Next Gemini 4 Models: Confirmed Dates vs. the Oxygen, Krypton and Helix Rumors
What Google shipped since September, which Gemini dates are officially scheduled, and how strong the evidence is for Oxygen, Krypton, Helix, Cobalt and a Gemini 4 Nano Banana.
Last updated: 2026-11-19. This page is a tracker and will be revised as Google publishes changes.
Google has not published a calendar of upcoming Gemini 4 models. The only forward-looking statements on record are a general one (“more models coming soon, including more efficient versions”) from Sundar Pichai on October 28, 2026, and a promise that Gemini 4 Neon for cyber reaches Fairwind Program members “in the coming days.” Everything else with a fixed date is a price change or a routing/shutdown change: Carbon and Argon double in price on December 1, Gemini 2.5 Pro and 2.5 Flash requests get rerouted on December 1, Omni 1.1 Flash reroutes on December 3, several TTS previews have an earliest shutdown of December 18, and Gemini 3.8 Flash doubles in price on January 1, 2027. Oxygen, Krypton, Helix, Cobalt and a Gemini 4-based Nano Banana are unconfirmed names. None has a Google model ID, price row or announcement.
Ofox publishes this article. It is documentation-based: we did not call any Gemini model for it. Official facts come from Google’s API docs, model cards and blog posts, checked on November 19, 2026. Rumors are attributed to the people who posted them and graded by how much evidence exists. The one thing we did run is the release-watch script near the end, which we executed against Google’s live documentation and the Ofox model list on November 19, 2026.
If you are choosing among the four models that already exist, read our Gemini 4 model selection guide and the Gemini 4 vs Claude vs GPT tier map instead. This page is about what comes next and how to know when it has really arrived.
What Google shipped from September 1 to November 18, 2026
The table counts model releases and access changes, not every feature. “Announced” and “generally available” are separate rows when Google separated them. Dates are the dates Google printed on the cited page.
| Date | Release | Where it became available | Source |
|---|---|---|---|
| Sep 2 | Gemini 3.8 Flash (gemini-3.8-flash) GA | Gemini API | Gemini API changelog |
| Sep 3 | Lyria 3.5 (lyria-3.5) GA | Gemini API | Changelog |
| Sep 15 | Gemini 3.8 Live and 3.8 Live Extended Thinking GA | Live API | Changelog |
| Sep 22 | Gemini 3.8 Flash TTS and Flash-Lite TTS GA | Gemini API | Changelog |
| Sep 30 | Gemini 4 Argon announced | Trusted cyber defenders via the Fairwind Program first | Google blog, Argon |
| Oct 6 | Nano Banana 2.1 (gemini-nano-banana-2.1) GA | Gemini API | Changelog |
| Oct 6 | EmbeddingGemma 2 (open model) | Open weights | Gemma releases |
| Oct 13 | Gemini 4 Argon (gemini-4-argon) GA | Gemini API; Pro and Ultra in the Gemini app | Changelog; Gemini app release notes |
| Oct 27 | Genie 4 announced | Research announcement | Pichai’s Q3 remarks, Oct 28: “yesterday we announced Genie 4” |
| Oct 28 | Gemini 4 Carbon announced | A subset of cyber defenders, then Fairwind partners | Google blog, Carbon |
| Nov 5 | Gemini Omni 1.2 Flash (gemini-omni-1.2-flash) GA | Gemini API | Changelog |
| Nov 10 | Gemini 4 Carbon (gemini-4-carbon) GA | Gemini API paid tier | Changelog |
| Nov 10 | Gemini Robotics 2.5 | Robotics partners only | DeepMind blog |
| Nov 11 | Gemini Decision 1 (gemini-decision-1) and the Decisions API | Gemini API | Changelog |
| Nov 11 | AlphaGenome 2 | Non-commercial research API; commercial research in private preview on Google Cloud | DeepMind blog |
| Nov 12 | Gemini 4 Hydrogen (gemini-4-hydrogen) GA | Gemini API, free and paid tiers; default model for the Antigravity agent | Changelog |
| Nov 18 | Gemini 4 Neon (gemini-4-neon) GA | Gemini API paid tier; Pro and Ultra in the Gemini app and Antigravity | Changelog; Neon announcement |
| Nov 18 | Genie 4 opened in Project Genie | Eligible Google AI Ultra subscribers; Pro members can join a waitlist | Google blog, Genie 4 |
Three practical notes on this list:
- Four Gemini 4 models exist today: Argon, Carbon, Hydrogen and Neon. Google’s deprecations page lists exactly those four IDs, each with “No shutdown date announced.”
