Gemini 4 Hydrogen vs Gemini 4 Argon: Cheaper, Faster, Slightly Smarter — and Twice as Likely to Invent a Fact

On the composite score that most people quote, Gemini 4 Hydrogen beats Gemini 4 Argon by one point. On the independent harness that produces it, Hydrogen finishes an index task in 13.5 minutes against Argon's 30.0, costs $3.16 against $4.97, and generates answers at 190 tokens per second against 65. That looks like a rout until you read the second table, where Argon hallucinates on 4.3% of Artificial Analysis's omniscience questions and Hydrogen on 8.5%, and where Argon wins both Terminal-Bench 4.0 and Terminal-Bench-Science. Gemini 4 Hydrogen and Gemini 4 Argon are the same underlying model with different thinking budgets bolted on, and the choice between them is not "new versus old" — it is "how much thinking do you want to pay for, and how much do you trust the answer without checking it."

The gap is a thinking budget, not an architecture

Google shipped Argon first. It was announced on 30 September 2026, reached general availability on 13 October, and was positioned as the efficient frontier model: the same 1,048,576-token unified context window as Gemini 4 Carbon, the same built-in tools, the same per-token price, but fewer thinking tokens burned per turn. Gemini 4 Hydrogen arrived on 12 November with the same window, the same tools, and a card that says plainly it is adapted from Carbon rather than trained separately.

What actually separates the two is the dial at the top of the request. Argon accepts low, medium and high and defaults to medium. Hydrogen accepts those three, defaults to high, and adds a fourth rung called xhigh. Everything else about the shape of the model is shared.

That matters because the headline comparison everyone runs is not like-for-like. Artificial Analysis publishes Hydrogen's 54 at xhigh and Argon's 50 at high. At Hydrogen's own default — high, the setting a request gets if you send nothing — the same harness scores it 51. So the honest one-point gap is 51 to 50, and the four-point gap is what you get by opting into the most expensive rung on one model and comparing it against the middle rung of the other.

Push the dial the other way and the ordering flips. Hydrogen at medium scores 41. If your workload runs at Google's recommended default for Argon, Hydrogen is not the better model — it is a cheaper one that thinks less.

Six lines that carry the decision

All figures below are from Artificial Analysis's own runs against Google's API unless marked otherwise. Per-token prices are Google's published rate card; both models are on the same card shape.

• Intelligence Index — Gemini 4 Hydrogen 54 at xhigh, 51 at high, 41 at medium; Gemini 4 Argon 50 at high, its ceiling. The comparable pair is 51 versus 50.

• Cost per completed index task — Hydrogen $3.16 at xhigh and $1.87 at high; Argon $4.97. The per-task gap is much wider than the one-point score gap, and it widens further at the default setting.

• Output speed — 190 tokens per second for Hydrogen at xhigh, 179 at high; 65 for Argon. Roughly a three-to-one throughput advantage that survives every thinking level.

• Time to first answer token — Hydrogen 37.8 seconds at xhigh and 16.7 at high; Argon 40.7. At the default setting Hydrogen starts answering 24 seconds sooner, which is the difference a chat surface notices.

• Verbosity — Hydrogen generated 304 million output tokens to complete the index, against 113 million for Argon and a 72.8-million median. Argon is already described by the same harness as "somewhat verbose"; Hydrogen is "very verbose", and that is what eats most of its per-token price advantage at xhigh.

• Price and tiers, per million tokens — Hydrogen $1.50 / $7.50 flat, cache reads $0.075, Batch and Flex at $0.75 / $3.75, Priority at $2.70 / $13.50, and a rate-limited free tier. Argon $2.00 / $10.00 through 30 November 2026, cache reads $0.10, Batch and Flex at $1.00 / $5.00, Priority at $3.60 / $18.00, no free tier. Both offer an invite-only Contributor tier; Hydrogen's is $0.75 / $3.75, Argon's $1.00 / $5.00.

Where Argon is still the better model

Three independent results and one hallucination rate say the newer, cheaper model is not a strict upgrade. They are the reason this page is not a recommendation to migrate.

• Terminal-Bench 4.0 — Argon 57.1%, Hydrogen 53.5%. The agentic coding benchmark where the older model still leads, and the one closest to what a long tool-using session actually does.

• Terminal-Bench-Science — Argon 52.9%, Hydrogen 49.5%. The same pattern in the science variant.

• AutomationBench-AA — Argon 77.5%, Hydrogen 76.1%. A narrow loss, but a loss, and it is the knowledge-work benchmark Google's own card quotes most heavily.

• Omniscience hallucination rate — Argon 4.3%, Hydrogen 8.5% at xhigh. Hydrogen answers a similar share of these questions correctly and invents an answer roughly twice as often when it does not know. On the same set at high the rate is 8.7%, so this is not an artefact of the top thinking level.

