
INPUT/OUTPUT PER M TOKENS
$3/$15
SWE SCORE IN ANTIGRAVITY
3×
CODING AGENT INDEX
#1
CHEAPER PER TASK
32-50%
Google has released Gemini 4 Neon, a frontier-class language model positioned as faster and cheaper than its predecessor Carbon while maintaining comparable safety guardrails. The model scores #3 on the Artificial Analysis Intelligence Index (61 at xhigh thinking), #2 on the Vals Index (70.9%), and #1 on the Artificial Analysis Coding Agent Index (58.6). It ships with a 1-million-token context window at $3 per million input tokens and $15 per million output tokens, with cached input tokens discounted 95% to $0.15 per million.
Extraction Map
The most revealing number in the announcement is the SWE-sweep benchmark split. On the standard mini-SWE harness, Neon scores 12.8%. Inside Google Antigravity — the company's proprietary agentic coding environment — that figure triples to 39.0%, a claimed state-of-the-art result. Google frames this as showing 'what the Antigravity harness unlocks.' Read differently, it shows that Neon's headline coding result is inseparable from Google's own tooling. Developers chasing that 39% will need to build on Antigravity, not just use the model. The benchmark sweep is impressively broad: #1 on Terminal-Bench Science (70.5%), RiemannBench (84.0%), HLE-Diamond (71.3%), Legal Research Bench, Tax Agent Bench, AutomationBench (56.3%), Vending-Bench 2 ($16,843), and SEC-bench Pro (90.5%). Neon also debuts at #2 on the new Agent Arena Index (1336 score across 4,200+ sessions) and #2 on steerability. The cybersecurity numbers are particularly strong: 76% on AutoBaxBench hard tier, 51.26% on WebExploitBench v1.2, and 90.5% on SEC-bench Pro — all claimed at roughly half the cost of competing models. Pricing is the strategic lever. At $3/$15 per million input/output tokens and a 32% lower cost-per-task than Carbon at the default thinking level, Google is explicitly competing on the Pareto frontier of capability versus cost. The four-tier thinking system (with xhigh adding deeper reasoning at 36%+ less than competitors' highest settings) gives teams granular control over their token budgets. This is a land-and-expand play: get developers building on cheaper inference, then upsell them into Antigravity for the benchmark-tripling harness. Safety claims are extensive and specific. Neon reports a 0.0% cheating rate on Terminal-Bench 4.0, a 2.45% shortcut rate on coding evaluations (lowest of any Gemini 4 model), 0.09% misaligned action rate on the agentic safety suite, and 99.2% on agent honesty evaluation. It matched a 0.0% attack success rate on Gray Swan's indirect prompt injection benchmark at highest thinking level. Fabrication rates are reported as the lowest yet in internal testing. Under-service dropped 43-56% and overreach dropped 15-35% versus Carbon. The announcement is a product launch dressed as a technical report. Every benchmark is selected to highlight Neon's strengths. The 'internal Google Antigravity testing' for fabrication rates means Google is grading its own homework on the metric most likely to erode user trust. The cyber capabilities — vulnerability finding, exploitation validation, patching — are gated behind the Fairwind Program for 'trusted defenders,' creating a controlled distribution channel for dual-use capability. What emerges is a model that is genuinely capable and aggressively priced, but architecturally designed to deepen Google ecosystem lock-in. The 3x performance multiplier inside Antigravity is not a neutral benchmark result — it is a business strategy. Developers who optimize for Neon's best performance will find themselves building inside Google's walls, paying Google's inference costs, and dependent on Google's harness for their production metrics.
Worth reading?
OPTIONAL
Score
56
/100
Mixed — genuine capability gains and lower prices expand developer access, but the architecture routes peak performance through proprietary tooling, making this more extractive than a pure API release.
What Is GRIN?
GRIN stands for Generativity, Resilience, Innovation efficiency, and Novelty openness. These four dimensions capture the essential characteristics of any memetic intelligence — whether a nation, a corporation, a religion, or a movement.
About GRINGenerativity
3/5
Resilience
2/5
Efficiency
4/5
Novelty
3/5
Extractive
3/5
Inverted
Does this create new capability or just redistribute existing value?
Real capability gains at lower cost expand access for developers and enterprises. But the 3x Antigravity multiplier means peak value creation is gated behind Google's proprietary harness.
Verified Facts (17)
Neon ranks #3 overall with a score of 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level
On the Vals Index, Neon ranks #2 (70.9%)
Neon ranks #2 on steerability on the new Agent Arena Index
Priced at $3 per million input tokens ($0.15 per million cached input tokens, a 95% discount) and $15 per million output tokens, with a 1 million-token context window
On SWE-sweep, it already leads on the mini-SWE harness (12.8%) and more than triples to a state-of-the-art 39.0% in Google Antigravity
#1 overall on the Artificial Analysis Coding Agent Index (58.6), with a score of 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0
Terminal-Bench Science (70.5%), RiemannBench (84.0%), and HLE-Diamond (71.3%)
State-of-the-art on AutoBaxBench hard tier (76%)
WebExploitBench v1.2 state-of-the-art (51.26%) at half the cost of competing large models
SEC-bench Pro (90.5%)
0.0% cheating on Terminal-Bench 4.0 and 2.45% shortcut rate on coding cheating evaluation
0.09% misaligned action rate on agentic safety suite and 99.2% pass rate on agent honesty evaluation
Debuts #2 globally with a score of 1336 on Arena's new Agent Arena Index across 4,200+ real-world agent sessions
Zapier's AutomationBench (56.3%) and Andon Labs' Vending-Bench 2 ($16,843)
32% lower cost per task than Carbon at default high thinking level
CWE-bench v2 70.7%, driving a 3.4% uplift over Carbon
Reducing under-service by 43%–56% and overreach by 15%–35% across different effort levels
More latest news
Read all