- Not all of them are callable everywhere. On Ofox,
google/gemini-4-hydrogenandgoogle/gemini-4-argonwere listed when we queriedGET https://api.ofox.io/v1/modelson November 19, 2026. Carbon, Neon, Decision 1, Omni 1.2 Flash and Genie 4 were not. For those, use Google’s native API. Our guides cover Neon on Google’s API, Carbon setup, Decision 1 and the Decisions API and how to get Genie 4 access. - Genie 4 and Robotics 2.5 are not developer API models today. Genie 4 sits behind a consumer subscription and Robotics 2.5 behind a partner program. Counting them as “API releases” inflates the release pace.
For a head-to-head on the two newest text models, see Gemini 4 Neon vs Carbon. For video, see Omni 1.2 Flash vs Seedance 2.5.
What Google has officially scheduled
These items appear in Google’s own documentation or announcements, with a date or a stated window. They are the only forward-looking Gemini events you can plan around.
| When | Event | What changes for you | Source |
|---|---|---|---|
| ”In the coming days” after Nov 18 | Gemini 4 Neon for cyber rolls out to trusted defenders | Only Fairwind Program members. Google says to contact your Cloud account team to request Fairwind access | Neon announcement |
| Nov 30 (last day) | Carbon and Argon introductory pricing ends; Google’s table says “through November 30, 2026” without a time zone | Last day at $2/$10 per 1M input/output tokens | Pricing |
| Dec 1 | Carbon and Argon Standard prices go from $2/$10 to $4/$20 per 1M input/output tokens; cached input from $0.10 to $0.20 | Batch, Flex and Priority rows for both models also double | Pricing |
| Dec 1 | gemini-2.5-pro requests automatically route to gemini-3.8-flash, billed at 3.8 Flash rates | Breaking changes include function calling and thought_signature validation, thinking_level replacing thinking_budget, removed sampling parameters | Changelog, Nov 2 |
| Dec 1 | gemini-2.5-flash requests automatically route to gemini-3.5-flash-lite, billed at Flash-Lite rates | Same class of breaking changes; gemini-2.5-flash-lite is not affected | Deprecations |
| Dec 3 | gemini-omni-1.1-flash requests route to gemini-omni-1.2-flash at 1.2 rates | Re-check clip length, reference images and cost | Changelog, Nov 5 |
| Dec 18 (earliest) | Shutdown of gemini-3.1-flash-tts-preview, gemini-2.5-flash-preview-tts, gemini-2.5-pro-preview-tts | Replacement: gemini-3.8-flash-tts or gemini-3.8-flash-lite-tts | Deprecations |
| Jan 1, 2027 | Gemini 3.8 Flash Standard price goes from $0.75/$3.75 to $1.50/$7.50; cached input $0.075 to $0.15; Batch from $0.375/$1.875 to $0.75/$3.75; Flex and Priority rows also double | This also affects anything routed to 3.8 Flash, including gemini-2.5-pro after Dec 1 | Pricing |
| Mar 15, 2027 (earliest) | Shutdown of gemini-2.5-flash-image | Replacement: gemini-3.1-flash-lite-image | Deprecations |
| Mar 31, 2027 (earliest) | Shutdown of gemini-2.5-flash-native-audio-preview-12-2025 | Replacement: gemini-3.8-live | Deprecations |
| May 7, 2027 (earliest) | Shutdown of gemini-3.1-flash-lite | Replacement: gemini-3.5-flash-lite | Deprecations |
| ”Soon,” no date | AlphaGenome 2 Skills in Google Antigravity | Scientific agent tooling, not a new Gemini model | AlphaGenome 2 blog |
Two details people get wrong about this table:
- “Shutdown date” means earliest possible date. Google’s deprecations page states that the listed dates “indicate the earliest possible dates” and that the exact date will be communicated with advance notice. Plan for the listed date; do not promise a customer that the old model will stop on that exact day.
- Routing is not the same as shutdown. On December 1 a request to
gemini-2.5-prowill still return a response, but from a different model, at a different price, with different API rules. That is harder to notice than an error. If you still send 2.5 Pro traffic, see our note on 2.5 Pro access in AI Studio and fix any missing thought signature errors before the switch.
