Gemini 4 Hydrogen vs Argon, Carbon, and 3.8 Flash

Gemini 4 Hydrogen is Google's fast agent tier at $1.50/$7.50 per million tokens; Carbon and Argon remain the depth models for hard reasoning — and Orchestrator is how Google wants you to mix them. Read the November 12 launch story first, then use this table to tune routers already running Argon or Carbon.

Update (November 12, 2026): Artificial Analysis cost per task and where Argon still wins

The Decoder's write-up pulls together Artificial Analysis (AA) data that adds cost per task to the comparison. Per that report, AA measures Hydrogen at about $3.16 per task at xhigh (54 index points), $1.87 at high (51), $0.80 at medium (41) and $0.29 at low (27). Xhigh uses nearly twice the output tokens of high, which drives the price. AA places Hydrogen on its Pareto frontier at high, medium and low.

Argon is not beaten everywhere. On Terminal-Bench 4.0 as run by AA, the report lists Hydrogen at 54% against Argon at 57.1% and Carbon at 65.2%. Hydrogen leads Argon on Agents' Last Exam v1.1 (71.6% vs 69.0%), OSWorld-2.1 offline subset (77.6% vs 74.7%), SWE-sweep (13.5% vs 4.3%) and Vending-Bench 2 (final balance $14,549 vs $13,718). Carbon stays ahead of both on Agents' Last Exam (77.4%), OSWorld-2.1 (81.5%) and Vending-Bench 2 ($16,484), but trails Hydrogen on SWE-sweep (9.6%). AA also reports an 8% hallucination rate for Hydrogen on its AA-Omniscience test. The Decoder attributes the launch claims and the Terminal-Bench 4.1 score (57%) to Google, and the index, hallucination-rate and cost figures to AA. The article does not attribute each table row, so check the original charts before you quote one.

TL;DR — which model for which job

Price per million tokens (published sources)

Cost-per-task example (illustrative only): a 40k input / 8k output agent step costs about $0.12 at Hydrogen list rates versus $0.16 at Argon intro input/output math — before caching, thinking surcharges, or batch discounts.

Model routing switchboard choosing fast and frontier Gemini tiers

Benchmark rows Google published for Hydrogen

These numbers are copied from Google's November 12 launch post, not re-run by explainx.ai:

If you maintain a frontier comparison hub, add Hydrogen as the Google fast slot, not as a replacement for Argon leaderboard rows.

Hydrogen vs Argon — product positioning

Argon shipped first as Google's Gemini 4 frontier narrative: million-token outputs, Fairwind cyber paths, and pricing aimed at maximum capability (launch coverage).

Hydrogen ships second as execution:

• Repository coding and debugging
• Interactive web development
• Multimodal agent workflows with four thinking levels

Google's surprising claim is that Hydrogen edges Argon on the Artificial Analysis Index at some thinking settings while staying cheaper — meaning some teams might downgrade routine steps without giving up score on vendor charts. Production teams should still keep Argon (or Carbon) routes for single-shot hard tasks until private evals agree.

Hydrogen vs Carbon — when depth still wins

Carbon remains the cyber-defender-first frontier story with its own pricing and access threads. Orchestrator explicitly pairs Carbon + Hydrogen, implying Carbon still owns coordination and judgment while Hydrogen runs workers.

Use Carbon when:

• You need the strongest Google chart row on deep reasoning benchmarks Carbon marketed at launch
• Compliance mandates the same model across long audit logs
• You are not on Antigravity and cannot automate switching

Use Hydrogen when:

• Steps are short, parallel, and tool-heavy
• Cost per merged PR dominates latency sensitivity

Hydrogen vs Gemini 3.8 Flash — migration stakes

Before November 12, many routers used 3.8 Flash as the Google fast default. Google's launch compares Hydrogen directly:

• 2.5× faster tasks at default high thinking
• Terminal-Bench 4.1 more than triples (57% vs 18.2%)
• Cleaner recovery when tools fail or search is empty

That is a strong reason to re-benchmark any workflow you tuned on Flash-specific quirks. It is not a reason to skip regression tests — Modern Web Guidance shows even frontier models fail half of web checks without skills.

Routing patterns that match Google's story

1. Single-model Hydrogen — cron jobs, codegen workers, spreadsheet agents with tight budgets.
2. Hydrogen verifier + Carbon author — Carbon drafts architecture; Hydrogen runs tests and linters (manual orchestration if you are off Antigravity).
3. Orchestrator automatic — Antigravity-only; monitor quota dashboards weekly.
4. Search vs API split — consumers hit Hydrogen in Search; your API stack may still be on Flash until IAM and billing catch up.

