{"slug": "google-deepmind-s-gemini-4-neon-targets-complex-agent-workflows-at-61-score", "title": "Google DeepMind's Gemini 4 Neon Targets Complex Agent Workflows at 61 Intelligence Score", "summary": "Google DeepMind released Gemini 4 Neon, a mid-tier paid model that scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, one point above Gemini 4 Carbon's 60 and behind Claude Fable 5.2 at 65 and GPT-6.2 Astra at 62. Neon is priced at $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 per million and a one-million-token context window, and Google claims 32% lower cost per task than Carbon and first place on the Artificial Analysis Coding Agent Index at 58.6. The model is available through the paid Gemini API, AI Studio, Antigravity and the Gemini app for Pro/Ultra subscribers with no free tier, while a cyber-focused variant rolls out to vetted defenders through the Fairwind Program.", "body_md": "# Google DeepMind's Gemini 4 Neon Targets Complex Agent Workflows at 61 Intelligence Score\n\nGoogle's new Gemini 4 Neon sits between Carbon and Hydrogen. It scores 61 on the Artificial Analysis Intelligence Index, just above Carbon, at $3/$15 per million tokens.\n\n- Google DeepMind released [Gemini 4 Neon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/?utm_source=alphasignal) , a mid-tier model sitting between Carbon and Hydrogen.\n- Neon scores 61 on the Artificial Analysis Intelligence Index, edging Carbon's 60 but trailing Fable 5.2 and GPT-6.2 Astra.\n- Priced at $3/$15 per million tokens. That's above Carbon's $2/$10 promo rate but below its $4/$20 standard rate.\n- Google claims 32% lower cost per task than Carbon and #1 on the Artificial Analysis Coding Agent Index (58.6).\n- Available on the paid Gemini API, [AI Studio](https://aistudio.google.com/docs/whats-new-gemini-4-neon?utm_source=alphasignal) , Antigravity and the Gemini app for Pro/Ultra subscribers. No free tier.\n- A cyber-focused variant rolls out to vetted defenders through the [Fairwind Program](https://deepmind.google/fairwind-program/?utm_source=alphasignal) .\n\nGoogle DeepMind has launched Gemini 4 Neon, a paid model built for frequent, complex work across software engineering, professional research, and science. It sits between the lower-cost Hydrogen model and the deeper-reasoning Carbon and Argon models, giving developers another option for balancing quality, latency, and cost.\n\nArtificial Analysis, a third-party model evaluator, gives Neon an Intelligence Index score of 61 at its highest `xhigh` thinking level. That places it third overall, behind Claude Fable 5.2 at 65 and GPT-6.2 Astra at 62, and one point above Gemini 4 Carbon at 60. Thinking levels control how much internal reasoning the model performs before answering, with higher settings generally increasing latency and billed tokens.\n\n## Neon slots above Hydrogen\n\nGoogle’s element-based Gemini 4 lineup now spans four models. Carbon and Argon handle deep reasoning, Hydrogen targets lower-cost everyday workloads, and Neon covers high-frequency tasks that still require substantial reasoning. The [developer docs](https://aistudio.google.com/docs/whats-new-gemini-4-neon?utm_source=alphasignal) position Neon for complex agent workflows.\n\nHydrogen scores 51 on the Artificial Analysis Intelligence Index at its default `high` setting and 54 at `xhigh`, compared with Argon’s 50. Neon reaches 61 at `xhigh`. In Google’s selected comparison table, Neon leads 16 of 19 reported results against the mid-tier Claude Opus 5.5 and GPT-6.2 Sol models.\n\n## Carbon’s promotion complicates the price\n\nNeon costs $3 per million input tokens and $15 per million output tokens. Cached input costs $0.15 per million tokens, a 95% reduction from its standard input rate. The model supports a one-million-token context window.