Compare Gemini 4 Argon, Carbon, Hydrogen, and Neon across reasoning, scientific work, coding agents, multimodal understanding, and cost per accepted result.
Gemini 4 Neon and Carbon both score 60 on the Artificial Analysis Intelligence Index at high reasoning effort. Yet the benchmark estimates 14.8 minutes per weighted task for Neon and more than half an hour for Carbon. Another comparison points to a different capability. Despite sharing the same roughly million-token context window, Carbon and Argon score 96.2% and 84.2% respectively on the long-context slice of GraphWalks in the same technical evaluation. Context capacity, the ability to follow relationships through that context, and the compute spent on completing a task each deserve their own measurement. Artificial Analysis provider results · Carbon model card
This article compares the published evidence for Argon, Carbon, Hydrogen, and Neon as of November 19, 2026. Modelflare did not rerun the evaluations. Here, model intelligence refers to observable capabilities such as extracting evidence, using intermediate results to constrain a conclusion, completing a task with tools, and recognizing when the available evidence supports only a qualified answer. Each capability needs its own tasks, scoring rules, and resource budget.
What the four models share and where they spend their effort
| Model | Google public API ID | Public availability milestone | Workloads to evaluate first |
|---|---|---|---|
| Argon | gemini-4-argon | Broader paid API availability on October 13 | Complex engineering, knowledge work, and document analysis with an existing Argon baseline |
| Carbon | gemini-4-carbon | Broader paid API availability on November 10 | Deep debugging, open-ended research, and deliverables requiring repeated verification |
| Hydrogen | gemini-4-hydrogen | Launched on November 12 across free and paid API tiers | High-volume extraction, bounded code changes, and agent steps with automatic acceptance checks |
| Neon | gemini-4-neon | Launched on the paid API on November 18 | Complex, recurring coding, knowledge work, and synthesis over several sources |
The dates refer to Google's public availability announcements. The four API models accept text, images, video, audio, and PDFs and produce text. The unified context window is 1,048,576 tokens. Each model page lists a default synchronous output limit of 65,536 tokens. Long decode can extend generation through continuations, within the unified window and the constraints of the selected interface. The public materials do not give a comparable parameter count, training budget, or inference compute budget by effort level for all four models. The task conditions are the basis for comparing them. Argon API specifications · Carbon API specifications · Hydrogen API specifications · Neon API specifications · Google long decode guide
DeepMind says Neon was adapted from Carbon. Google positions Carbon for deeper exploration and verification, Hydrogen for fast execution, and Neon for recurring work that calls for both depth and throughput. Those descriptions suggest useful starting configurations. The actual choice depends on how readily the task can be decomposed, how far a mistake can propagate, and how much human review costs. Neon model card · Carbon announcement · Hydrogen announcement
What the intelligence index measures
Artificial Analysis Intelligence Index v5.0.1 combines nine evaluations. Knowledge work, coding, science, and general capabilities each receive a 25% weight. That makes the index a useful summary of the suite. An application dominated by chart extraction, cross-referencing long documents, or tool calls with strict deadlines needs the relevant subtests and a separate estimate based on its own task mix. The table uses a single index version and retains the reasoning effort for each observation. Index methodology and version history · Provider results
| Model and reasoning effort | Index score |
|---|---|
Neon xhigh | 61 |
Neon high | 60 |
Carbon high | 60 |
Neon medium | 55 |
Hydrogen xhigh | 54 |
Hydrogen high | 51 |
Argon high (AI Studio) | 50 |
Neon low | 44 |
Neon gains five points from medium to high, then one point from high to xhigh. The corresponding weighted time estimates are 8.3, 14.8, and 21.5 minutes. The final step adds about 45% to the estimated time. Its value depends on which otherwise unsuccessful tasks it recovers. The index displays whole numbers, so a one-point gap has limited meaning without paired task outcomes and run-to-run variance. Carbon's 59 at its October evaluation belongs to v5.0. The current 60 comes from a newer scoring snapshot. Neon results by effort level · Early Carbon evaluation
Scientific reasoning spans answers, execution, and verification
HLE-Diamond covers demanding knowledge and reasoning questions. DeepMind's evaluations list 71.3% for Neon and 58.8% for Hydrogen. RiemannBench covers mathematical tasks, and Neon's card lists 84.0%. Terminal-Bench-Science adds execution in a terminal environment. Tool selection, successful analysis, and correct interpretation of intermediate outputs can all affect the result. A correct final answer and a reproducible scientific workflow call for different evidence. Neon model card · Hydrogen model card · Artificial Analysis benchmark definitions
