Google Releases Gemini 4 Neon and Omits the Top Two Models From Its Benchmark Comparison Table
The fourth Gemini 4 model in 49 days lands third on the intelligence index — behind rivals Google left off.

Google DeepMind launched Gemini 4 Neon on November 18, 2026, adding the fourth member of its Gemini 4 model family and extending a 49-day run of successive releases. The new model joins Gemini 4 Argon, Carbon, and Hydrogen in what has become an unusually compressed product cycle. Neon sits in the mid-tier — faster and more affordable than Carbon, positioned for enterprise, coding, and cybersecurity tasks.
But Google's own benchmark comparison table, published alongside the launch, omits the two models that rank above Neon on the Artificial Analysis Intelligence Index v5.0.1, the leading independent evaluation of large language models. Claude Fable 5.2 leads the index with a score of 65; GPT-6.2 Astra is second at 62. Neon scores 61 and ranks third overall. Neither Fable 5.2 nor GPT-6.2 Astra appears in Google's comparison.
Google also announced that Genie 4, the company's generative interactive environment model first introduced in October, was rolling out to Gemini Ultra subscribers around the same time as the Neon launch.
How Neon Fits Into the Gemini 4 Family
Google has released four Gemini 4 models in rapid succession: Argon on September 30, Carbon on October 28, Hydrogen on November 12, and now Neon on November 18. Each occupies a different position in the performance-price hierarchy.
Neon is priced at $3 per million input tokens and $15 per million output tokens — substantially below the rates charged for frontier-class alternatives. Artificial Analysis, which independently measures model performance and speed, recorded Neon's output at 110 tokens per second, roughly double Carbon's measured rate. That throughput matters for latency-sensitive enterprise applications, particularly in legal document review, automated coding, and high-volume agent workflows.
Google positions Neon as a middle layer: more capable than its lighter Gemini siblings, faster and cheaper than its own frontier models, and purpose-built for sustained professional use rather than single-query consumer tasks.
What the Benchmarks Actually Show
The selective comparison table that accompanied Neon's launch highlights scores on Terminal-Bench 4.1, Vals AI's enterprise benchmark suite, and AutoBaxBench hard tier — areas where Neon performs competitively. On Terminal-Bench 4.1, Neon scores 65.2% against Anthropic's Claude Opus 5.5 at 65.8%. The results are close, and the table omits Claude Fable 5.2, which ranks first on the broader intelligence index.
On the Artificial Analysis Intelligence Index v5.0.1, which aggregates scores across reasoning, coding, mathematics, and language tasks, Neon at xhigh thinking registers 61 points. That places it third overall — four points behind Claude Fable 5.2 and one point behind GPT-6.2 Astra. Neither model appears in Google's published comparison.
On the Vals AI enterprise index, which evaluates models on legal, financial, and business-domain tasks, Neon scores 70.9%, placing it second. Claude Fable 5.2 leads that index at 71.3%, a gap of 0.4 percentage points.
On the Artificial Analysis Coding Agent Index v1.1, Neon scores 58.6% — competitive at the high end of the mid-tier range, though Google's model card claim of a first-place ranking on this index warrants scrutiny, as Carbon scored 59.4% at high effort, suggesting configuration differences may affect where individual models land.
Read more: Artificial Analysis intelligence index leaderboard
Thinking Levels and What They Mean for Developers
Neon supports four thinking levels — low, medium, high, and xhigh — giving developers direct control over how much compute the model spends on each request. The Gemini API thinking documentation describes each tier and its associated token-budget behavior.
The xhigh setting is what produces Neon's highest benchmark scores, including the 61-point Artificial Analysis result. At lower thinking levels, latency drops and cost falls, but benchmark performance also declines. Developers targeting cost-sensitive pipelines can configure low or medium thinking; those prioritizing accuracy in complex reasoning tasks can use high or xhigh.
Google's developer documentation on thinking token budgets allows per-request configuration, so individual API calls can mix thinking levels across a single application. This granularity is useful for tiered workflows where some tasks demand maximum reasoning and others require only retrieval or summarization.
Neon's context window is 1,048,576 tokens — just over one million tokens — consistent with the rest of the Gemini 4 family and suited to long-document legal analysis, large-codebase agent tasks, and extended reasoning chains.
Neon for Cyber: A Cybersecurity Variant With Restricted Access
Google has also introduced a restricted version of Neon called Neon for Cyber, designed specifically for cybersecurity research and incident response. Access is gated through the Google Fairwind Program, which requires prospective users to apply and obtain authorization before the model is made available to them.
On AutoBaxBench's hard tier, Google reports Neon for Cyber scores 76%. AutoBaxBench evaluates models on binary exploitation, vulnerability analysis, and reverse engineering tasks. Google has not yet published independent third-party evaluation of this score, and external replication has not been completed at the time of publication.
The Fairwind access gate is consistent with how Google has previously managed high-risk model capabilities — limiting access to vetted research organizations while retaining the ability to monitor usage. Neon for Cyber is not available through the standard Gemini API.
Pricing Against the Competitive Mid-Tier
At $3 input / $15 output per million tokens, Neon sits in a competitive segment of the market. That pricing positions it well below frontier-tier models from Anthropic and OpenAI, while still commanding a premium over lightweight inference options. For enterprise customers processing large document volumes or running continuous agent loops, the per-token cost relative to output speed — 110 tokens per second — makes Neon a practical option for sustained workloads.
The Vals AI enterprise benchmark result of 70.9%, placing Neon second in a category defined by legal research, tax analysis, and financial document tasks, adds independent evidence for its positioning in professional services use cases. The 0.4-percentage-point gap behind Claude Fable 5.2 on that index is narrow, though Fable 5.2's higher score and first-place standing on the broader intelligence index give it a composite advantage across both general and enterprise evaluations.
Genie 4 Opens to Ultra Subscribers
Google's Genie 4 generative interactive environment model, first introduced in October 2026, began rolling out to Gemini Ultra subscribers around the time of the Neon launch. Genie 4 is designed to generate interactive environments — playable spaces, navigable 3D worlds, and dynamic simulations — from text or image prompts. It uses Gemini 3.8 Flash as its guiding agent, controlling behavior and narrative within the generated environment.
According to Google, the model is capable of producing coherent, interactable environments that persist across user actions. The rollout to Ultra subscribers makes Genie 4 available to paying users for the first time, though broad general availability has not yet been announced.
What Comes Next
Google has now released four Gemini 4 models in 49 days. Whether the pace continues into December — a fifth model, or capability updates to existing ones — has not been publicly indicated. The Neon launch coincides with broad industry pressure to maintain visible benchmark positioning, and the selective presentation in Google's comparison table is a reminder that vendor-published benchmarks should be read alongside independent evaluations from sources such as Artificial Analysis and Vals AI.
For developers and enterprise buyers evaluating Neon, the practical question is not whether it beats every model in every test — it does not — but whether its combination of throughput, context length, thinking-level flexibility, and mid-tier pricing fits their specific workloads better than the alternatives it was compared against and those it was not.