\n\n| Gemini 4 API pricing and best reported Intelligence Index scores |  |  |  | \n|---|---|---|---|\n| Model | Input per 1M tokens | Output per 1M tokens | Best index score | \n|---|---|---|---|\n| Gemini 4 Hydrogen | $1.50 | $7.50 | 54 | \n| Gemini 4 Neon | $3.00 | $15.00 | 61 | \n| Gemini 4 Carbon, introductory | $2.00 | $10.00 | 60 | \n| Gemini 4 Carbon, standard | $4.00 | $20.00 | 60 | \n\nUnder Carbon’s introductory pricing, Neon costs 50% more for both input and output. Google says Carbon’s promotional rates run through November 30, after which they rise to $4 per million input tokens and $20 per million output tokens. Neon then becomes 25% cheaper per token.\n\nGoogle’s comparison estimates that Neon costs 32% less per completed task than Carbon at the `high` thinking level. Output length helps explain the difference: Artificial Analysis recorded about 200 million Carbon output tokens during a full index run, compared with a 64 million median across tested models. Workload-specific reasoning and output lengths will determine whether that advantage carries into production.\n\n## Agentic coding drives the benchmark case\n\n- **Coding agents:** Neon ranks first on the Artificial Analysis Coding Agent Index with a score of 58.6. It records 65.2% on Terminal-Bench 4.1 and 46.2% at`high` on FrontierCode 2.0. Claude Opus 5.5 scores slightly higher on those two individual benchmarks, while Neon leads the broader composite.\n- **Harness sensitivity:** Neon scores 12.8% on SWE-sweep with the mini-SWE harness and 39.0% with Google Antigravity. A harness is the software loop that gives a model tools, feeds back results, and manages multistep work. The gap shows how strongly agent scaffolding can affect benchmark performance.\n- **Professional work:** Neon ranks second on the Vals Index at 70.9% and sits on its Pareto frontier, meaning the evaluated set contains no model that is both cheaper and more accurate on that index.\n- **Science:** Google reports leading scores of 70.5% on Terminal-Bench Science, 84.0% on RiemannBench, and 71.3% on HLE-Diamond.\n- **Live agents:** Neon ranks second on Arena’s Agent Arena Index with a score of 1336 across more than 4,200 agent sessions.\n\n### Neon can delegate to Hydrogen\n\nGoogle designed Neon to run alone or coordinate parallel Hydrogen subagents in Antigravity. Neon can plan a long-running job, split it into smaller tasks, and assign those tasks to cheaper Hydrogen workers. Carbon previously handled that planning role at a higher reported cost per completed task.\n\n## Safety results depend on safeguards\n\n- **Benchmark cheating:** Google reports a 0.0% cheating rate on Terminal-Bench 4.0, matching Carbon.\n- **Shortcut behavior:** Neon took prohibited shortcuts on 2.45% of tasks in Google’s coding-cheating evaluation, the lowest reported rate among Gemini 4 models.\n- **False refusals:** The model rejected 1.98% of legitimate Terminal-Bench 4.0 tasks and 0% of Terminal-Bench Science tasks.\n- **Prompt injection:** Neon recorded a 0.0% attack success rate on the Gray Swan evaluation at`xhigh` with Google’s real-time safeguards enabled.\n\nIndirect prompt injection hides malicious instructions inside content that an agent reads, such as a web page, email, or file. Google says Neon resisted every attack in the tested Gray Swan configuration, matching results achieved only by other Gemini models. The result applies to that benchmark, thinking level, and safeguard configuration.\n\nGoogle also plans to distribute a cybersecurity variant, Gemini 4 Neon for cyber, to trusted defenders through the Fairwind Program. Offensive-capable features are reserved for screened organizations and excluded from the public API.\n\n## Migration details can break clients\n\nGoogle has released Neon through the paid Gemini API tier, AI Studio, Android Studio, Gemini for business, Antigravity, and the [Gemini app](https://gemini.google.com/?utm_source=alphasignal) for AI Pro and Ultra subscribers. The API has no free tier.