Gemini 4 Hydrogen: Cost Math for Long-Running Incident Agents

On November 12, 2026, Google released Gemini 4 Hydrogen, which Google calls "a major standalone upgrade for teams migrating from Gemini 3.8 Flash." The Gemini API changelog lists gemini-4-hydrogen as generally available on the same day. The price is the headline for anyone who runs agents: $1.50 per million input tokens and $7.50 per million output tokens, "priced flat across its entire 1 million-token context window."
An incident agent is a long-context workload: every turn sends the logs, metrics and kubectl output read so far back to the model. This post works through the numbers, then gives rules for when Hydrogen should do the work and when Gemini 4 Carbon or another strong model should plan or review it.
What Google Shipped
The facts below come from the Hydrogen overview in the Gemini API docs, the model page and the pricing page.
| Gemini 4 Hydrogen | Gemini 4 Carbon | |
|---|---|---|
| Model ID | gemini-4-hydrogen |
gemini-4-carbon |
| Input per 1M tokens | $1.50 | $2.00 until Nov 30, 2026, then $4.00 |
| Output per 1M tokens (thinking included) | $7.50 | $10.00 until Nov 30, 2026, then $20.00 |
| Cached input per 1M tokens | $0.075 | $0.10, then $0.20 |
| Context window | 1,048,576 tokens, shared by input, thinking and output | 1M tokens |
| Default output limit | 65,536 tokens (up to 1M with long decode) | 64K / 1M |
| Thinking levels | low, medium, high (default), xhigh |
Not compared here |
| Caching | Implicit only, prompts of 4,096 tokens or more | Not compared here |
Google describes Hydrogen as built for "speed and efficiency for everyday coding and agentic workflows." It says the model "batches read and search tool calls in parallel" and "cuts fabricated claims by more than 5x compared to Gemini 3.8 Flash by flagging missing context instead of guessing." For an SRE agent, that last point matters more than raw speed: an agent that says "I have no logs for that pod" is safer than one that invents a cause.
The launch post also claims Hydrogen finishes tasks "2.5 times faster at its high default setting" (in its section on teams moving from Gemini 3.8 Flash, without naming the baseline), scores 57% on Terminal-Bench 4.1 against 18.2% for 3.8 Flash, and edges past Gemini 4 Argon on the Artificial Analysis Intelligence Index (51 against 50 at high, 54 at xhigh). These are Google's numbers in Google's launch post. The post links its benchmarks to leaderboards but does not say who ran most of the tests. The claim of "often completing tasks faster than models in its class" names no competitor. Treat all of it as a reason to test, not as a result.
Worked Example: One Incident Investigation
Take the same investigation we used in Where AI Agent Tokens Go: 40 tool calls, a context that grows from 20K to 150K tokens, about 3.4 million input tokens in total because each turn re-reads the context, and 60K output tokens.
| Model and price | No cache hits | 95% of re-reads hit the cache |
|---|---|---|
| Hydrogen ($1.50 / $7.50) | $5.10 + $0.45 = $5.55 | $0.24 + $0.23 + $0.45 = $0.92 |
| Carbon, intro price ($2 / $10) | $6.80 + $0.60 = $7.40 | $0.33 + $0.30 + $0.60 = $1.23 |
| Carbon from Dec 1 ($4 / $20) | $13.60 + $1.20 = $14.80 | $0.65 + $0.60 + $1.20 = $2.45 |
The cache column assumes 3.25M tokens are billed at the cached rate and the remaining 150K at the full input rate. For comparison, Anthropic's pricing page lists Claude Opus 5.5 at $4 / $20 and Sonnet 5.5 at $2 / $10, so after December 1 Carbon costs the same per token as Opus 5.5.
Three things stand out:
- Hydrogen is about 2.7 times cheaper than Carbon on the December price list, with or without cache hits. In November, Carbon's intro price hides most of that gap. Routing rules you tune this month will be wrong on December 1.
- Cache hits matter more than the model choice. Hydrogen without cache hits costs more than Carbon with them. Hydrogen supports only implicit caching, so you cannot pin a prefix. You can only keep the prefix stable and check the cached token counts the API returns.
- Flat pricing is not flat cost. No long-context surcharge means a 900K-token prompt pays the same rate per token as a 9K one. It still pays $1.35 of input for that one turn without a cache hit, and it pays it again on the next turn.
Planner and Executor: Google's Own Routing Advice
Google's docs suggest a split: "Have Carbon plan system interfaces or resolve hard roadblocks" and "delegate code edits, parallel subagent tasks, and test runs to Hydrogen." The migration section describes routing planning turns to gemini-4-carbon and execution turns to gemini-4-hydrogen in one session.
