at 27.4% ±5.4%, ahead of Muse Wonder 1.0 at 23.0%. Artificial Analysis's LAB-AA v1.1 uses its own harness, three judges and a rule that zeroes any task with a material hallucination. There Argon ranks 10th at 3.9%, with 3.82 material hallucinations per task, and Muse Wonder Full leads at 13.1%. [4][5]
What it means
Read "Gemini 4 Argon beats Muse Spark" as a roster change on a re-based benchmark, not a measured upset. Argon led Muse Spark Open on both versions. The models it replaced at #1 weren't run on v1.1, or at least aren't shown. Most models gained points under the fixed holdout and the new harness features (finish tool, nudge, compaction), so harness design moves results about as much as the choice of model does. Mercor's ±5.4-point interval also suggests a 2-point gap on 120 tasks is within noise.
For builders, the practical lesson is to evaluate inside your own agent harness. Watch tokens, latency and hallucination risk as closely as pass rate. Muse Spark Open takes almost four hours per task, and Argon's lead disappears once material hallucinations are penalized. If you're shortlisting models for legal work, rerun the leading candidates, including Muse Spark 1.2/1.3, on the v1.1 holdout or the public LAB set with your own tools, and don't mix v1 and v1.1 numbers.
Still open
- Why are Muse Spark 1.1, 1.2, 1.3 and 1.3 Max missing from the v1.1 board, and how would they score on it?
- How much of the score gain comes from the holdout fixes versus Vals' harness additions (finish nudge, compaction, tool-result cap)? No ablation is published.
- Why do several models show identical cost per test on v1 and v1.1 if every model was rerun from scratch?
- When was Harvey's holdout remediation done, and which tasks did the 17 removed criteria affect?