Gemini 4 Hydrogen vs Gemini 4 Carbon: One Family, Two Serving Profiles, and a Price Change That Arrives on Different Dates

Two models, six index points apart, 1.33x apart on the rate card today and 2.67x apart from 1 December, and behaving in a way that is hard to explain unless they are the same underlying capability run at two different speeds. G​oogle has not published a sentence saying G​emini 4 Hydrogen is derived from Gemini 4 Carbon. What it has published is a set of numbers that point the same direction: Hydrogen generates 188 tokens per second against Carbon's 72, reaches a first token in 40.55 seconds against 64.71, scores 54 on the Artificial Analysis Intelligence Index against 60, and finishes an index task for $3.16 against $5.48. That is not the shape of a small model next to a large one, the way a Flash next to a Pro used to be. It is the shape of one capability sold twice — once at its ceiling, once at a throughput setting — and priced as though the difference were architectural. G​emini 4 Hydrogen entered the catalogue on 12 November 2026 at $1.50 and $7.50 per million tokens. Gemini 4 Carbon, generally available since 10 November, sits at $2.00 and $10.00 until 1 December, then doubles to $4.00 and $20.00. Working out which one to call is mostly a question about which of those facts you are optimising for, and how much of the derivation story you are willing to take on inference.

Six index points, a 1.33x price gap today, and a 2.67x one from December

Artificial Analysis scores both models on the same harness, at each one's highest thinking setting, and the resulting pair is less dramatic than the marketing suggests.

• Intelligence Index — Gemini 4 Carbon 60 at high, ranked fourth of 78. G​emini 4 Hydrogen 54 at Extra High, ranked sixteenth. Six points, with Hydrogen ahead of Gemini 4 Argon's 50 on the same index.

• Output speed — Carbon 72 tokens per second vs Hydrogen 188. Hydrogen generates about two and a half times faster, which is the entire point of the model.

• Time to first answer token — Carbon 64.71 seconds vs Hydrogen 40.55 seconds at Extra High, dropping to roughly 16.7 seconds at its default High setting. Hydrogen starts sooner at every setting than Carbon does at its only one.

• Cost per completed index task — Carbon $5.48 vs Hydrogen $3.16. Roughly 58% of the price, for 90% of the score.

• Price per million tokens — Carbon $2.00 / $10.00 now, doubling on 1 December; Hydrogen $1.50 / $7.50 with no dated step on the published card.

Put the last two lines side by side and the interesting thing is not that Hydrogen is cheaper. It is that, on today's card, the cheaper model is cheaper per completed task by more than it is cheaper per token. Carbon burns more tokens and more wall-clock to reach a task completion, so the per-token advantage compounds rather than being eaten by retries. That is unusual, and it is the strongest evidence that these really are two serving profiles of one model rather than two different models.

Where the six points actually land

Six index points is an abstraction. G​oogle publishes the same benchmark family for both models against different comparison sets, which lets the losses be located rather than just counted. All figures below are vendor-reported and unreproduced; the comparison sets differ, so read the shape rather than the arithmetic.

• Agentic coding at scale — Carbon 61.2% on FrontierSWE 2 against GPT-6 Astra's 65.5%, and 65.2% on Terminal-Bench 4.0. Hydrogen 57.0% on Terminal-Bench 4.1, cheating-adjusted. Eight points of Terminal-Bench, on adjacent versions of the same benchmark.

• Software engineering sweeps — Carbon 9.6% on SWE-sweep; Hydrogen 13.5%. Hydrogen wins this one, and it is the only headline agentic benchmark where it does.

• Long-horizon computer use — Carbon 81.5% on OSWorld-2.1 against Claude Opus 5.5's 87.2%; Hydrogen 77.6% on the same benchmark. Carbon leads by four points and both lose to a competitor.

• Task accumulation — Carbon 77.4% on Agents' Last Exam and a Vending-Bench 2 balance of $16,484; Hydrogen 71.6% and $14,549. Carbon is better at the thing that requires holding a business together across many turns.

• Long context — Carbon 96.2% on GraphWalks across the 256k-to-1M range. G​oogle publishes no equivalent figure for Hydrogen, which is itself informative: the retrieval behaviour over a full million-token window is the one capability nobody has demonstrated for the cheaper model.

• Security work — Carbon has a published security cluster: CWE-bench 71.0%, ExploitGym 54.8%, SEC-bench Pro 88.2%. Hydrogen has none. If your workload is vulnerability triage, the choice is not a trade-off, it is a capability boundary.

The pattern across those six is consistent with the serving-profile reading: Hydrogen behaves like the same capability run fast, and the tasks it loses are the ones where being fast is not the constraint. It loses on endurance, on long-context recall, on multi-hour accumulation, and on the specialised security suite. It wins on raw sweep throughput. Whether that is post-training, a different serving configuration, or something else G​oogle has not described, the effect on a workload is the same — a speed setting showing up as a benchmark table.

1 December is the date that decides this, not today

Almost every comparison written about this pair in November will be wrong in December, because one card has a step in it and the other does not.

• Carbon today — $2.00 input, $10.00 output, $0.10 cache read, implicit caching only, no free tier. Batch and Flex at $1.00 / $5.00. Priority at $3.60 / $18.00.

• Carbon from 1 December 2026 — $4.00 input, $20.00 output, $0.20 cache read. Batch and Flex double to $2.00 / $10.00. Priority doubles to $7.20 / $36.00.

• Hydrogen, unchanged — $1.50 / $7.50 standard, $0.75 / $3.75 Batch and Flex, $2.70 / $13.50 Priority, $0.075 cache read. No dated increase anywhere on the card.

• The ratio — 1.33x on input today, 2.67x from December. On output, the same. On Batch and Flex, 1.33x today and 2.67x after.

