Gemini 4 Hydrogen: $1.50/$7.50 and When to Drop 3.8 Flash

Google DeepMind released Gemini 4 Hydrogen on November 12, priced at $1.50 per million input tokens and $7.50 per million output tokens.

The model is aimed at everyday coding and agentic execution rather than frontier reasoning, and Google says it beats one of its own frontier models on a composite intelligence index. It is available the same day across developer, enterprise and consumer surfaces.

Flat pricing across the full 1M context window

The pricing detail that matters for agent builders is not the headline rate but its shape. Google states Hydrogen is "priced flat across its entire 1 million-token context window" at those rates, so a long request costs a predictable multiple of a short one.

Hydrogen pairs a native 1 million-token input with up to 1 million output tokens, which Google positions for work over dense PDFs, spreadsheets, charts, audio and recorded video.

The benchmark line Google leads with

Google says Hydrogen surpasses Gemini 4 Argon on the Artificial Analysis Intelligence Index at its default high thinking level, 51 versus 50, and reaches 54 at its highest "xhigh" thinking setting.

That framing deserves care. Argon and Gemini 4 Carbon remain the company's frontier models for what it calls "deep, long-horizon reasoning on your hardest problems," and a single composite index is a poor proxy for long-context architectural work. The claim is that a cheaper execution model now edges a frontier model on one aggregate score, not that it replaces it.

Elsewhere Google reports Hydrogen ranks #5 on Agent Arena, leads its class on HLE-Diamond at 58.8%, ranks #1 on the Vals Tax Agent Bench at 52.3%, ties for #1 on the Vals Legal Research Bench at 55.8%, and scores 53.4% on the new Multimodal Knowledge Agent benchmark and 71.6% on Agents' Last Exam v1.1.

What changes if you are running 3.8 Flash

Google pitches Hydrogen primarily at teams migrating from Gemini 3.8 Flash, and the comparison numbers are the most concrete in the announcement:

• Tasks finish 2.5 times faster at the high default setting.
• Terminal-Bench 4.1 scores 57% versus 18.2% for 3.8 Flash.
• Fabricated claims are cut by more than 5x.

On token spend, Google cites early partner testing across complex graphs, SQL and data-heavy enterprise workflows where Hydrogen ran "2x to 2.8x faster while reducing token usage by 12% to 23%, compared to 3.8 Flash." Those are vendor-run partner tests, not independent measurements, and the workloads are described only in general terms.

Google also says the model batches file-read and search calls in parallel rather than inspecting files one at a time, and that it makes focused edits on issue resolution without rewriting unrelated files.

Orchestrator switches models inside one session

Alongside the model, Google introduced Orchestrator in Google Antigravity. Instead of one model running a task end to end, it "dynamically switches between models like Carbon and Hydrogen within the same session."

On Terminal-Bench 4.0, Google reports Orchestrator reached near-Carbon task success rates while consuming 24% less quota and completing 19% faster on average. The suggested parallel pattern puts Carbon in the coordinator role with Hydrogen running worker and verifier agents across isolated workspaces.

For teams that already hand-rolled a model router, that is the same architecture with vendor tooling attached — worth benchmarking against your own routing rules before adopting it.

Agent safety figures, self-reported

Every number in this section comes from Google's own evaluations. The company says Hydrogen reduces overreach in agentic tasks by up to 81% and under-service by up to 79%, and cuts misaligned actions and task shortcutting across tool-use and coding evaluations by 12x.

It reports a 96.7% honesty rate on an internal agent honesty evaluation, and that the model is up to 2x less likely to bypass permission boundaries when a task is impossible or blocked. Google also claims Hydrogen "preserves 99.9% chain-of-thought legibility with zero deceptive reasoning edits across more than 200,000 audited turns," and is more than 2.3x more resilient on the Gray Swan indirect prompt injection benchmark. False refusals on HealthBench Pro are reported under 0.4%.

It shipped into a crowded day

Hydrogen landed the same day NVIDIA expanded its Nemotron family with open-weight releases — Nemotron 3.5 Super VL at 120B, Nemotron 3.5 Bolt in 74B and 4B variants, and a 30B mixture-of-experts speech model — all published on Hugging Face. Teams weighing a hosted execution model against self-hosted weights now have both options dated to the same week.

Where to run it, and what to verify

Hydrogen is live in the Gemini API in AI Studio and the paid tier in Android Studio, in Google Antigravity for paid AI Pro and Ultra subscribers, and in Gemini for business via the API or model dropdown. In the Gemini app, Google notes free users reach Hydrogen through automatic model selection when it suits the prompt. Search is being upgraded with it globally.

The number to test first is the token reduction. Replay a hundred real production requests against both Hydrogen and 3.8 Flash and compare input tokens, output tokens and wall-clock time. The 12% to 23% figure is a partner result on unspecified workloads, and it either reproduces on yours or it does not.