The December 1 price change is big enough to have its own article: Gemini 4 Carbon and Argon December price change works through the arithmetic, and Gemini 3.8 Flash API pricing covers the 3.8 Flash baseline. Neon’s own rates did not change; see Gemini 4 Neon pricing.
What Google said about “more models”
The source most often cited for “Google will ship more Gemini 4 models soon” is Sundar Pichai’s Q3 2026 earnings remarks, published on October 28, 2026. The relevant words, checked on November 19, 2026, are:
- “Argon is just the start of the Gemini 4 series.”
- “We have more models coming soon, including more efficient versions to help developers optimize for speed and cost and to deploy across our billions of users.”
- “I would expect the next 12 months to see continued acceleration because of RSI.” The reason he gives is RSI; the sentence names no models and no release calendar.
That is the whole official forward guidance. The remarks give no count of models, no names and no month. Hydrogen (November 12) and Neon (November 18) shipped after these remarks, so part of “more models coming soon” has arguably already happened. Whether more efficient versions beyond Hydrogen are planned is not stated.
You may have seen claims that Google will release “a dozen models next month.” We searched Google’s blogs, the Gemini API changelog, the deprecations page and the earnings remarks and found no such statement. We could not trace where the figure started; release trackers that put rumored codenames, research models, consumer features and other labs’ models in one list make such counts easy to inflate. Treat any count of upcoming Google models as a third-party tally, not a Google plan.
The unconfirmed names, graded by evidence
We use three grades:
- Strong: a real Google artifact you could check yourself, such as a model ID returned by an API, a Cloud SKU with billing rows, or a docs page, but without an announcement.
- Weak: a screenshot or code reference from a Google product, reported by a named source, that nobody independent has reproduced.
- Very weak: model outputs or claims from social accounts, with no Google artifact, or with a retraction.
As of November 19, 2026, none of the names below reaches “strong.” None appears in Google’s model list, pricing page, changelog or deprecation tables. We checked those four pages for each name.
“First seen” is the earliest dated post we could find, in UTC. We could not open X posts directly (X blocks our fetcher without a login), so the wording of each X post below comes from search-engine previews and from the trackers we name, which we did read. Where a fact reached us only through a tracker or a news aggregator, the table says so.
| Name | First seen | Original reporter | Where | Claim | Evidence | Would confirm it |
|---|---|---|---|---|---|---|
| Oxygen | Name: Nov 15, 17:05. Arena outputs: from Nov 14 | @pankajkumar_dev; relayed on Nov 17 by AI信息Gap via Sina Tech | Anonymous Arena outputs served under the gemini-4-hydrogen label | Next Flash-class model, better design taste than Hydrogen at xhigh | Very weak | An oxygen model ID in AI Studio code or the Cloud SKU catalog |
| Krypton | Nov 9, 06:12 | @chetaslua; then @lyraxana and @Mr_Salio (timeline via ArtificialWatch) | Outputs posted on X with no stated source; an anonymous private label in Design Arena | A “first checkpoint” that takes about twice as long as Carbon on the same prompt and writes more code | Very weak (Arena claim retracted Nov 14) | Any Google artifact naming Krypton; a model card |
| Helix | Nov 10, 19:45 (commit dated Nov 10, 04:20) | @LuminaBench (timeline via ArtificialWatch) | A commit in a personal Antigravity Docker project whose owner says he works at Google; repository deleted by Nov 11 | Gemini 4 Flash-Lite tier; an unsourced price of $0.30/$2.50 per 1M tokens circulated from Nov 15 | Weak to very weak | A gemini-4-helix ID, docs page or SKU rows |
| Cobalt | Nov 18, 18:33 | @testingcatalog, analyzed by OrcaRouter | AI Studio front-end JavaScript listing models/cobalt next to models/neon | A named Gemini 4 model; purpose unknown | Weak | SKU billing rows or a model page |
| Gemini 4-based Nano Banana | Nov 18 (earliest dated source we found) | TestingCatalog | One sentence in TestingCatalog’s Neon article; no artifact shown | An image model upgrade “in preparation for quite some time” | Very weak | A new image model ID in the changelog |
Oxygen
The Oxygen story starts with Arena outputs that appeared on November 14 under the gemini-4-hydrogen label. The earliest post we found using the name is from @pankajkumar_dev on November 15, describing “Gemini 4 Oxygen (Flash)” as available on Arena under the Hydrogen label. A November 17 article by AI信息Gap, syndicated on Sina Tech, says testers first guessed the checkpoint was Krypton, then started calling it Oxygen, and shows it beside Hydrogen at xhigh in one side-by-side build. AGTP and other X accounts then repeated that Oxygen had leaked as Google’s next fast model.