What that means for a budget is that the two models are close enough in November to be argued about on capability, and far enough apart in December that the argument mostly ends. A team that picks Carbon in November on a six-point index lead is picking it at a price that no longer exists six weeks later, and the model it chose will be the more expensive of the pair by a margin that no benchmark gap justifies for most workloads.

There is a second asymmetry nobody mentions. Carbon has no free tier, and neither does Argon. Hydrogen does — rate-limited, on G​oogle's standard tier, with the same published prices applying to it. That is not a cost saving for a production system, but it is the difference between a prototype that runs on a laptop today and one that requires a billing account first.

The Antigravity arrangement tells you how G​oogle expects them to be used

G​oogle's own coding tool ships both, and the way it combines them is more informative than either rate card.

• Availability — Carbon is listed for AI Ultra and Enterprise only. Hydrogen is listed for Free and AI Plus, Pro, Ultra and Enterprise. Carbon is a plan-tier feature; Hydrogen is the floor.

• Default role — Hydrogen is the fixed implementation model inside Orchestrator. When Carbon does the planning, Hydrogen does the file edits; G​oogle's documentation is explicit that "once the plan is ready, G​emini 4 Hydrogen takes over to apply file edits".

• Reported result — the Antigravity team reports Orchestrator using 24% less quota and finishing 19% faster on Terminal-Bench 4.1 than running Carbon throughout, at success rates near Carbon's.

• The tell — nobody at G​oogle documents a configuration that runs Carbon on every step. The product they shipped puts Carbon on the planning turn and Hydrogen on everything else, and calls that the efficient arrangement.

If the vendor's own tool treats Carbon as a planning resource that is too expensive to spend on execution, that is a stronger signal about the intended split than any comparison table. Carbon is for the turn where a wrong decision is expensive. Hydrogen is for the two hundred turns where a slow decision is expensive.

Three configurations, and what each one costs per week

Take a concrete shape: a coding agent doing roughly 4,000 task turns a week, at an average of 12,000 input and 1,500 output tokens per turn, with a 30,000-token stable prefix cached.

• Carbon throughout — 48 million input tokens at $2.00 is $96; 6 million output at $10.00 is $60; cache reads on the 30,000-token prefix at $0.10 per million are negligible. Roughly $156 a week in November, roughly $312 from December.

• Hydrogen throughout — $72 input, $45 output, $117 a week, flat. A quarter cheaper than Carbon in November and 62% cheaper from December, when the same tokens cost $312 on Carbon's card.

• Carbon planning, Hydrogen executing — one planning turn in twenty means 5% of turns at Carbon rates and 95% at Hydrogen's, which lands near $119 a week in November and near $127 from December. Effectively Hydrogen's price with Carbon's judgment on the decisions that matter.

The third configuration is the one G​oogle ships, and the arithmetic says it costs about $2 a week more than running Hydrogen alone in November, and about $10 a week more from December, while buying back most of the six-point gap on the turns where it counts. That is a much better trade than it looks, and it is invisible if you compare the two models as substitutes rather than as a pair.

One caveat on the cache column: Hydrogen's cached input is $0.075 against Carbon's $0.10, but neither model supports explicit caching — both are implicit only, so the discount is a behaviour of the service rather than something you schedule. On a workload where you know exactly what should stay resident, that is a limitation both models share, and it is the reason the prefix above is priced as a hopeful estimate rather than a guarantee.

What to do with the model strings

The identifiers are gemini-4-carbon and gemini-4-hydrogen, and the second one is the trap. There is no gemini-4-flash, no gemini-4, and no flash token anywhere in Hydrogen's identifier — the community appends "(Flash)" to describe the slot, G​oogle does not put it in the string. A migration that pattern-matches on the old gemini-3.8-flash name will fail at the identifier rather than at the model, and it will fail silently if your client falls back to a default.

Where both models are reachable from one key, the split above becomes a routing policy rather than two integrations. OrcaRouter carries Gemini 4 Carbon and Gemini 4 Argon at provider list price with 0% markup, which means the 1 December doubling lands on our side the same day it lands on G​oogle's card — no renegotiation, no new contract. G​emini 4 Hydrogen is not in the catalogue; it comes through G​oogle's own API and AI Studio. So the workable arrangement for the three-way split is Hydrogen called direct for execution, with Carbon behind the router for planning turns and automatic failover across providers when either path degrades. That keeps the planner swappable — if Carbon's December price makes the planning turn uneconomic, Argon or a rival planner is a model-string change rather than a rewrite.

The honest summary, with the uncertainty attached

Six index points separate two models that behave like one capability at two speeds, and the cheaper profile is faster to start, faster to generate, and 58% of the cost per completed task. Against that, the expensive profile is the only one with published long-context retrieval, published security benchmarks, and a documented role as the planner in G​oogle's own tooling — and it is the one whose capability claims have been exercised across a wider set of evals.

What nobody has published is the thing that would settle it for a specific workload: a task-level comparison on the same harness at the same thinking setting, with the retry behaviour visible. G​oogle's two benchmark tables use different comparison sets, and the independent harness reports a single index number rather than the subscores that would localise the gap. Until someone runs both against the same task suite, the defensible position is the architectural one — use Carbon where a wrong plan is expensive, Hydrogen everywhere else — because that is what the vendor does, what the pricing rewards from December onward, and the only recommendation that survives the six points being partly an artefact of two different settings.

Watch three things. Whether Hydrogen gets a security benchmark cluster, because its absence is the clearest capability boundary in the pair. Whether anyone reproduces the Orchestrator quota claim independently, because the whole planning/execution split rests on it. And whether Hydrogen's card acquires a dated step, because a model with no expiry date in a family where two of three members double is either a deliberate anchor price or a rate that has not been set yet.