Released on November 19, 2026, Gemini 4 Neon is Google DeepMind’s new mid-tier reasoning model in the Gemini 4 family, positioned between Hydrogen and Carbon. It targets complex software engineering, knowledge synthesis, and scientific reasoning workflows. Its published pricing is $3 per million input tokens and $15 per million output tokens, paired with a 1,048,576-token context window and support for xhigh reasoning mode. Google’s benchmark results show clear strengths in specific domains, though the model does not outperform Claude Opus 5.5 across all evaluation tasks.
Before replacing a flagship model with Neon, teams should first map their workload characteristics. Benchmark scores are useful reference points, but end-to-end task success, token overhead from reasoning, and long-context retrieval quality often matter more than leaderboard rankings.
Model Family Positioning: Hydrogen, Neon, Carbon
Google splits the Gemini 4 lineup into three distinct tiers, each optimized for a different balance of speed, reasoning depth, and cost:
- Hydrogen: Prioritizes low latency and routine agent workflows. Standard pricing: $1.5 per million input tokens, $7.5 per million output tokens. It is the default pick for high-volume, simple chat and lightweight tool calls.
- Neon: Built for software development, knowledge work, and scientific reasoning. It supports four reasoning strength settings:
low,medium,high, andxhigh. This is the family’s general-purpose workhorse, capable of deep reasoning when needed while keeping token costs controlled for ordinary tasks. - Carbon: Reserved for the hardest mathematical, long-agent, and deep research tasks. Standard pricing is $4 per million input tokens, $20 per million output tokens. An introductory discounted rate of $2 / $10 remains available until November 30, 2026.
Neon is not simply a cheaper downgrade of Carbon, nor a longer-context variant of Hydrogen. It fills a critical middle slot: teams that need occasional deep reasoning can enable xhigh for complex steps, while most routine work runs at lower reasoning strength to cut token consumption.
Benchmark Performance Overview
The table below consolidates the official benchmark results from Google DeepMind, comparing Gemini 4 Neon against Claude Opus 5.5 and GPT-6.2 Sol. All scores are from the provider’s published test suite and are not independently validated.
| Category | Benchmark | Gemini 4 Neon | Claude Opus 5.5 | GPT-6.2 Sol |
|---|---|---|---|---|
| Capability Index | AA Index v5.0.1 | 61 | 59 | 57 |
| Vals Index v2.2 | 70.9% | 70.1% | 66.1% | |
| Agentic Coding | CursorBench 4.1 | 65.4% | 58.1% | — |
| Terminal-bench 4.1 (Cheating adjusted) | 65.2% | 65.8% | 61.5% | |
| FrontierCode 2.0 | 46.2% | 47.3% | 43.8% | |
| SWE | SWE-sweep (Provider harness) | 39.0% | 10.1% | 8.3% |
| SWE-sweep (Mini-SWE harness) | 12.8% | 3.6% | 11.6% | |
| Knowledge Work | AutomationBench | 56.3% | 42.5% | 41.3% |
| Agents’ Last Exam v1.1 | 75.8% | 71.6% | 73.0% | |
| Science & Math | RiemannBench | 84.0% | 80.0% | 80.0% |
| Terminal-Bench-Science 0.1 | 70.5% | 63.3% | 68.0% | |
| Reasoning | HLE-Diamond | 71.3% | 67.9% | 64.0% |
| Multimodal Understanding | MKA Climb WSR | 56.5% | 53.0% | 52.7% |
| Chartography | 73.5% | 66.3% | 75.1% | |
| Long Context | GraphWalks (256k to 1M, BFS F1) | 95.2% | 66.8% | 80.1% |
| Cybersecurity | CWE-Bench v2 | 70.7% | 67.1% | 66.0% |
| SEC-bench Pro | 90.5% | — | 88.3% | |
| WebExploitBench v1.2 | 51.3% | 37.1% | — |
Neon’s standout result is GraphWalks, the long-context retrieval benchmark, where it reaches 95.2%. This confirms its strength at locating and using information buried inside a 1M-token context window. It also leads on AA Index, CursorBench, SWE-sweep, AutomationBench, Agents’ Last Exam, RiemannBench, HLE-Diamond and CWE-Bench.
