Google's New AI Model Undercuts Anthropic on Price While Matching Flagship Performance
Gemini 4 Neon delivers Opus 5.5-level coding capabilities at 25% lower cost, positioning Google to capture cost-conscious enterprise customers.

Google has unveiled Gemini 4 Neon, a mid-range large language model that matches the coding performance of Anthropic's premium Opus 5.5 while undercutting it by 25% on API pricing. The release positions Google to capture cost-conscious enterprise customers seeking sophisticated AI capabilities without flagship-tier expenses.
Neon is now available through Google AI Studio, Android Studio, the Gemini app for Pro and Ultra subscribers, and Google's Antigravity platform. The model is priced at $3 per million input tokens and $15 per million output tokens, compared to Opus 5.5's $4 and $20 respectively, though it remains 50% more expensive than Claude Sonnet 5.5 and GPT-6.2 Sol at $2/$10.
Competitive Benchmark Performance
Neon scored 70.9% on the Vals professional tasks evaluation, placing second overall across tested models. On the FrontierCode 2.0 coding benchmark, it achieved 46.2% compared to Opus 5.5's 47.3%, demonstrating near-parity with Anthropic's flagship offering. The model notably outperformed Opus on Zapier's AutomationBench, scoring 56.3% versus Opus at 42.5% and Sol at 41.3%.
The model placed third on the Artificial Analysis Intelligence Index at its highest reasoning setting, edging past Google's own premium Carbon model. On domain-specific evaluations, Neon ranked first on Vals legal research and tax agent tasks, and scored 70.5% on Terminal-Bench Science compared to Sol's 68.0% and Opus's 63.3%.
Thai Tran, AI product lead at Glean, reported that Neon matched or exceeded Carbon's quality at three times the speed and 60% of the cost. Google itself reports a 32% cost reduction per task versus Carbon at default settings.
Speed and Agent Architecture
Neon streams approximately 110 tokens per second at its highest speed setting, 1.9 times faster than Carbon's 59 tokens per second, though it carries a 57-second time-to-first-token latency. The model's performance scales dramatically when deployed with optimized agent architectures: on the SWE-sweep coding evaluation, Neon scored 12.8% with a basic mini-SWE agent but jumped to 39.0% when paired with Google's Antigravity multi-agent system—a threefold improvement.
Enterprise Applications
Google and early users highlight several enterprise applications for Neon, including tracing unexpected charges across billing records, finding flaws in unfamiliar software, testing research hypotheses, and building spreadsheets that colleagues can inspect and change. These tasks require the model to follow evidence through several steps and produce artifacts that other people can use.
Thoughtful Lab founder Karina Nguyen said Neon helped her team understand complex software architectures quickly and spotted subtle vulnerabilities that other models missed in security evaluations. Cartwheel co-founder and chief scientist Andrew Carr described using Neon in machine-learning research, where it examined prior work, tested hypotheses, and incorporated feedback as experiments progressed.
Google positions Neon between its lightweight Hydrogen and heavyweight Carbon models, targeting users who need professional-grade capabilities without flagship pricing. The timing gives Google an opening with developers evaluating OpenAI's new GPT-6.2 models, as early users have reported challenges with Astra stopping early on end-to-end builds despite strong benchmark performance.
Sources
Follow Health AI Central
Keep up with health AI where you already read
On LinkedIn
The day’s strongest health AI stories in your feed, one or two each weekday.
On Google
Choose Health AI Central as a preferred source, and Google is more likely to show our reporting in Top Stories.
Reporting & accountability
How we maintain accuracy
Health AI Central separates reported facts, analysis, and public-source data. Readers can review our standards, request a correction, or consult the relevant research methodology before relying on a claim.