Google Ships Gemini 4 Neon, a Cheaper Near-Carbon Model at $3 Per Million Tokens
Google's Gemini 4 Neon costs $3 per million input tokens, ranks #3 on Artificial Analysis at 61, and tops legal, tax, and automation benchmarks with a 1M-token context window.

Why it matters
- Google released Gemini 4 Neon on Wednesday at $3 per million input tokens and $15 per million output tokens, with a 1M-token context window.
- Neon scores 61 on the Artificial Analysis Intelligence Index at xhigh thinking level, ranking #3 overall behind larger models.
- Neon ranks #1 on Legal Research Bench, Tax Agent Bench, Zapier AutomationBench (56.3%), and Vending-Bench 2 ($16,843).
- Neon is Google's fourth Gemini 4 release after Argon, Carbon, and Hydrogen, positioned as near-Carbon performance at lower cost.
- Cached input tokens are discounted 95% to $0.15 per million.
Google released Gemini 4 Neon on Wednesday, pricing the new mid-tier model at $3 per million input tokens and $15 per million output tokens — a deliberate play to deliver near-flagship performance at a fraction of the cost of its larger sibling, Gemini 4 Carbon.
The pricing details matter as much as the launch itself. Cached input tokens cost $0.15 per million, a 95% discount, and the model ships with a 1 million-token context window. That combination — long context, cheap cached reads, near-top capability — is aimed squarely at enterprises running high-volume agent and document workloads, where token economics decide which model a company can actually afford in production.
Where does Neon land on the benchmarks?
Neon enters a crowded leaderboard with respectable positioning. Google's announcement states it plainly: "Neon ranks ahead of comparable models and #3 overall with a score of 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, landing just behind larger, slower models."
That #3 placement, just behind bigger and slower competitors, is the core of the pitch. Neon is not trying to beat the flagships on raw capability. It's trying to get close enough that the price difference makes the decision for you. On the Vals Index it ranks #2 and sits on the Pareto frontier — the set of models where you can't get better performance without paying more elsewhere.
The full launch-day scorecard Google published:
- Artificial Analysis Intelligence Index: #3 overall, score of 61 at the xhigh thinking level
- Vals Index: #2, at 70.9%
- Legal Research Bench: #1
- Tax Agent Bench: #1
- Surge AI: #1 on most benchmarks across finance, accounting, spreadsheet modeling, and complex constraint following
- Chartography: 73.5%, for reading dense professional charts
- Agents' Last Exam v1.1: 75.8%, on par with leading models
- Zapier's AutomationBench: #1 overall at 56.3%
- Andon Labs' Vending-Bench 2: #1 at $16,843, a long-horizon business simulation
- Agent Arena Index: #2 globally with a score of 1336, across 4,200+ real-world agent sessions
- Terminal-Bench Science: 70.5%
- RiemannBench: 84.0%
- HLE-Diamond: 71.3%
The short version: Neon wins on specialized work — law, tax, spreadsheets, chart reading, automation — and stays competitive on general reasoning, where it trails only larger models.
Why is Google releasing models this fast?
Neon is the fourth Gemini 4 release in quick succession, following Argon, Carbon, and Hydrogen. That cadence is unusual even by the standards of the current AI release cycle, and it signals how Google is structuring its lineup: a family of models at different price and capability points rather than a single flagship with everything behind it.
The context here is competitive. Google has been working to catch up to Anthropic and OpenAI, particularly in the segments where those rivals have built strong enterprise mindshare. The Gemini 4 family is, by the source's assessment, "a step in the right direction." Neon fills a specific gap: users who want near top-of-the-line performance but need the efficiency and lower cost that a mid-tier model provides.
For buyers, a four-model family means more granular trade-offs. You can pick the cheapest model that clears your quality bar instead of paying flagship rates for every workload.
What does Neon actually do for enterprises?
Google positions Neon as a model for "software engineering, knowledge work, and science workflows" — three categories that cover a lot of expensive human labor. The enterprise pitch leans heavily on multimodal synthesis. As Google puts it: "Enterprise work lives in spreadsheets, PDFs, charts, audio, and video. Neon can synthesize across all of these inputs in a single 1 million-token context window and turns inputs into structured deliverables."
That single-window capability is the practical selling point. Instead of chaining multiple tools or models to process a mixed stack of documents, a company can feed everything into one context window and get structured output back. The #1 finishes on the Legal Research Bench and Tax Agent Bench suggest Google tuned the model for exactly this kind of document-heavy professional work.
The automation results point in the same direction. Winning Zapier's AutomationBench at 56.3% and Andon Labs' Vending-Bench 2 at $16,843 — a simulation that tests whether a model can run a business over a long horizon without losing the plot — indicates that Neon holds up on extended, multi-step tasks rather than just one-shot question answering.
How well does it behave as an agent?
Raw benchmark scores are one thing. Working with a model day to day is another, and Google made a point of highlighting usability alongside capability.
"Beyond just pushing on raw model capability, we're continuing to make our models better to work with," Google said in its announcement. "Neon ranks #2 on steerability (adapting cleanly to mid-task user corrections) on the new Agent Arena Index."
Steerability matters more as agents take on longer tasks. If a model can absorb a mid-course correction from a user without restarting or drifting, it's usable in real workflows. Neon also, per Google, "excels at multi-step instruction following and builds on our ongoing progress to curb fabrication, recording our lowest fabrication rate yet in internal Google Antigravity testing."
That fabrication claim deserves a caveat: it comes from Google's own internal testing, not an independent evaluation. Still, a falling hallucination rate is a meaningful trend for enterprise adoption, where a confident fabricated citation or number can be costly.
What about science and math?
The scientific and mathematical results are among the strongest in the launch. Google says Neon is "best in class on Terminal-Bench Science (70.5%), RiemannBench (84.0%), and HLE-Diamond (71.3%)," covering quantitative analysis, mathematics, and technical research problems.
HLE-Diamond is a stripped-down, harder subset of Humanity's Last Exam, one of the toughest general reasoning benchmarks currently in circulation. Scoring 71.3% there from a mid-tier price point is the kind of number that narrows the practical gap between frontier and mid-tier models. Combined with the #3 Artificial Analysis placement, it suggests the capability-per-dollar curve has shifted again.
What's the takeaway?
Gemini 4 Neon gives buyers a near-flagship model at mid-tier prices, with verified wins in legal, tax, financial, and automation benchmarks — and its real-world impact will hinge on whether independent evaluations confirm Google's internal fabrication and steerability claims as enterprises put the model into production.
Source: webpronews.com
More from Sophie Lindqvist
Show full bio
Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.
823 articles
Related articles
- Gemini 4 Neon runs at 110 tokens per second, undercuts Claude
- Google Opens Gemini 4 Neon to All Developers at $3/$15 per Million Tokens
- Gemini 4 Neon cuts task cost while intelligence gain stays modest
- Google Launches Gemini 4 Hydrogen at $1.50 Per Million Input Tokens
- Google Opens Gemini 4 Carbon to All Paying Users at $2 Per Million Tokens