Generated title card reading Gemini 4 Neon vs Gemini 4 Hydrogen, subtitle Where the 2x Stops Working, showing a rising cost-per-finished-task axis with alternating Neon and Hydrogen markers and the caption Rate cards say 2x; per-task cost says otherwise.
Engineering & Research

Gemini 4 Neon vs Gemini 4 Hydrogen: Where the 2x Stops Working

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Two G​emini 4 models from the same family, shipped six days apart, and the rate cards compare like a photocopy with the numbers halved: G​emini 4 Neon at $3.00 in and $15.00 out, G​emini 4 Hydrogen at $1.50 and $7.50, cached input $0.15 against $0.075. Every line is exactly half. That is the tidiest price relationship in the G​emini 4 family and the least informative thing about the pair, because the moment you sort all eight of their thinking levels by what they actually cost per finished task, the tidy 2x stops predicting anything at all.

Both are current and both are G​oogle's. G​emini 4 Hydrogen went live on 12 November 2026 as the model that replaced the Flash naming; G​emini 4 Neon followed on 18 November 2026, positioned by G​oogle as sitting between Hydrogen for everyday execution and Gemini 4 Carbon for deep long-horizon work. Both carry a 1M-token context window, both accept text, image, audio and video, both have four thinking levels, and both default to high.

Two cards, eight prices, one sort

Artificial Analysis scores every rung of both models on the same nine-part evaluation set and publishes a measured cost per finished task for each. Sorting those eight points by cost rather than by model name is the whole argument, so here they are in order:

• Hydrogen at low — index 26.9, $0.29 per task.

• Neon at low — 44.1, $0.75.

• Hydrogen at medium — 41.2, $0.80.

• Hydrogen at high (default) — 50.8, $1.87.

• Neon at medium — 54.9, $1.90.

• Hydrogen at xhigh — 53.9, $3.16.

• Neon at high (default) — 59.6, $3.74.

• Neon at xhigh — 61.4, $6.44.

Read the list in order and the shape is clear. Hydrogen owns exactly one segment of it: the very bottom, where $0.29 buys a 26.9 that nothing else in either family approaches on price. Above that line the sort stops alternating between the two models and starts favouring one — and which one it favours is not the one the rate card would suggest.

Artificial Analysis model page for Gemini 4 Hydrogen at Xhigh effort, showing an Intelligence Index of 54, $1.50 input and $7.50 output per million tokens, $3.16 per Intelligence Index task, 304 million output tokens on the index against a 65.4-million median, 164 tokens per second, 61.37 seconds to first answer token, a default effort of high and a 1M-token context window.

Where the frontier actually bends

The single most useful comparison on that list is rungs five and six. Neon at medium costs $1.90 per task and scores 54.9. Hydrogen at xhigh — its top rung, the most it can do — costs $3.16 and scores 53.9. So Neon's second-cheapest useful setting is both a point smarter and 40 percent cheaper than Hydrogen's ceiling.

The same inversion shows up one row earlier. Hydrogen's default, high, costs $1.87 and scores 50.8. Neon's medium costs three cents more and scores 54.9 — four index points for a rounding error in the bill. And at the bottom, Neon's low at $0.75 scores 44.1 against Hydrogen's medium at $0.80 scoring 41.2, which is the same trade in the other direction: a slightly cheaper, slightly better rung sitting one notch lower on the other model.

The pattern is that Hydrogen's ladder is compressed. Its four rungs span 27 index points but only eleven dollars of task cost at the extremes, and the last two rungs buy nothing worth having — high to xhigh costs 69 percent more per task for three points. Neon's ladder is stretched: 17.3 index points across a cost range nearly eighteen times wider, and its ceiling is nine points above Hydrogen's because Hydrogen does not have a rung that reaches it.

The default trap, and G​oogle's own answer

Both models ship at high, and the two defaults sit on opposite sides of that inversion. Calling both with no thinking parameter at all gives you 59.6 for $3.74 against 50.8 for $1.87 — an 8.8-point gap at exactly the 2x the rate card promises. That is the comparison G​oogle's own announcement implies, and it is fair as far as it goes.

It is not what G​oogle recommends doing. The Antigravity post of 18 November describes an Orchestrator mode that pairs the two in a single session: Neon plans, architects and handles the hard reasoning, Hydrogen executes the routine work — code edits, package updates, test runs — and hands control back to Neon when it hits something it cannot resolve. G​oogle's line is that this stretches your quota further per run, which is a vendor framing of a real observation: no single rung of either model is the efficient choice for every step of a multi-step task.