Carbon's card reports 64.8% for Carbon and 52.9% for Argon on the Artificial Analysis entry for Terminal-Bench-Science 0.1, a gap of 11.9 percentage points. Neon's 70.5% comes from a separate setup using the vision-enabled mini-swe-agent harness and averaging rewards across five attempts per task. The result should travel with its harness and aggregation rule. To judge whether more reasoning effort is worthwhile in research, ask each configuration to form a hypothesis on the same data, produce a reproducible analysis, and explain how an intermediate result changes its conclusion. Grade numerical accuracy, evidence provenance, and reproducibility individually. This reveals whether a configuration is better at finding a problem or simply better at presenting a finished report. Carbon model card · Carbon evaluation methodology · Neon evaluation methodology
Coding agents combine a model, a harness, and an acceptance standard
FrontierCode 2.0 contains 132 tasks designed by maintainers. Its criteria cover correctness, tests, scope discipline, and repository conventions. The weighted score, the proportion of runs that pass, and the average cost per run answer different questions about how many changes meet acceptance criteria, how thoroughly the work is done, and what it costs to produce. Cognition leaderboard · FrontierCode 2.0 methodology
| Model and reasoning effort | Harness | Weighted score | Pass rate | Mean cost per run |
|---|---|---|---|---|
Gemini 4 Neon high | Antigravity CLI | 46.2% | 50.8% | $4.38 |
Claude Opus 5.5 max | Claude Code | 47.3% | 51.8% | $12.82 |
GPT-6.2 Sol max | Codex | 43.8% | 48.2% | $1.29 |
These results compare systems using their own tools. In this snapshot, Neon and Opus have similar pass rates, and Neon's average run costs about a third as much. Sol costs less than Neon, with a pass rate 2.6 percentage points lower. This is enough to choose candidates for a follow-up. To isolate the model's contribution, pair the tasks and hold the visible tools, timeout, retries, workspace, and acceptance rules constant. Cognition leaderboard
DeepMind's SWE-sweep numbers show how much the harness can matter. Neon earns 39.0% in subtask-weighted reward in its provider harness and 12.8% in the vision-enabled Mini-SWE harness. That gap is 26.2 percentage points. The suite includes finding issues in a repository, and the public Antigravity CLI can delegate searches to subagents in parallel. The reported figures show a strong system effect, with the individual contributions of subagents, search strategy, and other harness differences unresolved. The same caution applies to Terminal-Bench 4.1. Hydrogen's 57.0% comes from five attempts per task at high effort. Neon's 65.2% comes from Artificial Analysis using Antigravity CLI. The 8.2-point gap combines changes to the model and the setup. Neon model card · Neon evaluation methodology · Hydrogen evaluation methodology
Long tasks also test whether an agent can keep following the delivery requirements. Carbon scores 77.4% and Argon 69.0% on the same 122-task Agents' Last Exam v1.1 setup in the Carbon model card. Both use the default ALE-Claw harness and a five-hour window, and DeepMind reports mean rubric scores. An application test can add a correction mid-task. Change one requirement after the first draft, then check whether the final result preserves the requirements that still apply and whether the agent rechecks the affected work. This measures delivery against the current instructions as well as task completion. Carbon model card · Evaluation methodology
Evidence tracking and multimodal understanding in a million tokens
GraphWalks embeds relationships in long text and asks models to follow multiple links. The Carbon evaluation reports breadth-first traversal F1 over 200 problems across the 256k-to-1M-token range. Carbon and Argon score 96.2% and 84.2% in that table. Neon reports 95.2% on the corresponding slice in its own model card. Those figures inform judgments about tracking relationships through long inputs. Contracts, consolidated accounts, and code standards also contain conflicting or superseded evidence. An application evaluation should therefore measure whether the model found the right passage, whether it used the version that governs the decision, and which material exceptions it missed. Carbon evaluation methodology · Carbon model card · Neon model card
For example, build three questions over the same material. Ask for a clause and its page number, then ask which appendix the clause references, then ask whether a newer appendix supersedes it. Move the evidence to the beginning, middle, and end while keeping the question and answer fixed. Add distractors of matching length. Report citation precision, recall of material evidence, and final-answer accuracy. These metrics distinguish a wrong page, a wrong version, and a reasoning error after reading the correct source.
