Google published a number this week that deserves more attention than it got: on Terminal-Bench 4.0, its new Orchestrator matched near-frontier task success while using 24% less quota and finishing 19% faster. Same results, less compute. The whole trick is switching between a big model and a cheap one mid-task.
The launch everyone covered was Gemini 4 Hydrogen, which shipped November 12. The interesting part is what Google shipped next to it: a router. And the routing pattern behind it is something you can copy in your own stack today, whether or not you touch a Google product.
What Hydrogen actually is
Strip the launch-page language and Hydrogen is a budget workhorse. It costs $1.50 per million input tokens and $7.50 per million output, and that price is flat across the entire 1M-token context window. No gradient pricing that punishes you for long documents.
The benchmark jump over the model it replaces is blunt. Gemini 3.8 Flash scored 18.2% on Terminal-Bench 4.1. Hydrogen scores 57%. Partners reported tasks finishing 2x to 2.8x faster with 12-23% fewer tokens. Fabricated claims dropped more than 5x. On the Artificial Analysis Intelligence Index it hits 51 at its default thinking level, 54 at the new xhigh setting — for reference, that default score sits ahead of Argon, one of Google’s own frontier models, at 50.
Read that again, because it’s the whole story: the cheap model now scores above a frontier model from the same family on a general intelligence index. That is why routing suddenly makes sense.
It comes with four thinking levels (low, medium, high, xhigh), so you pay for reasoning depth only when a task needs it. Google also published unusually concrete safety numbers: 81% less agent overreach, 12x fewer misaligned actions on tool-use and coding evals, 2.3x more resilience to indirect prompt injection on the Gray Swan benchmark, and 99.9% chain-of-thought legibility across 200,000 audited turns. It’s live in AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search, and Antigravity for paid Pro and Ultra subscribers.
The part that matters: Orchestrator
Buried next to the Hydrogen announcement is Orchestrator, now live inside Google Antigravity. Instead of picking one model for a session, it switches between Carbon (frontier) and Hydrogen (cheap) as the work unfolds. Planning a multi-step task? Carbon. Executing step four of a mechanical refactor? Hydrogen.
On Terminal-Bench 4.0 it hit near-Carbon success rates with 24% less quota burned and 19% faster completions.
I’ve seen teams build this by hand. A senior engineer writes a wrapper that sends the hard prompt to the expensive model and everything else to the cheap one, then spends a month tuning the threshold. Google shipping it as a product feature tells you the pattern has crossed from clever hack to default architecture. Frontier model as architect, budget model as executor.
Why this works now, and didn’t two years ago
Routing flops when the cheap model fails at the subtasks you hand it. Then you either retry with the big model (paying twice) or ship bad output (paying in bugs). Both eat the savings.
Hydrogen changes the math. A model at $1.50/$7.50 per million tokens that scores 51 on a general intelligence index and leads its class on HLE-Diamond (58.8%) fails rarely enough that the fallback path stays cheap. The economics only work because the floor rose. Every model family that ships a competent budget tier makes this pattern more accessible to people who don’t work at a lab.
One caveat I’d add: Google’s 24%/19% numbers are from its own benchmark on its own product. Treat them as directional, not contractual. Your workload’s split between planning and execution is what actually determines your savings.
How to copy the pattern without Antigravity
You don’t need Antigravity, and you don’t need Gemini at all. The pattern is model-agnostic. Here’s the shape of it.
Split by role, not by task. Most routing guides tell you to classify prompts: “summarization goes to the cheap model, coding goes to the big one.” That’s fragile. Real work mixes difficulty within a single task. Route by role instead: one model plans, decomposes, and reviews. Another executes the steps. This is closer to how Orchestrator works, and it survives tasks that don’t fit your classification table.
Send the plan, not the conversation. When the executor model starts a subtask, hand it the frontier model’s plan plus only the context it needs. Long chat history inflates input tokens on every hop, and stale context degrades output. A clean handoff is both cheaper and better.
Escalate on failure signals, not vibes. Define what “the executor struggled” means in code: a test that won’t pass after N attempts, output that fails your validation schema, a tool call loop. Those triggers hand the subtask back to the frontier model. Everything else stays cheap.
Dial thinking level per subtask. If you do use Hydrogen, the four thinking levels are a second routing axis. Mechanical extraction runs at low. Ambiguous debugging runs at high. Same model, tenfold difference in effective cost per token.
Measure the split before you optimize. Log which model handled what, and how many tokens each burned, for a week. I’d expect most teams to find that 70-80% of their tokens go to work a budget model can do. Until you’ve measured it, the routing layer is guesswork.
What to do this week
- Pull last month’s API bill and sort spend by endpoint or feature. Find the one workflow eating the most tokens.
- Estimate its plan-versus-execute split. If most tokens are mechanical execution, it’s your routing candidate.
- Write the dumbest possible version: frontier model produces a step list, budget model executes step by step, failures escalate back up. A weekend of code, no framework required.
- Run both paths in parallel for a week and compare cost and output quality. Keep whichever wins.
The teams already doing this cut their bills without shipping worse product. Google just made the pattern hard to ignore — the cheapest model in a family is now smart enough to execute what the smartest one plans, and the gap will keep widening in the wrong direction for anyone paying frontier prices for every token.
Source: Google’s Gemini 4 Hydrogen announcement (blog.google), November 12, 2026.