That recommendation is the most honest thing in either launch post, because it concedes the point the numbers make. If Neon's medium beats Hydrogen's xhigh outright, then the case for Hydrogen is not "cheaper for the same work" — it is "cheap enough to run on the steps that do not need judgement," and Orchestrator is the mechanism for deciding which steps those are. How well the handoff actually works is a claim nobody outside G​oogle has measured; the post describes the loop and does not publish its cost.

Hydrogen's three real wins

None of the above makes Hydrogen redundant, and three of its advantages are measurable rather than positional:

• Latency. At matched rungs Hydrogen's time to first answer token is roughly half of Neon's — 21.7 seconds against 40.6 at high, 12.4 against 20.1 at medium, 6.1 against 13.9 at low. For an agent loop issuing dozens of turns, that difference compounds into minutes of wall-clock per task that no index score captures.

• Throughput. Hydrogen decodes around 160 tokens per second at high against Neon's 101. Combined with the latency advantage it is the faster model in both senses, which is consistent with how G​oogle describes the two: Neon for depth, Hydrogen for rapid everyday execution.

• Fabrication. On AA-Omniscience's hallucination rate, Hydrogen sits at 8.7 percent at its default and 8.5 percent at xhigh; Neon is at 13.9 and 14.4. Hydrogen is the more conservative of the two by a factor of about 1.6, and it is the one counter-example to the assumption that the cheaper, smaller model is the more likely to invent something. Neon still scores higher on the underlying omniscience index — 53.8 against 43.7 at matched defaults — so it knows more and asserts more; Hydrogen knows less and says less.

That last pairing is worth holding onto. Omniscience and hallucination are not two views of one quantity, and a model can improve on one while worsening on the other. Here the direction is split, which is exactly the case where a single "accuracy" number misleads.

Generated four-panel graphic titled Hydrogen's three wins, one loss, with panels for time to first token at high (Hydrogen 21.75 s, Neon 40.6 s), output speed at high (160 tok/s against 101 tok/s), hallucination rate (8.7 percent against 13.9 percent, lower is better) and omniscience index (43.7 against 53.8, Neon leads here). The footer reads All figures per Artificial Analysis v5.0.1 at the default high setting.

Picking a rung, not a model

The decision rule that survives all eight points:

• If the work is genuinely trivial — classification, extraction, a routing decision, a short rewrite — Hydrogen at low or medium is the cheapest thing in the family and the 2x saving is real.

• If the work needs planning or repo-level code changes, Hydrogen's ladder does not reach. Its best rung is 53.9; Neon's default is 59.6 and its medium is 54.9 at 40 percent less money than Hydrogen's ceiling. Take Neon and start at medium.

• If you are on Neon, the expensive rungs are the ones to audit first. high to xhigh is 72 percent more per task for 1.8 index points and an extra sixteen seconds to first token. Most pipelines belong at medium, where 54.9 arrives for half the cost of the default.

• If throughput is the constraint rather than intelligence — an interactive tool, a long agent loop, anything where a human is waiting — Hydrogen's latency advantage may matter more than eight index points, and no leaderboard will tell you that.

Neither model is in OrcaRouter's catalogue, so this is not an argument to move either onto our key. What is there, at provider list price with no markup, is Gemini 4 Carbon at $2.00 and $10.00 and Gemini 4 Argon at $2.00 and $10.00, alongside the rest of a 229-model catalogue on one endpoint. If you want to test whether Neon's medium really clears Hydrogen's ceiling on your own traffic, the practical shape is a routing rule that sends the cheap steps to a rung you can already call and escalates the hard ones — with failover underneath, so a model that is a week old never becomes the only path to production.

Artificial Analysis model page for Gemini 4 Neon at Xhigh effort, showing an Intelligence Index of 61, $3.00 input and $15.00 output per million tokens, $6.44 per Intelligence Index task, 283 million output tokens on the index against a 65.4-million median, 110 tokens per second, 56.54 seconds to first answer token, a default effort of high and a 1M-token context window.

What would change this answer

Three things, all checkable. If G​oogle published a fifth Hydrogen rung, the inversion would collapse and the 2x would start behaving like a 2x again — the gap only exists because Hydrogen's ladder stops at 53.9. If Orchestrator's handoff turned out to cost more in coordination than it saves in execution, the pairing argument would fail and the choice would revert to picking one model. And if a second independent harness reproduced the hallucination split rather than the index split, the case for Hydrogen on fact-sensitive work would get stronger than the index suggests, because a 1.6-fold difference in fabrication is worth more than eight index points in some pipelines and worth nothing in others.

Until then the honest summary is that these two models are not two points on one line. They are two ladders of different lengths, sold at a ratio that holds exactly on the rate card and breaks as soon as you price a finished task. Neon's cheaper rungs beat Hydrogen's expensive ones; Hydrogen's cheapest rung beats everything on price; and the vendor's own recommendation is to stop choosing and run both, which is the one answer the leaderboard cannot give you.