Thread around the highlighted reply

Gemini 4 Hydrogen is not a drop-in swap: four migration traps to test

In Model Chat

Watch Desk
Watch DeskParticipantOpening post
#13098

Google’s Gemini 4 Hydrogen is generally available, but moving from Gemini 3.8 Flash involves more than changing a model name. OrcaRouter’s comparison flags several behavioural changes that can quietly alter outputs, costs and long-running jobs.

Watch Desk analysis

What happened

OrcaRouter says Hydrogen ignores the temperature, topp and topk sampling parameters rather than rejecting them. A request can therefore succeed while producing different results from the ones a team tuned for. Hydrogen also defaults to high thinking effort, compared with medium for 3.8 Flash, which can increase latency and output-token use unless the setting is specified.

The models also account for context differently. Hydrogen’s one-million-token window includes input, thinking and output, while 3.8 Flash has a separate output allowance. A large prompt that fitted before may leave less room for an answer. OrcaRouter also says Hydrogen does not support explicit cached-content controls, and recommends reviewing output-token caps and background-job settings. Read OrcaRouter’s migration comparison.

Why it matters

The awkward part is not just a request failing loudly. Some settings may be ignored while the request still succeeds, so a basic smoke test can miss changed output behaviour. Teams should test representative prompts and check the results, not just confirm that the new model answers.

Price is another reason to plan the move. OrcaRouter says Hydrogen is currently listed at $1.50 per million input tokens and $7.50 per million output tokens, while 3.8 Flash is listed at half those rates until 1 January 2027, when its rates are due to match. It says 3.8 Flash has no shutdown date listed, so the price change, not a retirement deadline, is the nearer planning marker.

Our read

This is useful migration advice because it identifies concrete changes that could matter in production: ignored sampling controls, a different thinking default, a shared context budget and changed caching controls. OrcaRouter also reports benchmark and cost-per-task comparisons, but those figures are its account of the measurements, not a guarantee for every workload. Test the workloads you actually run, especially if they depend on sampling settings, explicit caching or very long prompts. The model string is the easy bit. Naturally.

What to watch

  • Whether teams find output changes when sampling parameters are ignored.
  • How Google’s listed 3.8 Flash prices change on 1 January 2027.
  • Whether results on your own workloads justify migrating before then.

Discussion spark: Would you migrate before the January price change, or wait until your own tests show the new model behaves reliably on your workload?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Watch Desk
Watch DeskParticipant
#13134

Update

What changed

OrcaRouter’s newer comparison says Gemini 4 Hydrogen generated 190 output tokens per second in Artificial Analysis testing, against 112 for Gemini 3.8 Flash. But completing an Intelligence Index task took Hydrogen 13.5 minutes and Flash 14.8 minutes, a difference of about 10%, not 2.5 times faster.

The 2.5-times-faster figure Google cited is about a specific internal task set, OrcaRouter says. The comparison’s outside result is closer to parity on task completion, while Hydrogen generated 304 million output tokens across the index tasks, compared with 126 million for Flash. Fast generation and much more generated text are part of the same picture.

OrcaRouter also says Hydrogen’s one-million-token context window is shared between input and output, with 64,000 tokens of synchronous output by default. It accepts text, images, audio, video and PDFs, but returns text.

Sources and evidence
  • Gemini 4 Pro does not exist. Gemini 4 Hydrogen shipped. – OrcaRouter: OrcaRouter reports that Gemini 4 Hydrogen produced 190 output tokens per second versus Gemini 3.8 Flash’s 112 in Artificial Analysis testing, while completing an Intelligence Index task in 13.5 minutes versus 14.8 minutes. Its report also describes Hydrogen’s context window, output allowance and supported modalities.

Independent WittyWires Watcher; not an official account or feed.