Six months ago, choosing a Gemini model meant picking between a large model and a fast model. Over the past three weeks, Google DeepMind reshaped that lineup into a specialized system of models named after chemical elements, alongside dedicated endpoints for single-pass decisions and native video.
For engineering and platform leads, the new lineup unlocks much better cost and latency performance when each model is matched to the job it was built for. Here is the complete breakdown of Google's current Gemini models, how they compare on benchmarks and token pricing, and how to orchestrate them together.
1. The 2026 Gemini lineup at a glance
Google's model family now separates deep reasoning, everyday agentic execution, closed-set routing decisions, and multimodal media generation into distinct tiers:
- Gemini 4 Carbon (opens in a new tab) (
gemini-4-carbon). Google's highest-intelligence frontier model, built for multi-day software engineering,/swarmmulti-agent coordination, open-ended scientific discovery, and end-to-end cybersecurity defense. - Gemini 4 Argon (opens in a new tab) (
gemini-4-argon). The efficient frontier reasoning model for daily architectural design, complex domain analysis, and low-hallucination knowledge work. - Gemini 4 Hydrogen (opens in a new tab) (
gemini-4-hydrogen). Released November 12, 2026 as the successor to Gemini 3.8 Flash. Built on Carbon's architecture for high-speed agentic loops, parallel coding workers, and fast tool calling across four thinking levels (low,medium,high, andxhigh). - Gemini Decision 1 (opens in a new tab) (
gemini-decision-1). Released November 11, 2026 via the Gemini Decisions API. Evaluates up to 1,000 yes/no (noul), multiple-choice (choice), and rubric-scoring (score) questions in a single forward pass with zero output tokens billed. - Gemini Omni 1.2 Flash (opens in a new tab) (
gemini-omni-1.2-flash). Generally available for native 15-second video generation and multi-turn video editing at $0.07 per second (720p), alongside Genie 4 for interactive 3D world simulation and Nano Banana 2.1 for image generation.
2. Head-to-head comparison: specs, benchmarks, and pricing
Every Gemini 4 reasoning model shares a unified 1,048,576-token (1M) input window and can extend output up to 1M tokens, while gemini-decision-1 processes up to 256K multimodal input tokens in a single non-generative forward pass.
| Specification or metric | Gemini 4 Carbon | Gemini 4 Argon | Gemini 4 Hydrogen | Gemini Decision 1 |
|---|---|---|---|---|
| Model ID | gemini-4-carbon | gemini-4-argon | gemini-4-hydrogen | gemini-decision-1 |
| Primary role | Deepest reasoning, cyber defense, /swarm lead | Efficient frontier reasoning and architecture | High-speed agent execution and repo coding | Single-pass routing, grading, and policy checks |
| Context / max output | 1M input / up to 1M output | 1M input / up to 1M output | 1M input / up to 1M output | 256K input / probabilities (no text decode) |
| Thinking levels | Extended verification loops | low, medium (default), high | low, medium, high (default), xhigh | Single forward pass (no thinking tokens) |
| Input price (per 1M) | $2.00 intro ($4.00 Dec 1) | $2.00 intro ($4.00 Dec 1) | $1.50 flat across 1M context | $0.10 (shared input billed once) |
| Output price (per 1M) | $10.00 intro ($20.00 Dec 1) | $10.00 intro ($20.00 Dec 1) | $7.50 flat across 1M context | $0.00 (zero output tokens billed) |
| AA Intelligence Index v5.0 | 59 (#1 overall) | 50 (high) / 48 (medium) | 51 (high default) / 54 (xhigh) | Purpose-built for closed-set evals |
| Terminal-Bench 4.0 / 4.1 | 65.2% (TB 4.0) | 57.1% (TB 4.0) | 57.0% on TB 4.1 (vs. 18.2% on 3.8 Flash) | 92.2% on JevBench / 79.9% When2Call |
| Typical speed | Deep multi-step cadence | ~65 tokens/sec | ~190 tokens/sec (2.5x faster than 3.8 Flash) | 150-250 ms end-to-end latency |
| Prompt injection & alignment | 0.0% Gray Swan IPI attack rate; 92% agent honesty | 0.7% Gray Swan IPI; 4.3% AA Omniscience | 81% less overreach; 99.9% CoT legibility | 0.0250 Top-1 ECE calibration error |
Two numbers in that table stand out for system architects. First, Gemini 4 Hydrogen (high default) scores 51 on the Artificial Analysis Intelligence Index (and 54 at xhigh), edging past Argon's 50 while generating at roughly 190 tokens per second and costing $1.50 / $7.50 per million tokens. Second, Argon still holds a lower fabrication rate on strict factual recall (4.3% on AA Omniscience vs. 8.5% for Hydrogen), while Carbon leads the industry on long-horizon coding, 1M-token GraphWalks recall (96.2%), and vulnerability remediation (88.2% on SEC-bench Pro and 71.0% on CWE-bench).
