The short version: the model that sets your agent's bill is not the flagship you call a dozen times a day — it's the fast tier you call ten thousand times inside the loop. So the model choice that actually moves your invoice is which cheap-fast model sits at the bottom. As of November 12, there are three obvious candidates, and they line up cleanly on price:

Pick the cheapest one that clears each step's bar — and route the rest. Here's how to read the three.

Price first, because price is the whole point of this tier#

The reason you choose a fast tier at all is cost, so start there. Haiku is roughly 15x cheaper than Hydrogen on input and 20x cheaper than Sonnet. That multiplier is not a rounding detail when a single agent task fans out into hundreds or thousands of model calls — a plan step, a dozen tool calls, a re-read of context, a retry. The bottom tier is where almost all of an agent's tokens are spent, which is why getting it right beats almost any optimization you'll make at the top.

Google's new Hydrogen lands in the middle: more expensive than a mechanical workhorse, cheaper than a full mid-tier generalist, and pointed specifically at fast coding and agent loops — debugging, repository-level edits, web dev, multi-step tool use. It replaced Gemini 3.8 Flash on November 12 and Google says it's about 2.5x faster than Flash at default reasoning. The capability benchmarks making the rounds (an Artificial Analysis Intelligence Index around 51, a jump on Terminal-Bench) are mostly Google's own or early-partner numbers, so treat them as a reason to test, not a reason to trust.

Capability runs roughly the other way#

Here's the tension: by raw capability the order roughly reverses. Sonnet 5.5 is the strongest general reasoner of the three. Hydrogen is a fast specialist — excellent on the coding and tool-calling it's tuned for, less of an all-rounder. Haiku is the fastest and cheapest but hits a ceiling on hard, multi-step reasoning and long-horizon work.

That's why "which is best" has no answer and "best for what" has three:

If the step is…UseBecause
Mechanical: extraction, classification, routing, short tool callsHaiku 5.515–20x cheaper, and these steps don't need deep reasoning
Fast coding: repo edits, web dev, tool-heavy agent loopsGemini 4 HydrogenTuned and priced for exactly this, faster than its predecessor
Broader reasoning that isn't flagship-hardSonnet 5.5The strongest generalist here, without paying Opus/Fable rates
Long-horizon planning, gnarly debugging, high-stakes outputEscalate (Opus 5.5 / Fable 5.2 / Gemini Carbon)Only the hard 10% justifies $4/$20 and up

The real answer is a router, not a model#

The most important signal in the Hydrogen launch isn't the price or the benchmark — it's that Google shipped it inside its Antigravity coding tool with an Orchestrator that automatically routes each task to the right-sized model by difficulty. When the model vendor itself builds the router and sells it as the headline feature, that's the market telling you single-model selection is the wrong unit of decision.

So don't pick a model for your agent. Pick a default and an escalation rule: a cheap-fast tier (Haiku for mechanical loops, Hydrogen for coding-heavy ones) handles the high-volume majority, and anything that fails a confidence or complexity check gets kicked up to a heavier model. We've built this exact split step by step in cut your LLM bill with task-based routing and a three-tier Claude router; Hydrogen just gives you one more strong, cheap default to drop into the bottom tier, and Google's own Orchestrator is a preview of what you give up when you let the vendor route for you.

The one number that decides it#

The token price is only half the bill. The other half is retries and human cleanup. A model that's 15x cheaper but fails a third of the time on your workload is not a saving — it's a slower, more frustrating version of the expensive one. So the rule that survives every price change is this: use the cheapest tier that reliably clears each step's bar, and route everything else up. Benchmark all three — Haiku, Hydrogen, Sonnet — on your own agent's actual tasks, measure cost-per-successful-completion rather than cost-per-token, and let that number, not a leaderboard, set your default. If you're also weighing which paid coding assistant to run on top, we sort that out in which AI coding subscription a solo founder should pay for.