Google DeepMind's Gemini 4 Neon Ranks Second Globally at Half the Cost
Google's new Gemini 4 Neon takes #2 on the Vals Index at 70.9%, beating rivals on finance and legal tasks at under half the cost.
Takeaways
- Gemini 4 Neon hits #2 on Vals Index at 70.87%, trailing Claude Fable 5.2 by 0.4 points.
- Takes #1 on Vals Finance Index, Legal Index, Tax Agent Bench, and Legal Research Bench.
- Priced at $3/$15 per million input/output tokens with 1M context and 1M max output.
- Sits on the score-cost Pareto frontier on 7 of 14 Vals leaderboards.
- Uses roughly a third fewer tool calls than Hydrogen on Legal Research while writing longer answers.
- Available via Gemini API, Google AI Studio, Gemini app, and Antigravity.
Google DeepMind has released Gemini 4 Neon, which ranks second on the Vals Index with a score of 70.87%. It trails Anthropic’s Claude Fable 5.2 by 0.4 percentage points while costing less than half as much per task. Neon also leads the Vals Finance and Legal indexes.
Vals evaluates models across professional and agentic tasks, including research, coding, tax, finance, and legal work. Its results suggest Neon offers a strong price-performance balance for developers building tool-using systems, although individual benchmark gains vary by workload and reasoning setting.
Google places Neon between Carbon and Hydrogen in its element-named Gemini 4 lineup. According to the launch announcement, the model targets complex software engineering, knowledge work, and scientific tasks. Neon is Google’s fourth model release in seven weeks.
The price-performance edge
| Usage | Price per million tokens |
|---|---|
| Input | US$3 |
| Cached input | US$0.15 |
| Output | US$15 |
Among models with reported costs, Neon sits on the score-cost Pareto frontier for seven of the 14 Vals leaderboards where it appears. Those include the main Vals Index, the Finance and Legal indexes, Earnings Modeling Bench, Legal Research Bench, Tax Agent Bench, and Vibe Code Bench. In practical terms, no cheaper model on those boards matches its reported score.
Compared with Gemini 4 Carbon, released eight days earlier, Neon delivers similar performance at 17% to 32% lower cost, depending on the benchmark. Its specifications include a 1-million-token context window, a maximum output allocation of 1 million tokens, and support for text, images, video, and files as input.
Fewer searches, longer answers
Vals’ agent benchmarks show Neon using fewer tools than earlier Gemini 4 models on research-heavy tasks. On Legal Research Bench, Gemini 4 Hydrogen averaged about 220 tool calls and 88 case-law searches per question. Neon reduced both figures by roughly one-third while producing final answers averaging about 4,900 words.
On Tax Agent Bench, Neon averaged 32 web searches per task, compared with 59 for Hydrogen, and still produced longer responses. The results indicate that Neon can consolidate research with fewer retrieval steps, which can reduce latency and tool-related costs in production agents.
High versus xhigh
Vals evaluated Neon at Google’s high and xhigh reasoning levels. The Gemini 4 lineup supports four effort settings: low, medium, high, and xhigh. Higher settings allocate more computation to a request, usually increasing runtime and cost in exchange for deeper reasoning.
The high setting retained 98% of Neon’s xhigh Vals Index score at roughly half the cost and 40% less runtime, ranking sixth overall. On a question about Centene’s insurance rebates, xhigh produced a 1,177-word explanation. High led with the requested answer, earned full marks, and used 842 words.
Additional reasoning produced the largest gains on tax and coding tasks. At xhigh, Neon ranked first on Tax Agent Bench, scoring 7.3 points above its high configuration. It also ranked second on Vibe Code Bench, less than 0.1 point behind the leader. On Finance Agent, the high setting slightly outscored xhigh and ranked second.
Efficiency has limits
Neon’s shorter workflows did not improve every engineering benchmark. On Code Migration, it averaged 160 turns per task, compared with 234 for Carbon and 574 for Hydrogen, yet scored below both models. In this benchmark, turns represent steps in the model’s tool-driven workflow.
The result points to a workload-specific trade-off: reducing agent steps can lower cost and latency, while long-running migrations may benefit from the additional exploration performed by Carbon or Hydrogen. Teams evaluating repository-scale changes should test the models against representative codebases and tool configurations.
Leaderboard at a glance
| Benchmark | Score | Rank |
|---|---|---|
| Vals Index | 70.87% | 2 of 42 |
| Vals Finance Index | 68.31% | 1 of 45 |
| Vals Legal Index | 78.36% | 1 of 43 |
| Legal Research Bench | 56.25% | 1 of 99 |
| Tax Agent Bench | 54.89% | 1 of 90 |
| Vibe Code Bench v1.1 | 93.70% | 2 of 133 |
| BioMysteryBench | 80.00% | 4 of 34 |
| Terminal-Bench 4.0 | 61.62% | 7 of 69 |
Access and availability
Google launched Neon without a waitlist. Availability spans several products and subscription tiers:
- Gemini API: Access through Google AI Studio.
- Android Studio: Integration for development workflows.
- Gemini app: Access for AI Pro and Ultra subscribers.
- Antigravity: Access for paid subscribers using Google’s agentic development environment.
- Fairwind Program: Access planned for members, with no release date specified.
Artificial Analysis gives Neon a score of 61 on its Intelligence Index, one point above Gemini 4 Carbon. That places Neon third behind GPT-6.2 Astra and Claude Fable 5.2. The Artificial Analysis index uses a separate test suite, so its score cannot be compared numerically with the 70.87% Vals result.
Choosing a deployment setting
The high reasoning level provides the strongest starting point for production trials because it preserves most of the xhigh score while reducing cost and runtime. Tax reasoning and agentic coding workloads may justify xhigh, based on the larger benchmark gains in those categories.
Legal research, earnings modeling, tax analysis, and short-horizon coding agents align most closely with Neon’s strongest results. Repository-scale migrations and workflows that depend on long, stateful tool sequences warrant direct comparisons with Carbon and Hydrogen. Final model selection should account for application-specific accuracy, latency, token use, tool-call costs, and failure recovery.