That said, the model is not universally dominant. On Terminal-bench 4.1, Neon (65.2%) is nearly tied with Opus 5.5 (65.8%). FrontierCode 2.0 also slightly favors Opus. Chartography is led by GPT-6.2 Sol at 75.1%, above Neon’s 73.5%. The lesson is clear: Neon has distinct strong suits, but no single model wins every task category.
Pricing Breakdown: Where the $3 Input Price Creates Value
Neon’s standard pricing is $3 per million input tokens and $15 per million output tokens. It also has an implicit cache read rate of $0.15 per million cached input tokens. There is no free tier, and all reasoning tokens generated by low / medium / high / xhigh modes count toward billing.
It is important to avoid oversimplifying the cost comparison. Neon’s output price is higher than Hydrogen’s $7.5 per million output tokens. Enabling xhigh reasoning also adds substantial extra reasoning tokens, which can inflate the total bill. For short, latency-sensitive tasks, Hydrogen remains cheaper and faster.
The Neon advantage appears in workflows with long documents, codebases, multi-turn tool calls, and repeated prompt caching. We can run a simple estimate: a request with 1M uncached input tokens plus 200k output tokens costs roughly $6. If 800k of the input tokens hit the implicit cache, the cached portion is billed at $0.15 / 1M, which drastically reduces the total cost. This estimate excludes grounding, tool execution, and platform overhead; final invoices follow the API control panel.
The context window supports up to 1,048,576 input tokens. The default maximum output is 65,536 tokens, which can be extended to 1M tokens with long decode mode. This is well suited for loading entire code repositories, research papers, or bundled document sets into one turn. Still, developers should not fill the context window by default. Larger windows carry higher latency and increase the chance of irrelevant context diluting the model’s focus.
Who Should Test Gemini 4 Neon First
Three groups get the clearest practical value from Neon:
- Long-context engineering teams: Developers debugging code, refactoring large repositories, or running agent workflows that read specs, logs, and test suites continuously. The 1M window and strong GraphWalks retrieval performance reduce the need to split documents into many smaller chunks.
- Knowledge and research teams: Neon accepts text, images, audio, video, and PDF inputs. It works well for synthesizing mixed-source materials into structured reports, literature reviews, and analytical drafts.
- Cost-sensitive agent builders: Teams where Carbon is overkill, and Hydrogen repeatedly fails complex tasks that require deeper reasoning. Neon’s four reasoning modes give a tunable middle option.
For ultra-short chat, high-throughput simple classification, or tasks where every millisecond of latency matters, Hydrogen is still the better fit. For extremely hard mathematical proofs, autonomous long-running agent chains, and tasks where correctness is the top priority, Carbon remains the safer choice.
Rollout and Integration Guidance
Google is rolling out Neon gradually. It is available through the Gemini API, Google AI Studio, Gemini for business, Gemini App, and Antigravity for paid Google AI Pro / Ultra subscribers. Availability varies by region, account type, quota, and control panel settings. Seeing the model name in documentation does not guarantee immediate access to the same rate limits.
When integrating, the recommended workflow is controlled A/B testing. Lock a set of real production tasks, run them across your existing baseline model and Neon, and track completion rate, human edit count, total input and output tokens, tool call count, and end-to-end latency. It is critical to log which reasoning strength setting is used for each test. A “high score” for Neon may only apply to xhigh, while the “low price” comes from low mode. These two configurations are not interchangeable.
Managing this kind of multi-model fleet becomes easier when routing, quota enforcement, and logging are centralized. A unified API gateway helps teams route traffic across Hydrogen, Neon, Carbon and other model endpoints, standardize usage telemetry, and isolate test traffic from production; 4sapi provides that consolidation layer for teams running mixed Gemini workloads.
Final Takeaway
Gemini 4 Neon strengthens the mid-tier segment of Google’s model lineup. Its core selling point is the combination of a 1M-token context window, strong long-context retrieval, flexible reasoning controls, and a predictable price point that sits between Hydrogen and Carbon. It beats Opus 5.5 and GPT-6.2 Sol on many agentic coding, SWE, cybersecurity and long-context benchmarks, but it is not a universal replacement for flagship models.
The optimal strategy is workload segmentation: Hydrogen handles routine high-volume chat, Neon carries most complex daily engineering and knowledge work, and Carbon is reserved for the hardest long-horizon reasoning jobs. Before migrating production traffic, always validate on your own task set rather than relying only on public leaderboard numbers.