Limits and unknowns

• Vertex lag — GCP enterprise dropdowns may trail AI Studio (precedent on Carbon).
• Third-party routers — OpenRouter-style catalogs may miss Hydrogen or rename ids.
• Thinking defaults — higher thinking levels can erase Hydrogen's latency advantage.
• Enterprise Sheets — Workspace rollouts can change employee-visible behavior without API changes.

Hands-on next steps

1. Copy Google's $1.50/$7.50 row into your token economics spreadsheet.
2. Run 10 paired prompts on Hydrogen vs your current Flash alias.
3. Enable Orchestrator on one non-production Antigravity repo; compare quota to Carbon-only.
4. Read official demos for UI tasks that exaggerate Hydrogen's formatting strengths.

Worked cost scenarios (math only)

Assume list API rates with no caching and no batch discount:

Hydrogen wins on per-step math; Argon/Carbon may still win on steps avoided if a single deep pass replaces ten shallow retries. Measure both.

Thinking levels and latency tradeoffs

Hydrogen exposes four thinking levels, per Google's developer post, in developer docs linked from the launch post. Higher levels track Google's xhigh Index 54 marketing point but can erase the 2.5× speed advantage over 3.8 Flash.

Practical rule: default workers to mid thinking, escalate only when tests fail twice, and keep Carbon/Argon routes for escalation — similar to Claude Fable vs Opus routing.

Consumer vs API routing

How Hydrogen fits next to non-Google fast models

Teams comparing only inside Google should still sanity-check external fast tiers you already route — small-model comparison posts show how crowded the Flash class is. Hydrogen's pitch is Gemini 4 intelligence at Flash-like bills inside Google's agent stack, not universal domination on every third-party leaderboard.

Failure modes when downgrading Argon workers to Hydrogen

• Under-thinking on security patches — depth models catch subtle auth bugs; validate with agent security checklists.
• Over-trust in vendor honesty metrics — rerun hallucination tests on your docs.
• Orchestrator opacity — if engineers cannot see which model answered, incident reviews get harder; log model ids per step where Antigravity exposes them.
• Mixed regional availability — Vertex delays may strand one region on Flash while another uses Hydrogen.

FAQ for platform owners

Should we delete Flash routes? Keep them behind a flag until Hydrogen id is stable in every environment you use — staging, prod, and CI.

Does Hydrogen change Embeddings or RAG stacks? This launch targets generative/agent endpoints; re-validate retrieval pipelines separately.

Will Search Hydrogen answers match API outputs? Stein's post emphasizes fan-out and visuals; do not assume parity with AI Studio transcripts.

Edge tier models beside frontier cores in a Gemini routing diagram

Carbon intro pricing reminder

Carbon's consumer and CLI paths introduced time-limited promos separate from Hydrogen's flat $1.50/$7.50 story. If your router keys off lowest promotional token, recompute after promo expiry — Hydrogen may become the stable cheap tier even when Carbon promos undercut it temporarily.

Sheets and Search do not use your API router

Enterprises often discover employees consuming Hydrogen through Workspace or Search while engineering still pays for 3.8 Flash keys. Align admin console settings with cloud billing exports so FinOps is not surprised by parallel fast tiers.

Benchmark discipline

Google published multiple charts on November 12 — Terminal-Bench, AutomationBench-AA, HLE-Diamond, and Artificial Analysis rows. When you file internal comparisons:

1. Tag each cell with Google blog vs independent rerun.
2. Keep Terminal-Bench version numbers in filenames (4.0 vs 4.1 differ).
3. Store prompts under version control, as we recommend for eval harnesses.

When not to switch yet

• Compliance mandates Argon/Carbon-only until legal reviews Hydrogen data handling.
• Regional projects without Hydrogen ids in Vertex.
• Fine-tunes or LoRA stacks tied to Flash embeddings unrelated to Hydrogen.
• Latency SLOs measured on 3.8 Flash hardware affinity you have not retested.

Summary

Hydrogen is the default fast Gemini 4 worker in Google's November 12 story; Argon and Carbon remain depth; Orchestrator is the Antigravity-only glue. Price-sensitive agent loops should start Hydrogen evals this week, but production cutovers deserve the same rigor as any model migration.

Companion: Hydrogen launch guide · Showcase demos

Related reading

• Gemini 4 Hydrogen launch
• Pre-launch claimed vs verified leaks
• Gemini 4 Argon launch
• Carbon CLI access
• Hydrogen demos list
• Managed Agents sandboxes

All pricing and benchmark cells cite Google's November 12, 2026 posts unless noted. Re-validate before locking production routers.