Gemini 4 Hydrogen errors, limits and “Unavailable”

This is the most-reported Hydrogen problem on the Google AI Developers Forum since launch (dozens of threads from 12–14 November 2026, most from Google AI Pro accounts activated through the Reliance Jio 18-month offer, plus some directly paid Pro accounts). Antigravity’s Models documentation lists Hydrogen for Free, Plus, Pro, Ultra and Enterprise with no paid-only footnote, while Argon, Orchestrator (Argon) and Claude 5.5 carry a directly-billed / non-trial note. A forum Regular’s reply in one thread says Hydrogen is included in the baseline quota on all plans and suggests updating to Antigravity v2.31.2, opening Settings > Models, pressing the Refresh icon (or signing out and back in), and checking 18+ age verification and the Terms-of-Service country. Other users report it still shows Unavailable after that; one thread quotes the server labelling such accounts as Pro but returning Hydrogen as disabled. As of 14 November 2026 we found no Google statement confirming whether this is intended or a bug.

Source: Google Antigravity Docs, checked 14 November 2026 · Forum: Unfair Restriction of Gemini 4 Hydrogen for Jio-Bundled Google AI Pro Subscribers in Antigravity · Forum: [Bug] Gemini 4 Hydrogen locked for Jio Pro in Antigravity despite docs listing it on all plans · Forum: Jio 18-Month Google AI Pro Bundle Restricted from Gemini 4 Argon & Hydrogen in Google Antigravity

Several forum threads report this right after the 12 November rollout, from users of the official Antigravity app or VS Code extension. The reply in the thread says Hydrogen planning and execution in the official app is supported, to submit the official appeal form once (duplicates do not speed it up), that automated processing takes about 48–72 hours with no confirmation email and access restores on next sign-in if approved, and to remove third-party OAuth access in Google Account permissions before signing back in. Enterprise / Google Cloud accounts should contact Google Cloud Support. This is forum guidance, not a Google policy page.

Forum: [Appeal / Bug] False Positive 403 TOS_VIOLATION after Gemini 4 Hydrogen Plan Execution · Forum: [Appeal / Bug] False Positive 403 TOS_VIOLATION after Gemini 4 Hydrogen (Official VS Code Extension Only)

Google’s Hydrogen guide lists request fields that now fail validation: thinking_budget and thinking_level="MINIMAL" return 400 (use LOW, MEDIUM, HIGH or XHIGH); frequency_penalty, presence_penalty and candidate_count greater than 1 return 400; out-of-range temperature, top_p or top_k return 400 (in-range values are ignored, so remove them); max_output_tokens above 1,048,576 returns 400; and history payloads cannot end with a model turn or contain prefilled model turns.

Source: Google Cloud documentation, Last updated 12 November 2026 · Source: Google AI Studio / Gemini API docs, Retrieved 13 November 2026

Hydrogen binds thought signatures to the conversation prefix. Editing or compacting earlier turns on a same-model replay returns 400 by default; set prefix_mismatch_behavior="DROP" in ThinkingConfig (generation_config on the Interactions API) if your harness compacts history. In parallel function calls only the first functionCall part carries thought_signature — omitting it there, copying it onto sibling calls, or interleaving calls and responses returns 400. Pass back every signature from earlier turns unchanged; stripping them can cause MALFORMED_FUNCTION_CALL. Also declare tools natively instead of asking for raw-text JSON tool calls.

Source: Google Cloud documentation, Last updated 12 November 2026

A single synchronous response is capped at 65,536 tokens, and max_output_tokens counts thinking plus answer. For longer output set max_output_tokens above 65,536 (up to 1,048,576 minus input) and use background=true on the Interactions API or SDK automatic continuation (google-genai ≥ 2.34.0 for Interactions, ≥ 2.30.1 for GenerateContent). Status incomplete means the total budget ran out — raise max_output_tokens before continuing, or the continuation returns 400. On Google Cloud, long decode works only on the v1beta1 API. Old hard-coded caps such as 8,192 should be removed or raised.

Source: Google AI Studio / Gemini API docs, Retrieved 13 November 2026 · Source: Google Cloud documentation, Last updated 12 November 2026

Hydrogen does not support explicit context caching — remove cached_content from requests. Implicit caching runs automatically on shared prompt prefixes of at least 4,096 tokens and bills cache hits at $0.075 per 1M (95% off), so put static content first. Developers on the forum are reporting cases where a small change late in the prompt shows no reported prefix reuse; Google has not answered those threads yet, so check cached_content_token_count in your own usage data rather than assuming a hit rate.

Source: Google AI Studio / Gemini API docs, Retrieved 13 November 2026 · Source: Google Cloud documentation, Last updated 12 November 2026 · Forum: Gemini-4-hydrogen: exact repeats cache ~93%, but a one-character user-message change loses all reported shared-prefix reuse

Limits are per project (not per API key), measured as requests per minute, input tokens per minute and requests per day; RPD resets at midnight Pacific time. Google does not print a fixed Hydrogen RPM table — your active limits are shown in AI Studio and depend on usage tier. Paid tiers also have spend-based limits per rolling 10 minutes ($10 on Tier 1, $200 on Tiers 2–3) that return 429. The Batch API allows 3,000,000 enqueued tokens for Hydrogen. Retry 429 and 5xx with exponential backoff and jitter; do not retry 400, 402 or 403.

Source: Google AI Studio / Gemini API docs, Retrieved 14 November 2026 · Source: Google AI Studio / Gemini API docs, Retrieved 14 November 2026

Google’s troubleshooting guide covers four causes: verification needed (verify with a mobile device in AI Studio, or link billing); free tier unavailable for the project (a new key in the same project will not help — link a billing account); payment needed on the linked billing account; or an API key reported as leaked (create a new key in AI Studio). A 200 from models.list does not prove inference is enabled for the project.

Source: Google AI Studio / Gemini API docs, Retrieved 14 November 2026

Reported on the forum on 13 November 2026: the same Flex PayGo request works for Hydrogen on the Gemini Developer API and for gemini-3.8-flash on Vertex, but returns this 400 for Hydrogen on Vertex GenerateContent at location global; the Interactions API on Vertex accepted service_tier="flex". No Google reply yet. Until then, use standard tier or the Interactions API on Vertex, or Flex on the Gemini Developer API.

Forum: 400 “Flex API is not supported for model: gemini-4-hydrogen”

Single-user forum reports from 13 November 2026, not yet confirmed by Google: in AI Studio the Stop button and activity indicator can stay on after a response finishes and the prompt may not be saved; one user saw raw self-correction text appear inside a 114K-token multi-part output at xhigh; another hit repeated filter blocks in Antigravity on read-only USB hardware-recovery work. Google’s troubleshooting guide says BlockedReason.OTHER may mean the request is unsupported or violates terms, and recommends filing a bug from the guide.

Forum: Gemini Hydrogen – UI Issues, Instruction Following and Long-Context Reliability · Forum: Gemini-4-hydrogen Raw internal self-correction · Forum: False-positive safety filter blocks on Gemini 4 Argon & Hydrogen during read-only WinUSB hardware recovery