The migration between the three Gemini 4 models is one line of code; the routing decision is not. On Google's API the request schema does not change and the model string is the switch, so the cost of trying each model is small. What differs is the envelope — the ceiling, the price and the latency — and, on AIHubMix, which protocols each model documents.
The measured picture first. At high, Hydrogen scores 51 on Artificial Analysis's Intelligence Index at $1.87 per task and 8.3 minutes. Argon at high scores 50 at $4.97 and 30.6 minutes. Carbon at high scores 60 at $5.48, with 63.37 seconds to the first answer token. Carbon's $5.48 per task buys the highest score; Argon's $4.97 buys 50.
One schema, three ceilings
All three models take a 1,048,576-token window that includes input, thinking and output. All three support long decode up to that window minus the input, and all three use implicit caching only, with a 4,096-token minimum. On Google's API, the request schema does not change when the model ID does.
The limits inside that shared shape are where the models part.
| Capability | Carbon | Hydrogen | Argon |
|---|---|---|---|
| Context window | 1,048,576 tokens, input + thinking + output | 1,048,576 tokens, input + thinking + output | 1,048,576 tokens, input + thinking + output |
| Long decode | up to window minus input | up to window minus input | up to window minus input |
| Per-hop output | 65,536 / 32,768 | 65,536 / 65,536 | 65,536 / 32,768 |
| Effort ceiling | high | xhigh | high |
| Implicit cache minimum | 4,096 tokens | 4,096 tokens | 4,096 tokens |
Two rows carry migration risk. xhigh exists only on Hydrogen; Carbon and Argon stop at high and default to medium, so a configuration pinned to xhigh is not portable as a model-string change. And the decode ceiling is the window minus the input, which puts prompt size and answer length in competition for the same 1,048,576 tokens.
The gateway is where the shared schema stops
On AIHubMix the three are documented differently, and that is the first thing an integration meets.
| On AIHubMix | Carbon | Hydrogen | Argon |
|---|---|---|---|
| Documented input | text, image, audio, video, PDF | text, image, audio, video, PDF | text, image, audio, video, PDF |
| Documented protocols | Chat Completions, Responses, Anthropic Messages, native Gemini | Chat Completions | Chat Completions, Responses, Anthropic Messages, native Gemini |
All three list multimodal input. Carbon and Argon document all four protocol surfaces; Hydrogen is documented on Chat Completions only. An existing Anthropic Messages client, or a Responses-based tool loop, has a documented home on Carbon and Argon and no documented equivalent on Hydrogen. That is a contract difference rather than a quality difference, and it settles integrations before the benchmark is consulted.
The measured trade
Artificial Analysis's live snapshot at 2026-11-13T02:40Z covers eight configurations. It is Intelligence Index v5.0.1 with 78 ranked models, at standard list prices. The costs below are AA task costs, not AIHubMix bills.
| Model / effort | AA rank | AA Index | $/task | Min/task | First answer token |
|---|---|---|---|---|---|
| Carbon high | 4 | 60 | 5.48 | 28.1 | 63.37 s |
| Carbon medium | 6 | 58 | 4.68 | 24.3 | 21.82 s |
| Carbon low | 14 | 54 | 5.22 | 24.7 | 13.61 s |
| Hydrogen xhigh | 16 | 54 | 3.16 | 13.5 | 37.77 s |
| Hydrogen high | 21 | 51 | 1.87 | 8.3 | 18.89 s |
| Hydrogen medium | 39 | 41 | 0.80 | 3.7 | 11.80 s |
| Hydrogen low | 60 | 27 | 0.29 | 1.3 | 5.71 s |
| Argon high | 22 | 50 | 4.97 | 30.6 | 41.30 s |
Three comparisons carry most of the routing.
Carbon high against Hydrogen xhigh. Carbon is six index points ahead, costs $2.32 more per task (73.4% more), takes 28.1 minutes against 13.5, and waits 63.37 seconds for its first token against 37.77. The compare page has both listings side by side.
Hydrogen high against Argon high. Hydrogen is one point ahead, costs $3.10 less per task (62.4% less), and finishes in 8.3 minutes against 30.6; its first token arrives at 18.89 seconds against 41.30. On this text-heavy index, the cheaper model is ahead on all three axes; the compare page is the cost check.
Carbon high against Argon high. Carbon is ten points ahead for $0.51 more per task (10.3% more), and finishes slightly faster, 28.1 minutes against 30.6, even though its first token is slower.
Two anomalies matter when routing rules get written. Carbon low scores the same 54 as Hydrogen xhigh, but at $5.22 per task against $3.16 and 24.7 minutes against 13.5; it also scores below Carbon medium (54 against 58) at a higher cost per task ($5.22 against $4.68). And the same Hydrogen covers 27 to 54 across its own settings. A rule keyed on the model name without the effort level is under-specified.
