AI Aulis

AI NEWS IN ENGLISH · 20 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

41 stories · next issue tomorrow morning · How the paper is made →

← Issue 20 November 2026Paper

Gemini 4 Neon tops Vals AI’s Legal Research Bench with 56.25 per cent

AI-generated contentWritten by AI

blog.google, git-aarvin.github.io, latent.space, vals.ai, artificialanalysis.ai, x.com, arXiv · 8 sources · Sources published: 30 September–19 November 2026 · Some source publication dates are unknown

Gemini 4 Neon answered 56.25 per cent of Legal Research Bench questions fully correctly, but it leads the next models by less than one percentage point. In another Vals AI benchmark of legal work, Neon came eighth.

Vals AI updated its Legal Research Bench results on 18 November, and Gemini 4 Neon now tops the leaderboard. Gemini 4 Carbon is second and Muse Spark 1.3 Max third. Google cites the ranking in its Neon launch announcement, but Vals AI ran the evaluation.

In the benchmark, an AI agent researches a US legal question by searching case law and the web, then writes a reasoned answer. Lawyers drew up a rubric listing the points a correct answer must contain. An answer passes only if it meets every rubric item. The grader is the language model GPT 5.4. Vals AI said it agrees with the majority of three experts on 87.4 per cent of rubric items. Human graders average 83.4 per cent.

The table shows that less than one percentage point separates the top five models. Their cost per question, however, ranges from $0.59 to $23.06.

Model Fully correct Cost per question
Gemini 4 Neon 56.25% $7.37
Gemini 4 Carbon 55.77% $8.74
Muse Spark 1.3 Max 55.29% $0.59
Claude Opus 5 55.29% $6.76
Claude Fable 5.1 55.29% $23.06

Source: Vals AI, Legal Research Bench, 18 November 2026. Results are calculated on an unpublished test set of 208 questions.

The ranking does not show that Neon is better than the others at legal research. The test set has 208 questions, so Neon’s 0.48-percentage-point lead over Carbon amounts to one question. Vals AI researchers described the benchmark in an arXiv preprint published on 30 September. The preprint’s leaderboard listed 13 models, and only two of the 12 adjacent pairs differed with statistical significance. The authors said the leaderboard therefore shows tiers of models rather than an exact order.

The leading model also varies by area of law. Neon led on questions about health, administration, and constitutional and civil rights; Carbon led in criminal and family law.

In HLAB-Vals-1.1, Vals AI’s other benchmark of legal work, Neon fully solved 19.58 per cent of tasks and ranked eighth. The benchmark is based on Harvey’s Legal Agent Benchmark, in which an agent does a lawyer’s work offline using the documents, spreadsheets and presentations it is given. Gemini 4 Argon leads with 27.08 per cent, but the Latent Space newsletter reported that Argon has already been retired. The same model can therefore do well in legal research and reasoning yet fall behind in document work.

Neon is available on the Gemini API’s paid tier at $3 per million input tokens and $15 per million output tokens. In the Gemini app, Google AI Pro and Ultra subscribers can use it.

Report an error in this story