Within 2 weeks of Gemini 4 Argon's launch, Google has launched a new frontier AI model known as Gemini 4 Carbon. This is a significant release as Gemini 4 Carbon is the new top model in the Artificial Analysis Intelligence Index V5. With the new release, developers and security teams have access to a model that drastically improves usability (in coding), and closely meets safety requirements.
New model offers 1 million token context and lower pricing than competitors
The new model takes text, image, video, speech input and gives text output. Gemini 4 Carbon has a 1 million token context window for processing long documents. Google claims that Gemini 4 Carbon is a big leap forward compared to the model it launched just 2 weeks ago-Argon.
Gemini 4 Carbon is now the number one model for cybersecurity benchmarks and long horizon tasks. It has a 88.2 percent score in SEC-bench Pro, a benchmark for vulnerability discovery. Furthermore, the new model has 54.8 percent on ExploitGym, a benchmark for exploitation, and 71 percent on CWE-bench, a security patch benchmark.
According to Chrome's team, the model generates high-quality security patches 10x better than other public models. For generating secure code, internal benchmarks indicate that it has a success rate >2x higher than Gemini 4 Argon. It is around 23 percent cheaper compared to Anthropic Claude Opus 5.5 in standard Artificial Analysis Intelligence Index tasks.
The input tokens cost $4 for 1 million tokens. The output tokens cost $20 for 1 million tokens. Input tokens with cache are $0.20 for 1 million tokens (which is 95% discount).
The model is not yet publicly available. The model is still being reviewed by the US government.



SEC-bench Pro at 88.2 percent is impressive if the methodology holds up. I will believe the security claims when independent labs verify the results.
The pricing math in the article is slightly misleading. Input is $4 per million tokens but cached input drops to $0.20. That 95 percent discount is the real story for developers processing logs.
Does the 1 million token context window support the full multimodal pipeline or just text? I need to know if I can feed it video streams for security auditing without chunking.
It accepts text, image, video, and speech input. The output remains text only. You can process long documents but the model does not generate video or audio back.
still waiting for the us government review to clear so we can actually use this stuff lol
Is the cached input price of $0.20 per million tokens available for all users or just enterprise contracts? I am looking for cost effective options for my small team.
Google has not published specific tier restrictions yet. The public fact sheet lists global pricing. You should check the official developer console for the exact access rules once it launches.
The 10x better security patch claim comes from the Chrome team internal benchmarks. That is not an independent metric. Anthropic will likely release comparable numbers soon.