The problem is what happened on November 14 itself. According to ArtificialWatch’s log, @Mr_Salio, who had called that day’s Arena outputs Krypton, retracted and said they were Gemini 4 Hydrogen at a higher reasoning effort; @LuminaBench also said they looked like plain Hydrogen. We can’t tell whether the later Oxygen posts describe the same outputs. Arena outputs can’t be tied to a codename without a label from Google, and the Oxygen name has no trace in any Google artifact we could find. We grade it very weak. It may well exist; there is just nothing to verify yet.
Krypton
Krypton has the longest trail of the rumors and the most visible contradictions. On November 9, @chetaslua posted outputs he attributed to a “krypton-a version first checkpoint” and compared them with Carbon and Argon, without saying where he ran it. Other accounts amplified the posts, and @LuminaBench shared a screenshot of “Krypton” in a Design Arena providers list. ArtificialWatch checked Design Arena’s public registry and found krypton only as an anonymous private label with no lab attached.
On November 14 a further Krypton claim was retracted as Hydrogen. The same day, @pankajkumar_dev said Krypton would “mostly” launch next month because it needs more time; we saw that post only as quoted by Sumus, and it gives no source. That is the only Krypton date we found, and it is one account’s expectation. On November 16, AI信息Gap (via Sina Tech) reported that an Arena-monitoring bot had found a new gemini-4-carbon checkpoint and read it as Google testing Krypton under Carbon’s name. That reading is an inference, not an observation. Grade: very weak.
Helix
Helix is the most specific rumor, and its specificity is what weakens it. According to ArtificialWatch, a commit dated November 10 at 04:20 UTC in a personal Docker project for Google Antigravity, whose owner’s profile says he works at Google, pointed the Flash tier at Hydrogen and the Flash-Lite tier at gemini-4-helix. @LuminaBench flagged it at 19:45 UTC the same day, and the repository returned a 404 by 09:50 UTC on November 11. A personal project is not a Google product, so this is weaker than front-end code. Hydrogen went GA on November 12, which some read as validation.
From November 15, posts said Helix had been “spotted on Antigravity” with a price of $0.30 per 1M input tokens, $2.50 per 1M output and $0.03 cached. None of the posts we found shows where that table came from. Google’s pricing page lists exactly those three numbers for the existing gemini-3.5-flash-lite, and the same numbers as the text price of gemini-2.5-flash. A leaked price that matches current products could describe a planned successor at the same price, a relabelled screen or a fabricated table. It can’t tell those apart. Helix is also not a chemical element, which breaks the naming pattern of every Gemini 4 model so far. Grade: weak to very weak.
Cobalt
On November 18 at 18:33 UTC, TestingCatalog posted a screenshot of Google AI Studio’s minified JavaScript with a carousel config listing models/neon and models/cobalt. Neon went GA that day, which lends the screenshot some credibility. OrcaRouter’s write-up, published the same day, lists what is missing: no Cobalt ID in Google’s documentation, no changelog entry, and zero Cloud SKU billing rows. At the time of that write-up Neon also had no docs ID or changelog entry, but it already had 24 SKU rows. We grade Cobalt weak: it is a real product artifact, but it is a single screenshot, and a placeholder or a later-renamed model would leave the same trace.
A Nano Banana built on Gemini 4
TestingCatalog’s November 18 Neon article says a Nano Banana upgrade based on Gemini 4 has been “in preparation for quite some time.” No name, ID, date or price is given. The current image model is Nano Banana 2.1, GA since October 6, and gemini-3.1-flash-image is deprecated with no shutdown date. TestingCatalog has a good record on this family (it reported Neon’s SKU rows before GA), but this sentence comes with no screenshot, code or ID. Grade: very weak. If you build on Nano Banana today, nothing in Google’s docs suggests you need to wait.