For multimodal work, Chartography and LVBench illustrate the importance of input preparation. DeepMind reports 73.5% for Neon on Chartography and 75.1% for GPT-6.2 Sol in the same table. Neon scores 89.0% on LVBench. Gemini received low-resolution video at one frame per second, while Sol and Opus 5.5 had 500-frame and 600-frame inputs respectively due to API limitations. Video sampling and model choice should be held together in an application test. For an anomaly that lasts half a second, the sampling step may discard the evidence before the model can reason over it. For charts, grade the value, units, time range, and written conclusion independently. Neon model card and results · Multimodal evaluation methodology
Measure cost per accepted result
As of November 19, 2026, Google's standard Gemini API rates per million tokens are $1.50 input and $7.50 output for Hydrogen, and $3 input and $15 output for Neon. Argon and Carbon are $2 input and $10 output during the introductory period, which Google says expires November 30, and $4 input and $20 output afterward. On those per-token rates, Neon is 50% more expensive than either model during the introductory offer and 25% less expensive after it. These are direct Google API prices. Cache hits, billed reasoning output, tools, and retries affect the actual bill. Google API pricing · Argon introductory pricing · Carbon introductory pricing
For work with explicit acceptance criteria, divide the total cost of all attempts by the number of accepted results. Suppose two configurations each run 100 identical tasks. A spends $80 and passes 60, or about $1.33 per accepted task. B spends $110 and passes 90, or about $1.22 per accepted task. Pricing human review by the minute can widen or reverse the gap. Where errors have outsized effects, such as refunds, duplicate tool writes, or unattended execution, track severe failures separately and include timeouts, retries, and manual intervention in the total.
Reproduce a comparison on your own tasks
- Freeze the workload and acceptance criteria first. Grade code against behavior, compatibility, and scope. Grade research and knowledge work against factual evidence, correct inference, and completeness. Report rare severe errors separately.
- Use paired tasks. Record the model ID, effort level, tool versions, input preparation, source access, stopping rules, and total budget. Run a separate comparison with each vendor's native tools when the agent product as a whole matters.
- Measure both fixed budget and fixed quality. The first answers how much work gets done within a budget. The second estimates the spend required to meet a target acceptance rate. Measure time to the first visible result and to completion, and record medians, P95, and failure causes.
- Account for uncertainty. With a limited sample, estimate differences with a paired bootstrap by task, keeping repeated runs for the same task together. Freeze the task set before inspecting differences by difficulty, and reserve freshly collected failures for a future round.
- Escalate on observed failure modes. Start with Hydrogen for frequent steps with clear acceptance checks. Test Neon
highwhen one interaction must read sources, write code, and produce an artifact. Test Carbon where prolonged exploration or cross-module dependencies may drive rework. Keep existing Argon workflows as controls and compare regressions, completion time, and cost per accepted result.
For a cross-family comparison using the same coding leaderboard and additional reasoning evidence, see the companion analysis. GPT-6.2 Astra and Sol deep analysis
Originally published at modelflare.dev.
Disclosure: I'm part of the Modelflare team, which builds the AI API gateway mentioned in this article.

