In the November 18, 2026 snapshot of Vals Index 2.2, Gemini 4 Neon at XHigh scored 70.87% and ranked second out of 42 models, averaging $7.40 and 31 minutes 25 seconds per test. It scored 0.44 percentage points below first-place Claude Fable 5.2 and cost about 41% as much. Leaderboard results
Suitable comparisons: Model scores, API costs, and completion times on Vals AI's multistep work tasks, or the difference in overhead between Neon's two reasoning tiers.
Limits on extrapolation: The results do not establish how the model performs on legal, investment, or tax decisions, what its success rates or bills would be in production workflows, or its time to first token in chat.
Applicable model version: Google Gemini 4 Neon. Vals AI did not specify the exact deployment snapshot.
Test environment or client: Vals AI ran eight agent benchmarks. Tasks and tool configurations vary by benchmark. Evaluation methodology
Reasoning tier and parameters: The settings panel for Neon on the overall leaderboard lists Google as the provider, XHigh reasoning, default temperature and Top P, and a maximum output of 1,048,576 tokens. Vals AI also tested High separately. Overall leaderboard settings · High results
Vals Index 2.2 launched on October 19, 2026 and combines six private benchmarks and two public benchmarks. Vals AI calculates the arithmetic mean of the benchmark scores within each domain, then weights each domain by its industry's share of U.S. GDP as measured by value added, using figures from the U.S. Bureau of Economic Analysis. The index uses the information sector as a proxy for programming. The leaderboard snapshot is from November 18. Weights, version, and sample details
Vals Index = (8.0 × Finance + 5.6 × Programming + 1.2 × Legal + 0.5 × Tax) ÷ 15.3
The composite uses each benchmark's own scoring method. For Code Migration v1.1, the index uses a fixed subset of 50 of the 120 command-line migration tasks and all 10 COBOL tasks. The two groups carry weights of 75% and 25%, respectively. The leaderboard measures agent runtime as wall-clock time from task start to completion. Component scoring · Metric definitions
Vals AI reported that High scored 70.5%, used fewer tool calls, cost about half as much as XHigh, and took about 40% less time. The post reported cost and runtime only in relative terms. In the benchmark breakdown shared at the same time, XHigh scored 7.3 percentage points higher than High on Tax Agent, and High scored slightly higher than XHigh on Finance Agent. High comparison · Reasoning tier comparison by benchmark
| Rank | Model | Composite score | Average cost per test (USD) | Average tokens per test | Average runtime per test |
|---|---|---|---|---|---|
| 1 | Claude Fable 5.2 | 71.31% | 17.94 | 31.31 million | 43 min 42 sec |
| 2 | Gemini 4 Neon (XHigh) | 70.87% | 7.40 | 15.15 million | 31 min 25 sec |
| 3 | Claude Sonnet 5.5 | 70.85% | 16.18 | 78.59 million | 1 hr 26 min |
Neon's cost breakdown is $2.18 for input (including cache writes where applicable), $2.07 for cache reads, $1.71 for output, and $1.44 for reasoning. The amounts use the precision shown on the overall leaderboard. Overall leaderboard and cost breakdown
| Domain | Neon XHigh benchmark scores (at the precision shown in the component table) |
|---|---|
| Finance | Finance Agent v2.1 68.6%, Excel Modeling Benchmark v1.1 76.5% |
| Programming | Terminal-Bench 4.0 61.6%, Vibe Code Bench v1.1 93.7%, Code Migration v1.1 75.8% |
| Legal | Legal Research Bench 56.3%, HLAB-Vals-1.1 19.6% |
| Tax | Tax Agent Bench 54.9% |
The scores come from the index's Accuracy by Benchmark table. The Code Migration entry uses the fixed subset described above. Scores for all eight benchmarks
Neon and Sonnet 5.5 are 0.02 percentage points apart on the leaderboard, and Neon's standard error is 0.94 percentage points. The ranking reflects the order of the point estimates in this snapshot. U.S. GDP weights affect the composite score, and the same model scores quite differently on legal research and long-horizon legal deliverables. To assess reliability on a real task, look at results on similar tasks under the same scoring criteria. Neon results and standard error · Error and evaluation methodology
Expand Neon's entry on the overall leaderboard to see the settings and cost breakdown. Use the eight benchmark scores and the public formula to check the composite. The benchmark table shows one decimal place, so the recalculated score may differ slightly due to rounding. Each public benchmark can be rerun separately using the version, tools, and scoring settings documented on its page. A full rerun of the index requires the six private test sets and their run configurations. Vals AI uses only its non-public test sets for its private leaderboards. The validation examples it publishes for some benchmarks are suitable only for checking the evaluation workflow. Index methodology · Public and private test sets
Gemini 4 Neon