Older or adjacent names
- Barium: Business Insider reported on October 9 that Google staff were testing checkpoints codenamed Barium and Carbon (as logged by ArtificialWatch). Carbon then shipped, so the report had a real hit, but Barium itself has produced nothing since. Very weak as a release signal.
- Xenon and Radon: these appeared as anonymous labels in Design Arena with no lab attached; ArtificialWatch found that the one model there with the ID
radonbelongs to another company. Not Google evidence. - Gemini 5, Gemini Omni Pro, a new Gemini Pro: no Google statement or ID. Our own May 2026 article on Gemini 3.5 Pro’s expected release relied on a stated “next month,” and Google’s current model list contains no Gemini 3.5 Pro ID. Even an official-sounding timeline can change; an unofficial one has less to stand on.
Why “Gemini 4 Krypton” doesn’t mean “bigger than Argon”
Gemini 4 models are named after chemical elements: Hydrogen (atomic number 1), Carbon (6), Neon (10) and Argon (18). Google leaned into it in the Neon launch post, which describes a Neon-built 3D lab bench that pairs Bunsen flames with Carbon and Hydrogen and gas discharge tubes with Argon and Neon.
The atomic number is not a capability or size ranking. The release order and Google’s own positioning show this:
| Model | Atomic number | Gemini API GA | Google’s positioning |
|---|---|---|---|
| Hydrogen | 1 | Nov 12 | Speed and efficiency for everyday coding and agents; free tier |
| Carbon | 6 | Nov 10 | Most capable reasoning model, for the hardest long-horizon work |
| Neon | 10 | Nov 18 | Balances depth and efficiency for complex, high-frequency work |
| Argon | 18 | Oct 13 | The first Gemini 4 model |
Argon has the highest number and was released first. Carbon has a low number and sits at the top of the family. Neon, numbered between them, was positioned between Carbon and Hydrogen. So the rumored names tell you nothing about tier: Oxygen (8), Cobalt (27), Krypton (36) and Barium (56) could be any size. Krypton, Xenon and Radon are noble gases like Neon and Argon, which is a curiosity, not a prediction. The tier claims attached to these names (Oxygen as Flash, Krypton as larger) come from the people posting outputs, not from the name.
How earlier Gemini 4 leaks turned into releases
The useful question is not “is this rumor true” but “what did the rumor stage look like for models that did ship.” Two cases:
- Carbon: Business Insider reported the codename on October 9 (we saw that report only as logged by ArtificialWatch). Google announced Carbon on October 28 with limited access, and the Gemini API GA followed on November 10. A reported codename came first, then an announcement with gated access, then a changelog entry with a model ID and price.
- Neon: according to TestingCatalog, Neon appeared in a Gemini web app upsell on November 16, in the Google Cloud SKU catalog on November 18, and as a
gemini-pro-orchestrator-neonvariant in Antigravity CLI 1.5.11, before the GA entry the same day. Billing rows in the SKU catalog were the last signal before release.
Neither case had a published date until the day of release. Krypton is the counter-example: plenty of outputs, then a retraction. The pattern suggests a practical ladder, from weakest to strongest:
- Model outputs on Arena or social media attributed to a name.
- A name in product front-end code (AI Studio, the Gemini app, Antigravity builds).
- SKU billing rows in the Google Cloud catalog.
- A Google announcement, possibly with gated access (Fairwind, Ultra-only, partners).
- A Gemini API changelog entry with a model ID, a pricing row and a deprecations-table row.
Only step 5 means you can call the model from code. As of November 19, Cobalt is at step 2, Helix is somewhere between 1 and 2, and Oxygen and Krypton are at step 1.
How to watch for real releases
Watch the sources that sit at step 5, plus your provider’s model list. All of these are public except the last two, which need a key.
| Source | What it tells you | URL |
|---|---|---|
| Gemini API changelog | New model IDs, GA dates, deprecation and routing notices | https://aistudio.google.com/docs/changelog.md |
| Deprecations page | Every model with release date and earliest shutdown date | https://aistudio.google.com/docs/deprecations.md |
| Pricing page | Price rows and dated price changes | https://aistudio.google.com/docs/pricing.md |
| Model cards index | Dated cards for each Gemini model | https://deepmind.google/models/model-cards/ |
| Gemini API models list | Which model IDs your key can actually call | GET https://generativelanguage.googleapis.com/v1beta/models with an x-goog-api-key header |
| Ofox model list | Which google/* IDs are routable through Ofox | GET https://api.ofox.io/v1/models with your Ofox key |
The .md versions of Google’s docs pages are Markdown with a few template tags (such as {% id="11-18-2026" %} after each changelog heading), which makes them easy to diff. We did not have a Google API key for this article, so the Gemini models endpoint row is documentation-based; we did not call it.