The bill changes on December 1
Google's pricing page lists Hydrogen at $1.50 input, $7.50 output and $0.075 cached input, flat across the full context window. Carbon and Argon are $2 input, $10 output and $0.10 cached through November 30, then double. Hydrogen has no announced change.
| Model | Input | Cached input | Output | When |
|---|---|---|---|---|
| Hydrogen | $1.50 | $0.075 | $7.50 | now, flat across 1M |
| Carbon / Argon | $2.00 | $0.10 | $10.00 | through 2026-11-30 |
| Carbon / Argon | $4.00 | $0.20 | $20.00 | from 2026-12-01 |
| Hydrogen Batch / Flex | $0.75 | $0.0375 | $3.75 | now |
| Carbon / Argon Batch / Flex | $1.00 | $0.05 | $5.00 | through 2026-11-30 |
| Carbon / Argon Batch / Flex | $2.00 | $0.10 | $10.00 | from 2026-12-01 |
Official prices in USD per million tokens. The current gap is 25% per token; from December 1 it is 62.5% for input and output. If Carbon and Argon double while Hydrogen holds, their AA task costs would also roughly double at the same token mix — about $10.96 for Carbon high and $9.94 for Argon high. That is a computed estimate, not an AA measurement and not an AIHubMix bill.
AIHubMix lists all three at Google's current prices. The free tier is Google's only; the gateway is a paid route.
Routing: plan on one, execute on the other
Google's release notes give the division directly: "Route planning turns to gemini-4-carbon and execution turns to gemini-4-hydrogen."
Antigravity's Orchestrator is that sentence shipped as a product: "a dynamic, bidirectional approach between Carbon and Gemini 4 Hydrogen within the evolution of a single trajectory." Carbon explores dependencies, resolves ambiguity and plans; Hydrogen handles edits, package updates and tests; when Hydrogen gets stuck, the work returns to Carbon to repair the plan. Google reports "near-Carbon task success rates" with 24% less quota consumed and 19% faster completion on average. The benchmark is Terminal-Bench 4.0 in the launch post and 4.1 in Antigravity's post, so the version is unsettled. Orchestrator is not available through AIHubMix.
The portable form:
| Workload | Route | Reason |
|---|---|---|
| Hard debugging and architecture | Carbon high | 60 index, the ceiling |
| Execution turns, tool loops, tests | Hydrogen high | 51 at $1.87, 8.3 minutes |
| High-volume text | Hydrogen medium | 41 at $0.80 per task |
| Image, audio, video or PDF input | Any of the three | all list the same inputs |
| Responses or Anthropic Messages clients | Carbon or Argon | both list four protocols |
| Planning and execution in one trajectory | Carbon plus Hydrogen | Google's routing advice |
"Hard debugging" is not a precise threshold. The honest test is whether Hydrogen high or xhigh passes your evaluation, not whether the task sounds hard.
What is documented, what is not
- The AA figures are a single live snapshot: v5.0.1, 78 ranked models, 2026-11-13T02:40Z. They drift; re-check before routing production traffic.
- The snapshot covers Argon at
highonly, so Argon's other effort levels cannot be compared with Carbon or Hydrogen from these figures. - AA task costs are measured at standard list prices, not AIHubMix bills.
- Vendor benchmarks are self-reported. Hydrogen's model card reports Terminal-Bench 4.1; Carbon's reports Terminal-Bench 4.0. Do not compare the two rows directly.
- AIHubMix's protocol coverage differs by model. All three list multimodal input; Hydrogen is chat-completions-only in the current documentation.
- Hydrogen is a "short-term generally available (GA) release subject to a minimum 45-day retirement notice policy." Carbon and Argon have no listed shutdown date.
- No live tests were run for this post.
Keep reading: the Gemini 4 Hydrogen series
- If the routing question starts with leaving Gemini 3.8 Flash, the first post covers the launch, the price change and the benchmark trade-offs: Gemini 4 Hydrogen vs Gemini 3.8 Flash: Is the Upgrade Worth It?
- If you are choosing an effort setting before committing to a tier, the second post measures all four Hydrogen levels: Gemini 4 Hydrogen Thinking Levels: low, medium, high, xhigh
Sources
- Google AI Studio, Gemini 4 Hydrogen: what's new
- Google, Gemini 4 Carbon model page
- Google, Gemini 4 Argon model page
- Google Antigravity, Gemini 4 Hydrogen and Orchestrator
- Artificial Analysis, Gemini 4 Hydrogen — live snapshot 2026-11-13T02:40Z, Intelligence Index v5.0.1
Top comments (0)