The counter-list is real too, and it is why Hydrogen still wins the composite: FrontierCode 2.0 at 39.0% against Argon's 25.7%, CritPT agentic hard at 66.0% against 52.5%, HLE-Diamond at 58.8% against 56.7%, and a small edge on the engineering, legal and finance capability slices. Hydrogen is the better model on code repair and on hard competition-style reasoning, and the worse one on sustained terminal work and on knowing when to say it does not know.

That split has a practical shape. If a verifier, a test suite or a human reads the output before it matters, Hydrogen's hallucination rate is a cost you are already paying to catch. If nothing downstream checks the answer — a support reply, an extraction feeding a database, a summary a customer reads — Argon's 4.3% is worth more than Hydrogen's extra 40 tokens per second.

The 1 December step widens the gap without reversing it

Both models are on promotional cards with a hard expiry. Argon goes from $2.00 / $10.00 to $4.00 / $20.00 on 1 December 2026, with cache reads doubling to $0.20, Batch and Flex to $2.00 / $10.00 and Priority to $7.20 / $36.00. Hydrogen's published card carries no dated step at any tier, including the free one.

Run the arithmetic after the step and the per-token ratio goes from 1.33x to 2.67x in Hydrogen's favour. Per-task, Artificial Analysis recomputes its figure from the live rate card, and it has already published both versions: $4.97 while Argon's card read $2.00 / $10.00, and $9.95 on the run that read $4.00 / $20.00. The same harness puts Hydrogen at $3.16 across both. That does not make Argon the wrong model; it makes Argon a model you should be able to justify on accuracy rather than on habit.

It also makes the two models want different routing. OrcaRouter carries Gemini 4 Argon and Gemini 4 Carbon at provider list price with 0% markup, so the 1 December doubling appears on our side the same day it appears on Google's card rather than at the next invoice reconciliation, and a stack that keeps Argon for its accuracy checks pays the new rate transparently. Gemini 4 Hydrogen is not in our catalogue — it comes from Google's own API and AI Studio — so the configuration that matches this article is Hydrogen called direct for execution turns and Argon behind the router as the verifier, with automatic failover across providers covering the case where one path rate-limits mid-session.

What changes in your code if you switch

Google documents the migration as a model-string change with four behavioural differences, and only the first is obvious.

• Hop size — Hydrogen allows 65,536 tokens per non-streaming hop across REST and SDK calls; Argon allows 32,768 per non-streaming REST hop. Long single-shot generation gets simpler, and any code that chunked around the 32K limit can stop chunking.

• Thinking defaults — Argon defaults to medium and caps at high; Hydrogen defaults to high and adds xhigh. A straight model-string swap therefore silently raises both latency and thinking-token spend. Set thinking_level explicitly on the first request or your bill will tell you about it later.

• Caching — neither supports explicit caching; both discount implicit cache hits by 95% on prompts of 4,096 tokens or more. Argon's cache read is $0.10 and Hydrogen's $0.075, so a cache-heavy workload saves 25% on the cached portion before the December step and 62.5% after it.

• Sampling and tool calls — both ignore temperature, top_p and top_k rather than rejecting them, and both will return MALFORMED_FUNCTION_CALL if you prompt for raw-text JSON tool calls instead of declaring tools natively. Neither is a switch-specific problem, but both surface during one.

One further difference is not on the migration list and matters for agents: Argon's card publishes no knowledge cutoff, while Hydrogen's model card states March 2026 with a caveat that some domains are less current. If a recency-sensitive pipeline is the reason you are comparing these two, Hydrogen is the only one of the pair where you can read the date instead of testing for it.

The rule to apply

Switch from Argon to Hydrogen when the work is executed and then checked: test-driven coding, file edits, parallel subagents, retrieval-then-answer pipelines with a citation step, and anything where cost per finished task is the number on the dashboard. Run Hydrogen at high, not the default of your imagination — the published 54 requires xhigh, and the practical default is a 51 that costs $1.87 per task.

Stay on Argon when the answer ships unchecked, when the task is a long terminal session rather than a burst, or when you are inside three weeks of a launch and would rather not move a production path for a one-point score change. At 4.3% hallucination against 8.5%, Argon is buying you a verification step you would otherwise have to build.

What would change the answer is a third data point neither model has yet: a reproducible measurement of the thing Google actually advertised for Hydrogen — cleaner code and fewer fabricated claims, which the documentation puts at more than five times fewer than Gemini 3.8 Flash. Artificial Analysis's omniscience set measures the opposite direction for Hydrogen against Argon. Until someone runs both on the same task distribution, the honest position is that Hydrogen is the better economics and Argon is the better answer, and the 1 December price step makes that trade more expensive to defer rather than less.