
Gemini 4 Neon vs Carbon vs Hydrogen vs Argon: Which Should You Use?
On this page
- Start with Hydrogen for most new projects. It has the lowest paid API rates ($1.50/$7.50), the only free Developer API tier in the lineup, and scores 51 (High) to 54 (Xhigh) on the Artificial Analysis Intelligence Index at up to 170 tok/s.
- Use Neon for recurring engineering, writing, and research workflows. It hits 61 (Xhigh) and 60 (High) on the Artificial Analysis Intelligence Index and ranks #1 overall on the Coding Agent Index (58.64 at $6.86/task vs. Carbon's 55.47 at $9.75/task) while running nearly twice as fast as Carbon.
- Reserve Carbon (Intelligence Index 60) for high-stakes verification and deep reasoning where a missed edge case costs more than extra latency, and keep Argon (Intelligence Index 50) on your shortlist for specialist legal drafting, where it leads Vals' HLAB-Vals-1.1 benchmark.
- Watch the calendar: Carbon and Argon both double from $2/$10 to $4/$20 per million input/output tokens on December 1, 2026, and thinking tokens count as billed output.
If you're an agent: Set up Valyu and make your first search.
If you're a human: Create an account and get an API key.
MCP onboarding: Connect to https://mcp.valyu.ai/mcp and follow the MCP setup guide.
Four models in five weeks, and no single winner
Google shipped four Gemini 4 models in five weeks: Argon on October 13, Carbon on November 10, Hydrogen on November 12, and Neon on November 18, 2026 (changelog). Anyone trying to sort them into a neat ladder from "cheapest" to "best" runs into surprises immediately.
Neon matches Carbon on the Artificial Analysis Intelligence Index at High effort, edges past it at Xhigh, and takes the #1 spot on the Coding Agent Index at nearly twice Carbon's output speed. Hydrogen, the cheapest model in the lineup, already outscores Argon on the Intelligence Index while running at 165 to 170 tokens per second. Yet Argon, the earliest release of the four, still leads the whole family on Vals' legal work-product benchmark.
Picking the right model starts with the job. Default to Hydrogen for everyday execution, use Neon for recurring complex engineering, writing, and research, and make Carbon or Argon earn their place on specific workloads. That tracks Google's own model selection guide: Hydrogen for routine coding and high-volume interaction, Neon for complex engineering, knowledge, and science workflows, and Carbon for the hardest reasoning, verification, and long-horizon tasks.
All figures below reflect published documentation, pricing, and third-party benchmarks as of November 19, 2026 (no custom benchmark runs were performed here, and pricing or model behavior can change).
| Model | Start here when... | What should earn it the job? |
|---|---|---|
| Hydrogen | You need a fast, low-cost default for coding, routing, or interactive apps | Clearing your quality bar at the lowest standard rates of the four |
| Neon | You regularly run code migrations, long-form synthesis, writing, or complex research | More accepted work per dollar spent, including retries |
| Carbon | A subtle reasoning or verification miss would be expensive to ship | Fewer consequential errors or less reviewer rework on hard cases |
| Argon | You run a working Argon pipeline, or have specialist drafting tasks | Regression stability or a task-level win |
Call them through the Gemini API as gemini-4-hydrogen, gemini-4-neon, gemini-4-carbon, and gemini-4-argon (catalog). Only Hydrogen offers a free Developer API tier, and paid API rates are separate from consumer Gemini plans (pricing). Argon remains stable with no deprecation or shutdown announced (deprecations).
Start with Hydrogen, and give hard cases an exit ramp
The easiest way to overspend on an LLM stack is routing routine work to your heaviest model. Take a support-triage queue that assigns each ticket a topic, an urgency tier, and a destination.
Evaluate Hydrogen first on a reviewed slice of real tickets, including ambiguous requests, missing details, and prompt-injection attempts to override routing policy. Have it return a structured routing object with a policy reference and a human-review flag, while keeping actions like refunds and ticket closure outside its permissions.
Measure operational outcomes: wrong queues, unsupported urgency escalations, latency, and cost per accepted ticket. When Hydrogen handles the bulk of traffic cleanly and struggles on the ambiguous tail, route just those flagged tickets to Neon rather than upgrading the entire queue.
