Gemini 4 Argon tops Vals with leaner agent runs

Google’s Gemini 4 Argon has reached No. 1 on the Vals Index with a score of 68.9%, giving Google its first outright lead on the benchmark. The index tests work across finance, law, software development, and tax using tasks designed to approximate professional workflows. Argon’s result pairs the highest aggregate score with lower token use than several close competitors, which could reduce the cost of long-running agents.

The lead costs fewer tokens

Vals reports an estimated cost of $15.68 per task and a median completion time of about 46 minutes under its test configuration.

The 46-minute figure measures completion time for benchmark tasks that can include extended reasoning and tool calls. Interactive request latency will vary by workload. Google’s Logan Kilpatrick reported the introductory token rates, so teams calculating production costs should use the prices available to their accounts.

Vals describes Argon as moving the efficiency frontier: every cheaper model in its comparison scored lower, and Argon cost substantially less than its nearest accuracy rivals. On Vals Index tasks, it used roughly one-quarter of Claude Sonnet 5.5’s output tokens and one-third of its input tokens for comparable work.

Agent loops shrink

On Tax Agent, Argon averaged 19 turns compared with Sonnet 5.5’s 47, while producing answers more than twice as long. On Finance Agent v2, it generated about one-fifth as many tool errors. Shorter agent loops reduce tool round trips, retries, and context growth, all of which affect latency and cost in production systems.

Legal, finance, and code set the pace

Across the 22 benchmarks reported by Vals, Argon placed among the top five on 20. Its strongest results covered several distinct workloads:

Four benchmarks expose gaps

• Computer use: Argon ranked seventh of eight models on CUA-bench with 4.83%, limiting the evidence for agents that operate graphical interfaces.
• Medical scribing: Its 87.4% MedScribe score placed 15th, showing that a high raw score can still trail a tightly clustered field.
• Program synthesis: A 2.5% ProgramBench score points to weak performance on formal code generation from specifications.
• Terminal work: Claude Opus 5.5 retained the Terminal-Bench 4.0 lead with 66.4%, compared with Argon’s 57.6%.

Inside the Vals run

Vals tested Argon through Google’s provider using temperature 1, default top-p and top-k settings, and high reasoning effort. Temperature, top-p, and top-k control how the model samples candidate tokens, while reasoning effort governs how much computation it devotes to solving a task. Changes to those settings can affect accuracy, token consumption, and latency.

The benchmark configuration provided a one-million-token context window and capped output at 262,000 tokens. Google’s announcement advertises continuous reasoning and generation trajectories of up to one million output tokens, up from 64,000 in the previous generation. The leaderboard therefore reflects the 262,000-token cap used by Vals rather than the model’s full advertised output ceiling.

Access starts behind a gate

Argon’s lead ends a run of Vals Index wins by flagship models from Anthropic and OpenAI. Gemini 4 also returns Google’s focus to its highest-capability tier after a year centered on faster, lower-cost Flash models.

Google is beginning a phased rollout with trusted cybersecurity partners while it evaluates the model’s safety with the U.S. government. Broader availability through the Gemini API is expected later, although Google has provided no release date.

Where Argon fits first

Teams running legal or financial document pipelines, code migrations, terminal agents, and other tool-heavy workflows have the clearest reasons to evaluate Argon. Its combination of domain accuracy, fewer agent turns, lower tool-error rates, and reduced token consumption could improve cost per successful task.

Computer-use agents, medical scribing, and formal program synthesis require separate evaluations because Argon trails on the corresponding benchmarks. A production bake-off should also measure end-to-end completion rate, tool reliability, latency, applicable token prices, rate limits, and data-governance requirements. Those results will determine whether the leaderboard advantage carries into a specific system.