A release-watch script you can run on a schedule
The script below stores a snapshot and prints only what changed: a new top changelog entry, new model IDs mentioned in the changelog, new or changed deprecation rows, and new google/* IDs on Ofox if OFOX_API_KEY is set. It exits with code 1 when something changed, so a cron job or CI step can alert on it. It uses only the Python standard library.
"""Watch Google's Gemini API changelog, deprecations page and an OpenAI-compatible
/models list for new model IDs. Stores state in a JSON file and prints only changes."""
import json, os, re, sys, urllib.request
STATE = os.environ.get("WATCH_STATE", "gemini_watch_state.json")
DOCS = {
"changelog": "https://aistudio.google.com/docs/changelog.md",
"deprecations": "https://aistudio.google.com/docs/deprecations.md",
}
ID_RE = re.compile(r"`((?:gemini|lyria|veo|imagen)-[a-z0-9.\-]+)`")
def get(url, headers=None):
req = urllib.request.Request(url, headers={"User-Agent": "release-watch/1.0", **(headers or {})})
with urllib.request.urlopen(req, timeout=60) as r:
return r.read().decode("utf-8")
def snapshot():
snap = {}
log = get(DOCS["changelog"])
snap["latest_changelog_date"] = re.search(r"^## (.+?)\s+\{%", log, re.M).group(1)
snap["changelog_ids"] = sorted(set(ID_RE.findall(log)))
dep = get(DOCS["deprecations"])
snap["deprecation_rows"] = sorted(set(re.findall(
r"\* `([a-z0-9.\-]+)`\n \* ([^\n]+)\n \* ([^\n]+)", dep)))
key = os.environ.get("OFOX_API_KEY")
if key: # optional: what your router exposes today
data = json.loads(get("https://api.ofox.io/v1/models", {"Authorization": f"Bearer {key}"}))
snap["router_google_ids"] = sorted(m["id"] for m in data["data"] if m["id"].startswith("google/"))
return snap
def diff(old, new):
out = []
if old and old.get("latest_changelog_date") != new["latest_changelog_date"]:
out.append(f"changelog has a new top entry: {new['latest_changelog_date']}")
for k in ("changelog_ids", "router_google_ids"):
added = sorted(set(map(str, new.get(k, []))) - set(map(str, old.get(k, []))))
if added and old:
out.append(f"new in {k}: {', '.join(added)}")
added_rows = set(map(tuple, new["deprecation_rows"])) - set(map(tuple, old.get("deprecation_rows", [])))
if added_rows and old:
out.append("deprecation table changed: " + "; ".join(" | ".join(r) for r in sorted(added_rows)))
return out
if __name__ == "__main__":
old = json.load(open(STATE)) if os.path.exists(STATE) else {}
new = snapshot()
changes = diff(old, new)
json.dump(new, open(STATE, "w"), indent=1)
if not old:
print(f"baseline saved: top changelog entry {new['latest_changelog_date']}, "
f"{len(new['changelog_ids'])} model IDs in changelog, "
f"{len(new['deprecation_rows'])} deprecation rows, "
f"{len(new.get('router_google_ids', []))} google/* IDs on router")
print("\n".join(changes) if changes else "no changes")
sys.exit(1 if changes else 0)
We ran it twice on November 19, 2026 at 03:25 UTC with OFOX_API_KEY set. Actual output:
baseline saved: top changelog entry November 18, 2026, 116 model IDs in changelog, 88 deprecation rows, 22 google/* IDs on router
no changes
no changes
The first run saves the baseline; the second run, seconds later, correctly reports no changes. An independent reviewer re-ran the same file at 03:33 UTC that day from an empty state file and got the same three lines. Your counts will differ as Google edits the pages. That run tells you the parser works against today’s page format. It does not prove it will catch every future release: Google can change the Markdown layout, and the regex would then need updating. If the script crashes on a None match, that is the signal to look at the page by hand. The ID regex also picks up any backticked model string in the changelog, including old ones, so the first diff after a new entry can list more than one ID.