Applied to the investigation above, with Carbon at December prices and no cache hits:
- Carbon writes the plan in 4 turns at about 25K context (100K input, 12K output): $0.40 + $0.24.
- Hydrogen runs 35 tool-calling turns (about 3.15M input, 45K output): $4.73 + $0.34.
- Carbon reviews the findings once at 150K context (150K input, 3K output): $0.60 + $0.06.
The total is about $6.36, against $5.55 for Hydrogen alone and $14.80 for Carbon alone. The planner and reviewer add about 15% to the cheap run. Whether that is worth it depends on how often the cheap run alone gets the root cause wrong.
Cost per Solved Task, Not per Token
A wrong root cause costs more than tokens. It costs an engineer's time, and sometimes a second incident. A simple model:
expected cost per incident = model cost + failure rate × cost of a failure
Assume a failure means 30 minutes of on-call time, and value that at $50. The success rates below are hypothetical. Measure your own.
| Setup | Model cost | Success rate (example) | Expected cost |
|---|---|---|---|
| Hydrogen alone | $5.55 | 70% | $5.55 + 0.30 × $50 = $20.55 |
| Carbon plans and reviews, Hydrogen executes | $6.36 | 82% | $6.36 + 0.18 × $50 = $15.36 |
| Carbon alone | $14.80 | 85% | $14.80 + 0.15 × $50 = $22.30 |
The break-even is easy to compute: the $9.25 price gap between Carbon and Hydrogen equals 18.5 points of success rate at $50 per failure. If Carbon does not fix that many more incidents than Hydrogen on your alerts, Hydrogen is the cheaper agent. This is the reason we score models on real incidents in SRE Bench rather than trust a single index number.
Routing Rules for an SRE Agent
| Task | Model | Why |
|---|---|---|
| Collect logs, metrics, events; run read-only commands | Hydrogen | Many turns, large context, low reasoning per turn |
| Summarize a long log or trace dump for the main agent | Hydrogen | Pure reading, where the flat price helps most |
| Form the investigation plan for a novel or multi-service incident | Strong model | A bad plan wastes every later turn |
| Second opinion on the root cause before it goes to a human | Strong model | One short turn, high value |
| Propose a change to production (rollback, scale, config) | Strong model, then a human | The cost of a wrong action dwarfs token cost |
| Security incident analysis | Strong model, and check the vendor's policy | Google says Carbon spends more thinking tokens on "security audits" |
Traps Before You Switch
Thinking tokens are output tokens. The $7.50 output price includes thinking, and Hydrogen defaults to high, not medium. Google advises reserving xhigh for problems "where high falls short." Log output tokens per run before and after a level change.
minimal is rejected. The docs say Hydrogen "rejects minimal with an error." If your agent configuration sets a thinking level of minimal from an older Flash model, the first call to Hydrogen fails. Use low instead.
The free tier trains on your data. The pricing page marks "Used to improve our products" as Yes for the free tier and No for paid. Incident data contains hostnames, customer IDs and sometimes secrets. Use a paid key for anything that touches production.
Implicit caching needs a stable prefix. It starts only at 4,096 tokens, and anything that changes the start of the prompt (a timestamp in the system prompt, a changed tool list) turns cache hits into full-price input.
Where Akmatori Fits
Akmatori supports Google as an LLM provider, and in Settings → AI models you can type a model name that is not in the built-in list yet, such as gemini-4-hydrogen. Test it on a staging alert before you rely on it. The API reference shows what you need to compare models: GET /api/incidents/:uuid returns the agent runs with their token count and execution time, and PUT /api/settings/llm/:id sets the model and thinking_level of a configuration.
One LLM configuration is active at a time for all agent runs (PUT /api/settings/llm/:id/activate), so Akmatori does not split planning and execution turns between models inside a run. To compare Hydrogen with a stronger model, make each one active for the same class of alerts and compare cost and correct root causes per incident.
Conclusion
Gemini 4 Hydrogen makes the reading-heavy part of incident work cheap: $1.50 per million input tokens with no surcharge up to 1M tokens. It does not make the context free, it bills thinking as output, and its benchmark claims are Google's. Use it to execute, keep a stronger model for plans and reviews, and decide with cost per solved incident on your own alerts.
If your team wants AI-assisted incident workflows on your own infrastructure, Akmatori helps SRE teams investigate alerts, coordinate response, and automate safe infrastructure actions. Powered by Gcore for global infrastructure reliability.