Artificial Analysis benchmarks: Neon vs Carbon, Hydrogen, and Argon
For engineering teams picking a primary workhorse, independent data from Artificial Analysis (November 19, 2026 snapshots) puts the four models into sharp focus. Across both the Intelligence Index v5.0.1 (spanning knowledge work, coding, science, and general reasoning) and the Coding Agent Index (where all four models are evaluated inside the same Antigravity CLI harness), Neon delivers flagship accuracy with faster throughput and lower average task cost than Carbon.
| Model (top evaluated setting) | AA Intelligence Index | AA Coding Agent Index (Antigravity CLI) | Mean Coding Task Cost | Output Speed |
|---|---|---|---|---|
| (Xhigh) | 61 (60 at High) | 58.64 (#1 overall) | $6.86 | 110 tok/s (101 at High) |
| (High) | 60 (58 at Medium) | 55.47 | $9.75 | 56 tok/s |
| (Xhigh) | 54 (51 at High) | 52.32 | $4.45 | 170 tok/s (165 at High) |
| (High) | 50 (48 at Medium) | 45.50 | $6.84 | 66 tok/s |
Three numbers stand out immediately:
- Neon leads the Coding Agent leaderboard: At 58.64, Neon (Xhigh) sits #1 overall on Artificial Analysis's Coding Agent Index, ahead of Carbon (High) at 55.47, while cutting average task cost by nearly 30% ($6.86 vs. $9.75) and running nearly twice as fast in output throughput (110 tok/s vs. 56 tok/s).
- Thinking depth creates real tiers inside each model: Neon steps cleanly across reasoning levels on the Intelligence Index (61 Xhigh, 60 High, 55 Medium, 44 Low), while Hydrogen spans 54 (Xhigh), 51 (High), 41 (Medium), and 27 (Low). Even at default High effort, Neon ties Carbon High at 60 while Hydrogen beats Argon High (51 vs. 50).
- Watch verbosity on heavy reasoning runs: Artificial Analysis flags Neon (Xhigh) and Hydrogen (Xhigh) as verbose on the Intelligence Index (283M and 304M output tokens across the full suite vs. 205M for Carbon High and 113M for Argon High). Because thinking tokens are billed as output, dialing Neon down to High or Medium can trim spend sharply when you do not need maximum search depth.
Put Neon on the migration; make Carbon earn the escalation
Picture a shared API change spanning several services where a silent compatibility break hurts downstream teams. Run Neon against a pinned repository snapshot, migration spec, and tool budget, asking for a dependency map, isolated branch edits, executor-run tests, and a list of unresolved risks.
Bring Carbon into the review for the highest-risk verification pass, such as confirming that backward-compatibility invariants hold across edge cases. That follows Google's positioning of Carbon for hard verification (selection guide) and gives you a clear threshold: escalate only when reviewers actually catch fewer bugs or spend less time reworking patches.
Check your harness before blaming the model. In Google's Neon announcement, the same Neon model jumped from 12.8% on SWE-sweep with mini-SWE to 39% with Antigravity. Even as a vendor-reported result, the takeaway is clear: better tools, sharper context, and tighter execution loops often matter more than a model swap.
Argon vs Neon: why legal drafting and legal research split
Composite leaderboards blur the specialist edge cases that decide domain rollouts. On the Vals Index 2.2 (updated November 18), all four models finish within 2.4 points: Neon at 70.87%, Carbon at 70.45%, Argon at 69.39%, and Hydrogen at 68.49%.
Look at the underlying legal and tax tasks, though, and the ranking flips:
| Vals task | Neon | Carbon | Argon | Hydrogen |
|---|---|---|---|---|
| HLAB-Vals-1.1 legal work-product | 19.6% | 17.5% | 27.1% | 11.3% |
| Legal Research | 56.3% | 55.8% | 54.8% | 53.8% |
| Tax Agent | 54.9% | 51.8% | 49.2% | 52.3% |
Argon leads by 7.5 points on HLAB-Vals-1.1 legal work-product (27.1% vs. Neon's 19.6%), while Neon leads Legal Research (56.3%) and Tax Agent (54.9%). Vals does not disclose thinking settings, and benchmark percentages measure test performance rather than autonomous real-world coverage.
