Gemini 4 Carbon: benchmarks, pricing and when developers get it

Gemini 4 Carbon is Google's new flagship model, announced on October 28, 2026, two weeks after Gemini 4 Argon became available to everyone. It already ranks #1 on the Artificial Analysis index.

And you can't call it.

No public endpoint, no model ID, no date. There's a pretty benchmark table, a launch price and a promise of "as soon as possible". Meanwhile, the timeline has already decided Claude is dead. Again.

This post separates three things: what Google published about Gemini 4 Carbon, what third parties measured on their own (including where it loses), and how to set up a watcher that pings you the minute the model shows up in the Gemini API.

TL;DR

• What it is: Google's new top model, above Gemini 4 Argon. Focused on cybersecurity and long-running agentic tasks.
• Status as of 10/29/2026: announced on 10/28, with access only for a group of cyber defenders through the Fairwind Program. No public API.
• Price: US$ 2 / US$ 10 per million tokens (input/output) during the introductory period, cached input at 95% off. After that, US$ 4 / US$ 20.
• Benchmarks: #1 on the Artificial Analysis Intelligence Index v5.0 (59). But it's #2 on the Vals Index, behind Claude Sonnet 5.5, and loses to Claude Code with Opus 5.5 on the coding agent index.
• What to do: keep the endpoint watcher running and the model fallback ready in your code.

Context: from Argon to Carbon in 15 days

The timeline is short and fast.

On September 30 Google announced Gemini 4 Argon, the first model in the Gemini 4 family. Cyber defenders first. On October 13 it became available to everyone: Gemini API, AI Studio, Antigravity and Google Cloud. The full comparison with Opus 5.5, including cost per agentic session, is in the post Gemini 4 Argon vs Claude Opus 5.5.

On October 28 came Gemini 4 Carbon, signed by Koray Kavukcuoglu, chief AI architect at Google DeepMind. Google calls it "our strongest model with state-of-the-art cybersecurity and long horizon capabilities".

Long horizon here means tasks that run for a long time. According to the official post, internal teams saw Carbon run for days on engineering problems: root-causing subtle bugs across large codebases and low-level systems, pushing through dead ends without giving up.

And the rollout plan is the same as Argon's. First a subset of cyber defenders, then trusted partners through Fairwind, and only then developers, businesses and consumers. Google also says it's taking part in the U.S. government's voluntary pre-release testing process. If Argon's pattern repeats, that's about two weeks between announcement and open API. But that's an observed pattern, not a promise.

Benchmarks: what Google says and what third parties measured

This is where the confusion lives. There are Google's own numbers, numbers Artificial Analysis verified, and numbers Vals measured from scratch. Mixing the three is how the timeline ends up at "Carbon crushed everyone".

The jump over Argon (Google's numbers)

Source: Google's official post and the table reproduced by VentureBeat. One detail Google itself notes: the cyber evaluations ran without guardrails. That's the raw model, not what you'll get through the API.

Google also cites two internal results nobody outside can check: more than 2x the success rate on secure code generation compared to Argon, and patches "more than 10X better" than another lab's model, which it doesn't name.

Against the competition

VentureBeat counted: Carbon has the top score in 15 of the 17 rows of Google's chart. Now the rows that matter if you write code with agents:

• Artificial Analysis Intelligence Index v5.0: Carbon 59, Claude Opus 5.5 58, GPT-6 Astra 57, Argon 50. A lead, but by one point.
• FrontierSWE 2: Astra 65.5%, Opus 62.3%, Carbon 61.2%. Carbon in third.
• OSWorld offline (computer use): Opus 87.2%, Carbon 81.5%.
• AA Coding Agent Index: Claude Code with Opus 5.5 scores 62. Carbon via Antigravity CLI scores 59, tied with Claude Code with Sonnet 5.5.

And the third-party-only measurement: Vals ran Carbon on Vals Index v2.2 and it came in #2 of 29, at 70.36%, behind Claude Sonnet 5.5 (70.85%). On Vals' Terminal-Bench 4.0 it's #4, at 63.1%. On the Vals Finance Index and Legal Research Bench, it's #1.

