If you have been following Google’s AI releases this month, you have probably seen two names appearing in the same headlines: Gemini 4 Argon and Gemini 4 Carbon. Search results make them sound like two products you can compare and choose between. They are not — at least not yet. One is a public model you can pay for today. The other is an internal Google experiment that may never ship at all. Here is what each name actually refers to, what is confirmed and what is still just a leak, and what it means for you if you are deciding what to build on.
Argon vs Carbon: the short version
Argon is the product; Carbon is the lab experiment. Argon was announced on 30 September 2026 and went generally available on 13 October 2026 in the Gemini API, the Gemini app (paid tiers), and GitHub Copilot. Carbon, by contrast, surfaced only through a Business Insider report dated 9 October 2026: it is an internal checkpoint Google engineers are testing on the company’s own coding platform, with no public endpoint, no pricing, and no announced release date. If you want to use one today, the answer is Argon — there is no other choice on the table.
What Gemini 4 Argon actually is
Argon is Google’s current flagship Gemini 4 model, and it is fully public. It was first handed to vetted cybersecurity partners under Google’s Fairwind program at the end of September, then opened to paid API developers and Google AI Pro and Ultra subscribers in October. Google’s own product lead, Tulsee Doshi, described it as a “powerhouse” and the first entry in a broader family of upcoming models (reported by Cryptoverse News).
What you can verify about Argon right now:
- Availability: generally available since 13 October 2026 — in the paid Gemini API, the Gemini app for Pro and Ultra subscribers, and GitHub Copilot for Pro+, Max, Business and Enterprise users (per GitHub’s changelog, via Postcutoff’s 13 October roundup).
- Pricing: $2 per million input tokens and $10 per million output tokens through 30 November 2026, with cached input at 95% off; rates double from 1 December (per a widely shared breakdown by Sandeep Raiza — this is an individual summary, not a Google announcement, so treat the exact figures as reported until Google’s pricing page confirms them).
- Context: a 1,048,576-token context window with up to 1 million tokens of output (reported).
- Benchmarks: Google claims leads on DeepSWE v1.1 (77.9%) and the Vals Index (68.9%), but independent numbers are more mixed — the Artificial Analysis Intelligence Index scores it around 52.6, behind Claude Opus 5.5 at 57.6, and Endor Labs’ independent security benchmark found Argon on Cursor solving 29% of security tasks at a $765 run cost versus Opus 5.5 on Claude Code at 33.5% for $116 (Endor Labs, 21 October 2026).
- Agents: Google’s own researchers had four Argon agents beat Minecraft’s Ender Dragon with no human intervention — 27 hours, 2,783 programs written, 1.35 billion input tokens (Engadget).
For the full backstory on Argon’s launch, see our Gemini 4 Argon launch explainer.
What Carbon is — and what it is not
Carbon is not a product. It is a codename from a leak. The entire public record of Carbon is a single Business Insider report (byline Hugh Langley, 9 October 2026), based on internal documents, screenshots and employee chats. According to that reporting: Google has been testing three Gemini 4 checkpoints internally — Argon, Barium and Carbon — on an internal coding platform called Jetski (Google’s internal name associated with the Antigravity coding environment, per TestingCatalog). One employee reportedly said Carbon “feels like Opus 5.5” for coding; another added “Carbon is really good!”.
Two caveats matter here. First, those are employee impressions from internal chats, not benchmarks — there is no public benchmark, no model card, and no endpoint for Carbon. Second, a DeepMind employee has publicly hinted that the fast iteration pace comes from recursive self-improvement — AI models helping build their successors — which suggests Google is running a rolling checkpoint cadence rather than preparing a big-bang “Carbon launch”. For the full deep dive, read our Gemini 4 Carbon explainer.