For legal teams building contract issues registers or reviewing deal packets, dropping Argon just because Neon sits higher on overall indices leaves a strong contender on the table. Compare Argon and Neon side by side on redacted packets inside your approved environment, scoring clause references, quoted evidence, missing-document flags, and how cleanly each separates contract text from interpretation. If the workflow also needs live legal authority, test retrieval and research separately.
One quick clarification from Google's Argon documentation: where Google notes faster responses and fewer tokens per turn for Argon, the comparison is against Carbon, not Neon.
Gemini 4 API pricing: Carbon and Argon double on December 1
The single biggest line item on Google's pricing page is a calendar date. Through November 30, 2026, Carbon and Argon carry introductory rates of $2 per million input tokens and $10 per million output tokens. On December 1, both double to $4 / $20.
Hydrogen stays at $1.50 / $7.50 and Neon stays at $3 / $15 under the published schedule. Billed output includes thinking tokens.
| Model | Input / output per 1M through Nov 30 | Input / output per 1M from Dec 1 | Example request through Nov 30 | Same request from Dec 1 |
|---|---|---|---|---|
| Hydrogen | $1.50 / $7.50 | $1.50 / $7.50 | $0.0675 | $0.0675 |
| Neon | $3 / $15 | $3 / $15 | $0.135 | $0.135 |
| Carbon | $2 / $10 | $4 / $20 | $0.09 | $0.18 |
| Argon | $2 / $10 | $4 / $20 | $0.09 | $0.18 |
The example columns fix usage at 20,000 input tokens and 5,000 billed output tokens (including thinking), with no caching, tools, or retries. For Hydrogen, 0.02 × $1.50 + 0.005 × $7.50 = $0.0675.
December 1 flips the relative token math. In late November, Neon looks pricier per token than Carbon ($0.135 vs. $0.09). From December 1 onward, Neon is 25% cheaper per token than Carbon and Argon ($0.135 vs. $0.18). On Artificial Analysis's full Intelligence Index suite, that shows up clearly: Neon (Xhigh) averages $6.44 per task versus $7.76 for Carbon (High) at its $4/$20 standard rate, while Hydrogen (Xhigh) comes in at $3.16 and Argon (High) at $4.97 (measured at its $2/$10 introductory rate).
Two pricing details are easy to miscalculate in forecasts:
- Discounts only apply to qualifying traffic: Implicit cache hits get a 95% input discount on actual hits, and explicit caching is unsupported. Batch and flex tiers list 50% rates but behave differently from interactive calls (pricing).
- Track cost per accepted task: Thinking volume and retry rates can flip a token-rate comparison completely. Measure
cost per accepted task = total spend / accepted tasks, counting failed attempts, retries, billed thinking, and paid tools in the numerator and each passing task once.
Speed and thinking settings: know which wait you are measuring
Speed across Gemini 4 depends heavily on whether you measure streaming throughput, wait time before the first answer token, or total wall-clock time per task. Artificial Analysis's November 19 snapshots separate all three:
- Output throughput: Hydrogen leads at 170 tok/s (Xhigh) and 165 tok/s (High), followed by Neon at 110 tok/s (Xhigh) and 101 tok/s (High), Argon at 66 tok/s (High), and Carbon at 56 tok/s (High).
- Time to first answer token (includes thinking): At High effort, Hydrogen responds fastest (20.08s), followed by Neon (40.64s), Argon (48.55s), and Carbon (72.86s). Turning on Xhigh extends that initial thinking pause to 45.22s for Hydrogen and 56.51s for Neon.
- Average time per Intelligence Index task: Hydrogen finishes in 8.6 minutes (High) or 15.2 minutes (Xhigh), Neon in 14.8 minutes (High) or 21.5 minutes (Xhigh), Argon in 29.5 minutes (High), and Carbon in 36.6 minutes (High).
Thinking controls also vary across the lineup (catalog):
- Carbon and Argon support
low,medium(default), andhigh. - Neon and Hydrogen support
low,medium,high(default), andxhigh.
Benchmarking Neon at its default (high) against Carbon at its default (medium) compares two different levels, and even matching labels can consume different compute. Test the exact setting you plan to ship alongside the acceptance rate it produces.
Shared architecture: multimodal in, text out, one context budget
All four models share the same underlying API surface (catalog):
- Multimodal input, text output: All four accept text, images, audio, video, and PDFs, and return text only. None include image generation, audio generation, or the Live API, so summarizing a recording fits well while real-time voice needs another model.