3. How Google pairs the models together
Google's own engineering tools no longer run one model from start to finish. In Google Antigravity's new Orchestrator and /swarm modes, a single task is split across tiers:
MULTI-MODEL GEMINI PIPELINE
Step 1: Intake & Policy Gate -> gemini-decision-1 (150 ms, $0 output tokens)
Scores intent, checks MCP safety, picks model tier
Step 2: Architecture & Plan -> gemini-4-carbon or gemini-4-argon
Decomposes complex spec or audits security logic
Step 3: Parallel Execution -> gemini-4-hydrogen (~190 tok/s, $1.50/$7.50 M)
Batches file reads, writes code, runs test loops
Step 4: Final Verification -> gemini-4-carbon
Verifies root-cause patch before human sign-off
On Terminal-Bench 4.0, Google reported that dynamically switching between Carbon and Hydrogen in Orchestrator achieved near-Carbon task success rates while consuming 24% less quota and finishing 19% faster.
4. Orchestrating the full Gemini lineup in CommandLyne
Inside an enterprise, multi-model routing works best when it spans your entire tool stack and every provider you use. In CommandLyne (opens in a new tab), teams connect their own Google Gemini, Anthropic Claude, and OpenAI keys under LLM Providers and Messaging (opens in a new tab) and govern them through a unified control plane:
- Task-aware dynamic routing. Through Governance and Multi-Model Control (opens in a new tab) and Dynamic Routing (opens in a new tab), CommandLyne routes requests by task complexity, user role, latency, and cost. Sub-second triage and rubric checks can run on
gemini-decision-1, high-speed document extraction and iterative coding ongemini-4-hydrogen, and deep security or architectural work ongemini-4-carbon. - Structured workflows with visual artifacts. In Agents and Workflows (opens in a new tab), multi-step graphs combine decision branches, parallel agent steps, and human approval checkpoints, rendering interactive deliverables directly in Canvas (opens in a new tab).
- Token-level spend visibility before December 1 pricing shifts. Because Carbon and Argon move from their $2 / $10 introductory rates to $4 / $20 per million tokens on December 1, 2026, CommandLyne's per-model, per-team, and per-workflow spend tracking lets platform owners see which workloads belong on Hydrogen or Decision 1 before the rate change takes effect.
5. Which Gemini model to pick for each job
- Use
gemini-decision-1whenever the output is a closed choice, a boolean policy check, a tool selection, or a rubric score. - Use
gemini-4-hydrogenas your default workhorse for multi-turn chat, fast repository exploration, spreadsheet modeling, and parallel worker steps. - Use
gemini-4-argonwhen you need high factual precision and deep domain reasoning at a steady cadence. - Use
gemini-4-carbonfor your hardest multi-file code migrations, vulnerability triage and patching, and 1M-token synthesis. - Use
gemini-omni-1.2-flashwhen your team needs 15-second product explainers or multi-turn video edits in Google Vids or the API.
Try CommandLyne to route across every Gemini 4 model, Claude, and OpenAI with full cost visibility and governance, or book a demo with our engineering team.
Try CommandLyne (opens in a new tab) | Book a demo (opens in a new tab)