Run it once a day. Combined with the five-step ladder above, a changelog hit is the only result that should change your code.
What to do in your code when a new Gemini model lands
When a new Gemini 4 model reaches the changelog, a few things are worth doing before switching traffic. None of them requires the rumor stage.
Keep the model ID in configuration, pinned to a stable ID. Google’s model version docs say stable IDs “usually don’t change,” while a -latest alias is hot-swapped with each release after a two-week email notice. For production, pin gemini-4-neon or gemini-4-hydrogen and change it deliberately.
Read the new model’s defaults, not just its name. Gemini 4 models do not share thinking defaults: Carbon and Argon default to medium, Hydrogen and Neon to high and add xhigh. A new model can change your latency and cost even with the same prompt. Our thinking levels guide shows how to set the level explicitly.
Check the price row and the free tier separately. Of the four Gemini 4 models, only Hydrogen has a free tier on the Gemini API (details here). Carbon and Argon carry a dated price change. A new model’s launch price may also be introductory; look for “through [date]” wording in the pricing table.
Run your own eval set before moving traffic. Google’s launch tables mix harnesses and settings. Neon’s own announcement reports 12.8% on SWE-sweep with the mini-SWE harness and 39.0% in Google Antigravity. Compare models on your tasks at fixed thinking levels. Our Argon to Neon migration guide walks through one such switch.
Plan for routing, not only for shutdowns. As December 1 shows, Google can keep an old ID working while pointing it at another model. Keep the deprecations page in your watch list, record which model ID each request used, and if the response carries a model or version field, log it and alert when it changes.
Make failure handling model-agnostic. New models arrive with their own rate limits and early errors. Retry on 429 and 5xx with backoff, and keep a fallback model configured. The Gemini 4 API errors guide and the older Argon error reference list the common ones. If you route through Ofox, its error-handling guide and rate-limit documentation describe the retry codes and current default limits.
Check availability on your gateway before announcing support. A Google release doesn’t make a model available on every gateway on the same day. On Ofox, check GET /v1/models or the model page, for example Gemini 4 Hydrogen or Gemini 4 Argon, where current prices and supported protocols are shown. Ofox’s Gemini API documentation covers the native and OpenAI-compatible paths. Ofox also still lists google/gemini-2.5-pro and google/gemini-2.5-flash; we have not verified how those IDs behave on Ofox after Google’s December 1 routing change, so check the model page before relying on them.
Gemini 4 posts in this series
- Choosing a model: model selection, Neon vs Carbon, tier map vs Claude and GPT, hallucination rates
- Cross-vendor: Neon vs Claude Opus 5.5 vs GPT-6.2 Sol, Carbon vs Claude Fable 5.2 vs GPT-6.2 Astra, Hydrogen vs Claude Sonnet 5.5 vs GPT-6.1 Terra, Hydrogen vs Claude Haiku 5.5 vs GPT-6 Luna, Argon benchmarks
- Setup and pricing: Hydrogen API, Hydrogen in coding tools, Argon access, Argon pricing, Carbon task cost, Computer Use, 3.8 Flash to Hydrogen
- Other Google models: Omni 1.2 Flash API
Related Docs
Frequently Asked Questions
- When will Gemini 4 Oxygen be released?
- There is no release date. As of November 19, 2026, Google has not mentioned a model called Oxygen in any announcement, model card, pricing page, changelog or deprecation table. The name first appears in social posts on November 15 about anonymous Arena outputs served under the gemini-4-hydrogen label from November 14. The account that first called those November 14 outputs Krypton later said they were Gemini 4 Hydrogen at a higher reasoning effort.
- Is Gemini Krypton bigger than Carbon?
- Nobody outside Google can show that. The claim rests on side-by-side outputs posted from November 9 by an account that did not say where the model ran. On November 14 a later Krypton claim was retracted, and the same day another account said Krypton would "mostly" launch next month without giving a source. No Krypton model ID exists in Google docs, and an element number says nothing about model size.
- Will there be a Gemini 4 Flash-Lite?
- Google has not announced one. Sundar Pichai said on October 28, 2026 that more models are coming, including more efficient versions, without naming them. The Helix rumor describes a Flash-Lite tier, but its leaked price is identical to the existing Gemini 3.5 Flash-Lite price, which weakens it as independent evidence.


