For development teams, the useful questions are practical: what does it cost, where can it run, and how often does its output survive testing?
Pricing is published—and access has boundaries
Google’s announced API rates are $1.50 per million input tokens and $7.50 per million output tokens, with flat pricing across the million-token context window. That provides a concrete starting point for evaluation, rather than an unspecified promise of affordability.
Google’s developer announcement lists the Gemini API, Google AI Studio, Android Studio and Gemini for business as access routes. Antigravity access requires a paid Google AI Pro or Ultra subscription; the launch announcement also specifies the paid tier for Android Studio. Free Gemini app users receive Hydrogen through automatic model selection when Google considers it appropriate, not a guaranteed manual selection option.
The distinction matters: availability across several products does not mean identical access or billing in each.
A coding agent, not just a code generator
Google describes Hydrogen’s developer workload as repository coding, debugging, interactive web development and multimodal tasks. It offers four configurable thinking levels, allowing developers to choose between different reasoning depths. Those are intended capabilities, not guarantees that a generated patch is production-ready.
The model card adds an important technical boundary: Hydrogen is adapted from Carbon and has a combined input-output context window of up to one million tokens. Default synchronous output is capped at 64,000 tokens; longer output, up to one million tokens, uses background long decode. A large context window should therefore not be mistaken for an unrestricted million-token response in an ordinary synchronous request.
For teams evaluating large-repository workflows, that distinction belongs in the design review—not in the footnotes after implementation.
The benchmarks show a mixed picture
Google DeepMind’s model card reports a 13.5% SWE-sweep score for Hydrogen, compared with 3.1% for Claude Sonnet 5.5 and 5.8% for GPT-6.1 Terra. But its Terminal-Bench 4.1 table puts Hydrogen at 57.0%, below Sonnet 5.5’s 63.9%; it lists no corresponding Terra result. These are published evaluation figures, not a universal ranking of coding assistants.
The sensible reading is narrower than “Google wins coding.” Hydrogen warrants testing against the tasks a team actually performs. A model that excels at discovering repository bugs may not lead on every terminal-driven workflow.
What teams should measure before switching
The following is an editorial evaluation checklist, not a Google setup procedure:
- Use representative issues. Include small fixes, ambiguous bug reports and changes spanning multiple files.
- Define success before running the agent. Require the intended fix, passing tests and no unrelated changes.
- Measure cost per accepted change. Record retries and human review time alongside token charges.
- Keep the toolchain consistent. Compare models using the same repository snapshot, permissions and test commands.
- Review failures separately. Distinguish incorrect code from tool failures, incomplete work and misleading completion claims.
Google’s own model card supports that caution. It acknowledges hallucinations, occasional slowness or timeouts, and increased token use at higher thinking levels. It also lists a March 2026 knowledge cutoff and says the full safety and security report is forthcoming.
The bottom line
Hydrogen’s launch gives developers a clearly priced model to evaluate for iterative coding work. It does not establish that replacing an existing assistant will lower a particular team’s total costs.
The meaningful benchmark is the accepted patch: correct, tested, reviewable and delivered without excessive retries. If Hydrogen improves that outcome, its efficiency argument has substance. If not, a lower token bill is merely a cheaper route to another debugging session.