Argon vs Carbon: the checkpoint lineage table
This is where the naming gets confusing, because internal codenames do not map cleanly to public names. The same Business Insider reporting that surfaced Carbon also revealed that the model the public knows as “Argon” was internally the Barium-B checkpoint. Here is every claim about the Gemini 4 checkpoint line, tagged by how solid it is:
| Internal codename | Claim | Status |
|---|---|---|
| Barium-B | The checkpoint selected for the public Argon release | Reported (Business Insider, 9 Oct 2026) |
| Carbon | Separate, possibly more capable checkpoint tested internally on Jetski | Reported (Business Insider, 9 Oct 2026) |
| Carbon | Feels like Claude Opus 5.5 for coding | Unverified — one employee impression, no benchmarks |
| Carbon | Will ship publicly, or ship as an Argon update | Unknown — no release date, no announcement |
| Argon | First of a broader family of upcoming Gemini 4 models | Confirmed — Tulsee Doshi, Google DeepMind |
The honest version: Google is running multiple checkpoints in parallel and releasing whichever one crosses its internal quality bar. “Carbon” may be the next checkpoint to cross that line, or it may be an experimental branch that never leaves the lab.
Which one will you actually get?
If you are buying or building on a Gemini 4 model today, the answer is Argon — it is the only option. It is generally available in the API, the Gemini app’s paid tiers, and GitHub Copilot; Carbon is inaccessible to everyone outside Google and might never be accessible. Practical upshot: evaluate Argon on its actual numbers — the $2/$10 launch pricing through 30 November, the mixed independent benchmarks, the ~1M-token context — not on what a future checkpoint might do. If you are choosing between Argon and rivals like Claude Opus 5.5, judge the models you can run, not the ones you can only read about.
What developers should do while the checkpoint treadmill runs
Google has moved to a rolling release model: the “launch” is just the moment a checkpoint crosses the external line, and the next one is already in internal testing. As the Models desk at Capernaums put it in its 9 October brief, the practical consequence is that pinning a specific model version is becoming a liability — your eval suite should run against checkpoints, not against major version names, because the ground shifts underneath you every quarter or faster (Capernaums Models brief). For buyers, that means negotiating platform fees around continuous updates rather than version locks.
Frequently asked questions
Is Gemini 4 Carbon a model I can try?
No. Carbon is an internal Google checkpoint described in a single October 2026 leak. There is no public endpoint, no API access, no app switch, and no confirmed plan to release it.
Is Carbon better than Argon?
Unknown. The only signal is one Google employee’s impression that Carbon “feels like” Claude Opus 5.5 for coding — an internal-chat remark, not a benchmark. No public performance numbers exist for Carbon.
When will Carbon be released?
There is no announced release date, and Google has declined to comment on the reporting. It may ship as a future Gemini 4 update, it may ship under a different name, or it may never leave internal testing.
Can I use Gemini 4 Argon for free?
Not currently. Argon is available through the paid Gemini API, GitHub Copilot’s paid tiers, and Google AI Pro and Ultra subscriptions. Free accounts and the free API tier are not included (per widely shared reports — Google’s own pricing page is the final word).
Why don’t Google’s internal names match the public ones?
Google tests multiple checkpoints in parallel under codenames and releases whichever crosses its bar. Per the October 2026 reporting, the public “Argon” was internally the Barium-B checkpoint — so the codename “Argon” as used inside Google and the product name “Gemini 4 Argon” are not the same thing.
Sources and methodology
Claims in this article rest on: Google’s own statements (Tulsee Doshi, GitHub Copilot changelog, Argon system-card reporting); the single Business Insider report on Carbon (9 October 2026, byline Hugh Langley, via TestingCatalog, kie.ai and Orcarouter’s summaries); independent benchmarks (Endor Labs 21 October 2026; Artificial Analysis Intelligence Index); and trade reporting (Engadget’s Minecraft agents piece, Capernaums Models brief). Where a claim comes from employee impressions or individual summaries rather than official sources, it is labelled as reported or unverified above. Verified 21 October 2026. Not hands-on tested — Carbon is inaccessible and this is a comparison of public records, stated plainly.
Govind Dheda covers AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.