Honest read: it's a genuine frontier model, and in security it looks like it's alone at the front. In coding agents, where most of us live, it ties with what already exists. Not the blowout the timeline is selling.

The number the release didn't highlight: hallucination went up

Google's post says Carbon cuts "unsupported claims" in half compared to Argon. That's an internal test.

The external test tells another part of the story. On Artificial Analysis' AA-Omniscience, which asks factual questions and measures when the model makes things up instead of admitting it doesn't know, Carbon gets 58% of questions right (Argon got 48%). But the hallucination rate went from 4% to 11%.

Both things fit together. Carbon answers more and abstains less. When it answers more, it guesses more. The overall score goes up because it gets more right, and the share of wrong guesses goes up with it.

To be fair: 11% is still far below Opus 5.5 and GPT-6 Astra, which sit around 50% on the same measurement, per VentureBeat. Carbon is still one of the models that makes things up the least. But Argon made things up even less.

Why does this matter for your product? If you have a RAG system or an agent that needs to say "not found" when the context doesn't have the answer, switching from Argon to Carbon can make exactly that behavior worse. Capability benchmarks don't catch it. Your eval does, if it has "answer isn't in the document" cases.

Pricing: same as Argon, for now

Source: official post and VentureBeat.

Example math: 10 million input tokens and 1 million output tokens cost US$ 30 at the introductory price and US$ 60 after. Same model, double the bill.

Two gotchas. First, Google didn't say when Carbon's introductory price ends. If you're budgeting for 2027, put US$ 4 / US$ 20 in the spreadsheet. Second, price per token isn't cost per task. A model that runs "for days" on a task burns tokens for days. The announcement doesn't include token usage, retries or execution time, so there's no way yet to know whether Carbon is cheaper per solved task.

Tracker: how to know when Gemini 4 Carbon lands in the API

The Gemini API ListModels endpoint is the fastest source there is. A model shows up there when it becomes available to your key, often before the announcement makes the rounds. Same method as the Gemini 3.5 Pro tracker, adjusted for the 4 family.

Step 1: see what exists today

curl -s "https://generativelanguage.googleapis.com/v1beta/models?key=${GEMINI_API_KEY}&pageSize=200" \
  | jq -r '.models[].name' \
  | grep -i 'gemini-4'

Today this should return models/gemini-4-argon and no Carbon. Google hasn't published the model ID yet. Following Argon's pattern, the guess is gemini-4-carbon, but it's a guess. That's why the filter is gemini-4 and not the exact name.

Step 2: a watcher that alerts on its own

#!/usr/bin/env bash
set -euo pipefail

SNAP="${HOME}/.cache/gemini-4-models.txt"
mkdir -p "$(dirname "$SNAP")"

atual=$(curl -sf "https://generativelanguage.googleapis.com/v1beta/models?key=${GEMINI_API_KEY}&pageSize=200" \
  | jq -r '.models[] | select(.supportedGenerationMethods[]? == "generateContent") | .name' \
  | grep -i 'gemini-4' \
  | sort -u || true)

if [ ! -f "$SNAP" ]; then
  printf '%s\n' "$atual" > "$SNAP"
  echo "initial snapshot: $(wc -l <<<"$atual") gemini-4 models"
  exit 0
fi

novos=$(comm -13 "$SNAP" <(printf '%s\n' "$atual") || true)

if [ -n "$novos" ]; then
  echo "new gemini-4 models in the API:"
  printf '%s\n' "$novos"
  printf '%s\n' "$atual" > "$SNAP"
fi

In cron, every 15 minutes:

*/15 * * * * /usr/local/bin/gemini4-watch.sh | mail -s "Gemini 4 in the API" you@example.com

Swap mail for a curl to a Slack or Discord webhook. Since the script only prints something when a new model shows up, you don't get spammed.

Step 3: confirm before celebrating

Showing up in the list doesn't mean it works for you.

