Benchmarks · Real-world & economic value
Vals Index
A single score for a model's likely economic usefulness: how well it does agentic work in finance, coding, law and tax, weighted by each sector's share of US GDP.
Vals AIReleased 2026Live
Vals AI runs dozens of benchmarks of professional work, most of them on private task sets it builds with domain experts. The Vals Index rolls the agentic ones into a single number meant to approximate a model’s economic impact. Rather than averaging its components equally, it weights each sector by its share of US GDP, so finance counts for more than half the score and tax for about 3%. The index is run entirely by Vals, which also publishes each model’s cost and run time per task.
It is revised often. Version 2, in August 2026, replaced SWE-bench Verified with Vals’s own coding and legal tests. Version 2.1 in September added a tax sector and moved to Terminal-Bench 4.0, and version 2.2 in October updated four components. Vals says scores are not comparable across versions. That makes claims about who “leads the Vals Index” depend on the date: Gemini 4 Argon led v2.1 at 68.9%, but was fourth on v2.2 a few weeks later.
At Gemini 4 Carbon’s launch the top of v2.2 was tight: Claude Sonnet 5.5 at 70.85%, Carbon at 70.36% and Claude Opus 5.5 at 70.18%, all within a point. Google’s comparison table showed Carbon first because it left Sonnet 5.5 out.
The set
A composite of eight benchmarks, six of them private, all run by Vals AI itself. In version 2.2: Finance (Finance Agent v2.1, Excel Modeling Benchmark v1.1), Coding (Terminal-Bench 4.0, Vibe Code Bench v1.1, Code Migration v1.1), Legal (Legal Research Bench, Harvey's Legal Agent Benchmark) and Tax (Tax Agent Bench). Each sector score is the average of its components, and the sectors are combined with weights from US GDP shares: (8.0 × Finance + 5.6 × Coding + 1.2 × Legal + 0.5 × Tax) / 15.3. Reported as accuracy, 0–100, with cost, tokens and run time per task alongside.
Where it stands
Revised often — v1.1 in May 2026, a rebuilt v2 in August, v2.1 in September, v2.2 in October — and scores do not carry across versions. Most components are private, so the index can be checked only against Vals's own board. Launch posts from Google cite it. Vals has not published the original launch date; the earliest archived board is v1 as of May 2026, and the rebuilt v2 was announced with the company's Series A in August 2026.
How the top score changed hands
Vals Index v2.2current
- September 2026Gemini 4 Argon68.90% (v2.1)Led version 2.1 (board as of 18 October 2026), the first Gemini model to top the index; fourth on v2.2.
- October 2026Gemini 4 Carbon70.36%Second on v2.2 at launch, 0.49 points behind Sonnet 5.5.
Current best: Claude Sonnet 5.5 — 70.85% Vals AI board as of 27 October 2026: Sonnet 5.5 70.85%, Gemini 4 Carbon 70.36%, Claude Opus 5.5 70.18%, Gemini 4 Argon 69.24%. Google's Carbon launch table, which omitted Sonnet 5.5, showed Carbon first.
In the timeline · 3 entries
Google releases Gemini 4 Carbon to cyber defenders a month after Argon
Artificial Analysis ranked it first on its new v5.0 index at 59, a point above Claude Opus 5.5; introductory pricing matches Argon at $2/$10 per million tokens.
Models & capabilities · Security & misuse
Google makes Gemini 4 Argon generally available
Google's safety report says a chain-of-thought penalty applied by mistake in training more than tripled the shortcuts Argon took without mentioning them, from 11.2% to 36.4%.
Models & capabilities · Safety & alignment
Google releases Gemini 4 Argon to cyber defenders before a wider launch
The output limit rises from 64,000 to one million tokens and introductory pricing is $2/$10 per million; Artificial Analysis placed it level with GPT-6 Astra, behind Anthropic's newest models.
Models & capabilities · Security & misuse