- One unified 1,048,576-token budget: Input, thinking, and output share a single 1M-token pool (Neon, Carbon, Hydrogen, Argon). Very large input packets leave less room for thinking and output.
- Long-decode hops: Approaching a million generated tokens requires continuation hops within remaining context. Broad summaries mention 65K hops; the long-decode guide lists
65,536tokens per hop for Neon and Hydrogen, and32,768non-streaming or65,536streaming for Carbon and Argon. - Tools need your runtime: Function calling, structured outputs, search grounding, and computer use are supported across all four, provided your app supplies the tools, executor, and permissions.
Current research: fix your evidence before buying more reasoning
A November launch date does not bring November knowledge. Every Gemini 4 model card (Neon, Carbon, Hydrogen, Argon) lists a March 2026 knowledge cutoff and notes hallucination risks. Moving from Hydrogen to Carbon will not surface a paper, filing, or release published in October.
Suppose your team needs to decide whether a new technical result should shift next quarter's roadmap. Hand Neon and Carbon the exact same retrieved source packet (primary paper, setup details, and limitations) and ask for a decision memo that separates observed results, author claims, and proposed next experiments, with supporting passages attached to consequential claims.
Where Valyu Search and DeepResearch fit
- Targeted verification: Use the Valyu Search API when you know what needs checking. Retrieve the specific paper, filing, or benchmark conditions and feed the exact passages and URLs to Gemini so your app controls what enters context.
- Open-ended investigation: Use Valyu DeepResearch for broader roadmap questions, such as tracing an original finding, checking independent replications, and flagging differences from your deployment. Audit the cited brief first, then pass that fixed packet into Neon and Carbon to compare recommendations.
Preserve URLs, publication dates, and verbatim passages, and instruct the model to rely only on supplied sources, flag missing evidence, and separate source statements from inference. Because no four-way citation-precision benchmark or measured Gemini-plus-Valyu score is published here, score retrieval coverage, citation faithfulness, synthesis quality, latency, and cost separately so you know whether a gap comes from search or from the model.
FAQ: choosing Gemini 4 without overbuying
Which Gemini 4 model should I use for coding?
Start with Hydrogen for everyday coding and fast interactive loops, then move to Neon for complex multi-file engineering; Neon (Xhigh) currently ranks #1 overall on Artificial Analysis's Coding Agent Index (58.64) inside the Antigravity CLI (selection guide). Gate any escalation to Carbon behind repository-specific tests and review metrics so the extra spend clearly pays for fewer bugs.
Is Neon better than Carbon?
For most recurring engineering and agentic workloads, yes. On Artificial Analysis, Neon matches Carbon on the Intelligence Index at High effort (60 vs. 60), edges ahead at Xhigh (61), outscores Carbon on the Coding Agent Index (58.64 vs. 55.47) at lower average task cost ($6.86 vs. $9.75), and streams nearly twice as fast (110 tok/s vs. 56 tok/s). From December 1 onward, Neon is also 25% cheaper per token. Carbon still earns a spot in evaluations for high-stakes reasoning and verification where reviewers confirm it catches hard edge cases Neon misses.
Is Argon deprecated, and should I migrate off it?
No. Google has announced no deprecation or shutdown for Argon (deprecations). Even though Hydrogen and Neon outscore it on Artificial Analysis's Intelligence Index, keep Argon in your comparison if an existing pipeline passes your regression suite or if you run specialist legal drafting where Argon leads Vals' HLAB-Vals-1.1 benchmark (Vals). Account for its rate doubling to $4/$20 on December 1.
How should I make the final call?
Define acceptable output first. Compare representative tasks using identical tools, permissions, and side-effect checks, repeating runs where variance matters. Report acceptance rate, latency, and total spend together, using December 1 rates for any Carbon or Argon deployment continuing past November.
Most teams do not need a single model to win every benchmark. They need a fast default for routine work, a capable workhorse for recurring complexity, and an escalation tier that proves it earns its price.
More from the blog





Valyu Add
Join 12,000+ professionals and knowledge workers.
Valyu Add is a free weekly research briefing for builders, investors and operators. Every issue is sourced, cited and verified with Valyu DeepResearch.