• Minimal generateContent call. A restricted-access model can show up and return 403. With Carbon starting on an allowlist, that's quite likely.
• Price on the official page. Check whether the introductory price applies to your account and until when.
• Model card and limits. Vals lists 1M context and 1M max output, but Google hasn't documented that. Don't size anything on a number that isn't in the model card.

What to do while Carbon isn't out

Don't sit and wait. Keep model swaps one .env line away.

In Laravel with Prism:

// config/ai.php
return [
    'principal' => env('AI_MODEL', 'gemini-4-argon'),
    'fallback'  => env('AI_MODEL_FALLBACK', 'gemini-3.7-flash'),
];

namespace App\Services;

use Illuminate\Support\Facades\Log;
use Prism\Prism\Enums\Provider;
use Prism\Prism\Prism;

class GeradorDeResposta
{
    public function gerar(string $prompt): string
    {
        $modelos = [config('ai.principal'), config('ai.fallback')];

        foreach ($modelos as $modelo) {
            try {
                return Prism::text()
                    ->using(Provider::Gemini, $modelo)
                    ->withPrompt($prompt)
                    ->asText()
                    ->text;
            } catch (\Throwable $e) {
                Log::warning('model unavailable, falling back to the next one', [
                    'modelo' => $modelo,
                    'erro' => $e->getMessage(),
                ]);
            }
        }

        throw new \RuntimeException('no Gemini model responded');
    }
}

The day Carbon opens up, the change is AI_MODEL=gemini-4-carbon (or whatever ID Google publishes) plus an eval run. Not a refactor.

And run the eval for real. Pick 30 to 50 real cases from your domain and include cases where the right answer is "I don't know" or "it's not in the context". With Carbon's hallucination going up compared to Argon, that's the kind of regression you only see if you look for it.

Limitations and caveats

• No technical report. Google hasn't published a technical report, architecture, parameter count or a Carbon-specific safety report. What exists is Argon's.
• Most numbers are Google's. Only the AA Index placement, the Coding Agent Index, Omniscience and the Vals Index were measured by third parties. The rest is self-reported, and the cyber tests ran without guardrails.
• Conflicting specs. Vals lists 1M max output. Argon has a documented 65,536 default output. Until the model card ships, treat the output limit as unknown.
• AA v5.0 is a new suite. 60% of tasks are private and the suite swaps benchmarks compared to the previous version. Don't compare Carbon's 59 with scores from older versions of the index.
• This post is dated. Written on 10/29/2026. If you're reading later, run the curl from step 1 before trusting the access status.

Quick FAQ

Can I use Gemini 4 Carbon yet? Only if you're in the Fairwind Program's cyber defender group. For developers, businesses and consumers, Google says "coming soon", with no date. It's not in the public Gemini API or AI Studio.

What's Carbon's model ID in the API? Not published yet. Following Argon's pattern (gemini-4-argon), it's likely gemini-4-carbon, but don't hardcode it. Use an environment variable and the watcher filtering on gemini-4.

Carbon or Argon? Today, Argon, because it's what exists in the API. When Carbon opens up, it's mainly worth it for security (finding and fixing vulnerabilities) and long tasks. For RAG that needs to admit "I don't know", test first, because Argon hallucinates less on AA-Omniscience.

Should I switch from Claude to Carbon? For coding agents, the independent numbers don't justify it: Claude Code with Opus 5.5 is ahead on the Coding Agent Index, and Sonnet 5.5 is ahead on the Vals Index. For security audits, Carbon is the strongest candidate. Decide with your eval, not the vendor's chart.

Conclusion

Gemini 4 Carbon is real, it's strong and, in cybersecurity, it looks like it's alone at the front. It's also locked down, with no model ID, no technical report and a final price twice the announced one.

Being #1 on an index by one point isn't a blowout. And answering more without hallucinating more is math Google hasn't closed yet.

The method is the same as the Gemini 3.5 Pro tracker: facts have a vendor URL and a third-party measurement, rumors have a screenshot. Put the watcher in cron, the model in .env and "I don't know" cases in your eval. The day Carbon opens up, you change one variable while the timeline is still arguing about benchmarks.

This isn't hype. It's engineering.