- Runs at roughly 110 tokens per second, nearly 2x Carbon, with lower base pricing
- Beats Anthropic's competing model on coding at a fraction of the cost
- Model card claims 100% blocking of indirect prompt-injection attacks on Gray-Swan
AI CURATED HOT STORY · 11/19/2026
Google Launches Gemini 4 Neon with 69% Lower Cost per Task
Google has launched Gemini 4 Neon, the fourth model in its Gemini 4 series, positioned as a mid-tier agentic model focused on fast end-to-end task completion and cost efficiency. On the Arena Agent Index, Neon (High) ranks second overall with a score of 1336, achieving a median cost per task of just $2.82—69% lower than the top-ranked Claude Fable 5.2 and 53% lower than Claude Opus 5.5—while placing in the top three for user satisfaction, steerability, and task completion.
Technically, Neon supports a 1 million token context window and four reasoning levels, with API pricing at $3 per million input tokens and $15 per million output tokens, and cached input at $0.15. It tops the Artificial Analysis Coding Agent Index with a score of 58.6, surpassing Claude Opus 5.5, and can orchestrate multiple Hydrogen sub-agents in parallel within the Antigravity development environment. The model runs at about 110 tokens per second, twice as fast as its predecessor Carbon, and blocks 100% of indirect prompt injection attacks on the Gray-Swan benchmark.
The release signals Google's accelerating iteration cadence, with intervals shrinking from 28 days to 6 days, and a rapidly consolidating product line. By offering near-frontier agentic performance at mid-tier pricing, Neon directly challenges Anthropic and OpenAI, potentially reshaping cost-performance competition in the agent market and accelerating enterprise AI adoption.
Hourly heat curve
18 fully observed hoursThe line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.
Coverage timeline
Total 16 reports · Latest first- The DecoderT2·78 pts
- Latent SpaceT1·78 pts
- Four Gemini 4 models shipped in about seven weeks with shifting naming and tiers
- Neon targets the mid-tier with better speed and cost than Carbon, beating it on some benchmarks
- Argon retired and legacy Pro/Flash discontinued as the lineup consolidates
- Marktechpost (@Marktechpost)T2·78 pts
- Neon fills the mid-tier gap between Carbon and Hydrogen, defaulting to high thinking with an xhigh option
- Acts as an orchestrator over parallel Hydrogen subagents in Antigravity
- Self-reported results lead Opus 5.5 by wide margins on GraphWalks and SWE-sweep but trail slightly on Terminal-bench and FrontierCode
- Closed weights at $3 input and $15 output per 1M tokens
- GREG ISENBERG (@gregisenberg)T2·58 pts
- Gemini 4 Neon debuts at No. 1 on the Artificial Analysis Coding Agent Index
- Coding performance is roughly on par with Opus 5.5 at about half the cost per task
- Speed and price appeal to all-day agent users, though Claude is still preferred for the hardest problems
- 宝玉 (@dotey)T2·78 pts
- Neon scores 61 on Artificial Analysis' composite index, ranking third and nearly matching in-house Carbon (60) while being cheaper and faster
- API pricing is $3 per million input tokens and $15 per million output tokens, with cached input at $0.15, roughly 40% of GPT-6.2 Astra and Claude Fable 5.2
- The Gemini 4 family has shipped four variants in seven weeks (Argon, Carbon, Hydrogen, Neon), with Argon's overlapping positioning and deprecation status left unclear
- Hacker News AIT2·58 pts
- Gemini 4 Neon hits 127 tokens/s, roughly 2.8x Carbon's speed
- Benchmark gains are incremental; speed is the headline improvement
- Projected win rate against Opus 5.5 is about 48.3%, placing them in the same tier
- IT之家 · 智能时代T2·88 pts
- Ranks first on Artificial Analysis' coding agent index at 58.6, ahead of Claude Opus 5.5
- Supports a 1M-token context window spanning spreadsheets, PDFs, charts, audio and video
- Offers four reasoning tiers from Low to xhigh, with cached input at $0.15 per million tokens
- Acts as an orchestrator in Antigravity, coordinating Hydrogen sub-agents on long-running tasks
- Artificial AnalysisT1·62 pts
- Neon scores 61 on the Intelligence Index, just one point above Carbon and third overall
- Cost per task falls ~17% and time ~40%, yet output tokens rise ~15%
- The cost advantage stems mainly from a 25% price cut, not efficiency gains
- Arena.ai (@arena)T1·78 pts
- Release gaps shrank from 28 days to 6, signaling near-exponential iteration speed
- Gemini 4 Neon (High) ranks #2 on the Agent Arena Index at 1336 points, reshaping the Pareto frontier
- Neon costs $2.82 median per task, 53% less than Claude Opus 5.5 while scoring higher
- Artificial Analysis (@ArtificialAnlys)T1·74 pts
- Gemini 4 Neon scores 61 on the Intelligence Index, one point above Carbon and third overall
- Cost per task falls to $6.44, about 17% lower, with time per task down roughly 40% to about 21 minutes
- Savings come from $3/$15 per 1M token pricing rather than fewer tokens, since output tokens rise about 15%
- Rohan Paul (@rohanpaul_ai)T2·78 pts
- Scores 61 on the Artificial Analysis Intelligence Index for third place, priced at $3/$15 per 1M tokens
- Costs 32% less per task than Gemini 4 Carbon at default high setting, roughly 50-80% less than top rivals
- Ranks first on the Coding Agent Index at 58.6, with SWE-sweep rising from 12.8% to 39.0% inside Antigravity
- OpenRouter (@OpenRouter)T1·74 pts
- Gemini 4 Neon is live on OpenRouter via Google Cloud, positioned between Carbon and Hydrogen
- Google claims the top spot on the Artificial Analysis Coding Agent Index
- Google reports 32% lower cost per task than Carbon
- Arena.ai (@arena)T1·78 pts
- Gemini 4 Neon (High) ranks #2 overall on Agent Arena with 1336 points, Google DeepMind's best placement on the index
- Median cost of $2.82 per task is 53% below Claude Opus 5.5 (Max) and 69% below Claude Fable 5.2 (Max)
- Top-three across praise-vs-complaint (#2), steerability (#2), and task completion (#3)
- Google DeepMindT1·88 pts
- Gemini 4 Neon scores 61 on the Artificial Analysis Intelligence Index, ranking #3 overall behind larger, slower models
- Ranks #2 on the Vals Index on the Pareto frontier and #2 on Agent Arena steerability
- Targets fast end-to-end task completion in software engineering, knowledge work, and science workflows
- Hacker News AIT2·88 pts
- Gemini 4 Neon scores 61 on the Artificial Analysis Intelligence Index, ranking third overall behind larger, slower models
- It ranks second on the Vals Index and sits on its Pareto frontier, and second on the Agent Arena steerability ranking
- Positioning emphasizes fast end-to-end task completion across software engineering, knowledge work, and science workflows
- The DecoderT2·78 pts
- Gemini 4 Neon is live in the API at $3/$15 per million tokens, above Carbon and Argon's launch pricing and double Hydrogen
- Google positions it as frontier-level intelligence for complex software engineering, knowledge work, and science workflows
- It is Google's fourth Gemini 4 model in four weeks, with benchmarks still pending