What happened
On 2026-10-28 Google announced Gemini 4 Carbon, calling it "our strongest model with state-of-the-art cybersecurity and long horizon capabilities". It builds on Gemini 4 Argon, which Google's own post says launched two weeks earlier.
The same day, Artificial Analysis ranked it #1 on its Intelligence Index v5 with a score of 59 — the first time a Google model has topped that overall index. That is the headline everyone shared.
Here is the part the headline leaves out: you can't use it. Access starts with a subset of cyber defenders, then widens to trusted partners through Google's Fairwind Program. Developers, businesses and consumers get it "as soon as possible", and Google gives no date. There is no model ID in the post, and no mention of the Gemini API, AI Studio or Vertex AI. Artificial Analysis says the same thing in plainer words: it is rolled out to selected users only and is not publicly available.
So this is a benchmark launch, not a product launch. That changes how a developer should read it.
How it actually does on the things you'd build with
A top spot on an overall index does not tell you whether a model is good at your job. The coding numbers are more honest, and they are closer than the #1 suggests.
On Artificial Analysis's Coding Agent Index, Carbon paired with the Antigravity CLI scores 59, against 62 for Claude Code with Claude Opus 5.5 (max), which stays first. On WebDev Arena — the one that matters most for frontend work — Carbon gained 89 Elo to reach 1768, and that only puts it at #4. On Terminal-Bench 4.0 in Artificial Analysis's test it scores 64%.
Read together, that is a model that is excellent in general, close on coding, and not ahead on web work. If you are already building a frontend with Claude Code, there is nothing here that says switch on day one — even leaving aside that you couldn't switch today if you wanted to.
The price, and the cost per task
The introductory price is $2 per million input tokens and $10 per million output tokens — the same as Gemini 4 Argon. Cached input tokens get a 95% discount. A footnote says that after the introductory period the price rises to $4 / $20 per million, with no end date given for the discount.
Per-token price is the wrong number to compare, though, because a model that reasons for longer spends more tokens to finish the same task. Artificial Analysis measured the end-to-end cost: $7.76 per Intelligence Index task for Carbon at standard pricing, against $10.04 for Claude Opus 5.5 (max) and $6.25 for GPT-6 Astra (max). Carbon also burns fewer output tokens per task than Opus — 123k against 173k.
That is the genuinely interesting result for anyone running an agent on a budget: Carbon is cheaper per finished task than Opus, and GPT-6 Astra is cheaper still. The per-token sticker price barely moved from Argon; the cost that lands on your bill depends on how many tokens the model spends thinking, and that is a different number.
The cybersecurity angle is about reviewing code, not writing it
Google built this one for security, and the benchmarks are all defensive: 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploitation), 71.0% on CWE-bench (patching). It also reports a 0.0% attack success rate on Gray Swan's indirect prompt injection benchmark, against 0.7% for Argon, and claims a more-than-2x uplift in secure-code generation over Argon from its own internal tests.
If you ship agent-written code, the useful framing is not "a model that writes safer code" but "a model that is good at finding the bug the other model wrote". The prompt injection result matters specifically for agents that read web pages or tool output and can be steered by text they didn't expect to trust. But note the source: these are Google's own reported results, and the AA-Omniscience test puts Carbon's hallucination rate at 11% against 4% for Argon — higher, not lower, than the model it replaces.
What is not known yet
A fair amount, and it is worth being honest about it:
- When developers get it. No date, no model ID, no API mention.
- How the launch discount ends. Artificial Analysis says there is no confirmed end date for the introductory pricing.
- Speed and latency. No figures published.
- Whether the higher hallucination rate shows up in real coding work, or stays a benchmark artifact.
- Whether any of the numbers hold up. They come from Google and from a test run with selected access. Nobody outside the rollout can reproduce them yet — a point Nathan Lambert made on Bluesky when the scores first appeared.
My take
This is Amar. I've learned to split a model launch into two questions: is it good, and can I use it. Carbon answers the first loudly and ducks the second, and I think the gap between those two is the real story this week.
I'm not moving my frontend agent off Claude Code on the strength of a 59 — partly because Opus still edges it on the coding index I care about, and mostly because a model I can't call isn't a choice I can make. What I am filing away is the cost-per-task number. If a model finishes the same work for fewer tokens, that shows up on a real bill in a way a benchmark rank never does, and that is the thing I'll actually test the day Carbon opens to the API. Until there's a model ID and a price I can pay, Gemini 4 Carbon stays on the watchlist, not in the stack.