\n\n### Gemini 4 swaps keep the schema\n\nExisting Gemini 4 integrations can switch to the `gemini-4-neon` model identifier without changing the request schema. A basic Python request looks like this:\n\n``` python\nfrom google import genai\n\nclient = genai.Client()\ninteraction = client.interactions.create(\n    model=\"gemini-4-neon\",\n    input=\"Audit this retry queue for race conditions.\",\n    generation_config={\"thinking_level\": \"high\"},\n)\n\nprint(interaction.output_text)\n```\n\nNeon defaults to `high` reasoning and supports `xhigh`, matching Hydrogen’s available range. Carbon and Argon default to `medium` and cap at `high`. Moving to `xhigh` can increase completion time and the number of billed thinking tokens.\n\n### Gemini 3 clients need four changes\n\n1. Remove `temperature` ,`top_p` , and`top_k` . Neon ignores values within the accepted range and returns an HTTP 400 error for values outside it.\n2. Remove explicit context caching. Neon supports implicit caching, which applies automatically to eligible prompts containing at least 4,096 tokens.\n3. Set client timeouts to at least 15 minutes or use background execution. Difficult prompts can run for several minutes.\n4. Declare tools through native function calling. Asking the model to write raw JSON tool calls can produce a `MALFORMED_FUNCTION_CALL` error.\n\nGoogle bills every token the model processes, including thinking tokens, grounding instructions, and function declarations inserted by the system. Evaluations at `xhigh` should account for those additional tokens as well as longer latency.\n\n## Independent testing remains thin\n\nMany launch figures come from Google’s own runs or its selected benchmark set. Artificial Analysis provides an external composite score, but broader independent tracking was incomplete at launch. OpenRouter’s comparison pages, which incorporate Artificial Analysis data, still listed Neon’s Intelligence Index as unavailable.\n\nProduction costs will depend on prompt length, reasoning level, tool use, retries, and successful task completion. A representative evaluation should record accuracy, latency, input tokens, output tokens, thinking tokens, tool failures, and total cost per completed task. Carbon’s temporary promotional rate also belongs in any near-term comparison.\n\nGoogle’s [launch post](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/?utm_source=alphasignal) and [model card](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-4-Neon-Model-Card.pdf?utm_source=alphasignal) contain the full benchmark tables and safety disclosures.", "url": "https://wpnews.pro/news/google-deepmind-s-gemini-4-neon-targets-complex-agent-workflows-at-61-score", "canonical_source": "https://alphasignal.ai/news/google-deepmind-s-gemini-4-neon-targets-complex-agent-workflows-at-61", "published_at": "2026-11-18 21:34:04+00:00", "updated_at": "2026-11-18 22:15:14.665581+00:00", "lang": "en", "topics": ["large-language-models", "ai-agents", "ai-products", "artificial-intelligence"], "entities": ["Google DeepMind", "Gemini 4 Neon", "Gemini 4 Carbon", "Gemini 4 Hydrogen", "Gemini 4 Argon", "Artificial Analysis", "Claude Fable 5.2", "GPT-6.2 Astra"], "also_reported_by": [], "takeaway": "Gemini 4 Neon scores 61 on the Artificial Analysis Intelligence Index at xhigh thinking versus Carbon's 60, at $3/$15 per million tokens with cached input at $0.15 per million and a one-million-token context window.", "recommendation": null, "alternates": {"html": "https://wpnews.pro/news/google-deepmind-s-gemini-4-neon-targets-complex-agent-workflows-at-61-score", "markdown": "https://wpnews.pro/news/google-deepmind-s-gemini-4-neon-targets-complex-agent-workflows-at-61-score.md", "text": "https://wpnews.pro/news/google-deepmind-s-gemini-4-neon-targets-complex-agent-workflows-at-61-score.txt", "jsonld": "https://wpnews.pro/news/google-deepmind-s-gemini-4-neon-targets-complex-agent-workflows-at-61-score.jsonld"}}