Kavukcuoglu said its biggest strengths are cybersecurity and long-horizon tasks, the kind of jobs that involve many steps or run for extended periods. The model has already taken the top spot in the Artificial Analysis Intelligence Index, ahead of Claude Opus 5.5.
Google’s own comparison table shows Carbon leading on 15 of 17 benchmarks against rivals, including GPT-6 Astra and Opus 5.5. Every figure in that table comes from Google’s own testing. GPT-6 Astra still leads on FrontierSWE 2, a software engineering test, and Opus 5.5 keeps its lead on OSWorld-2.1, which measures how well a model can operate a computer.
One change in method stands out. Google now reports Terminal-Bench 4.0 scores from Artificial Analysis, an independent testing firm, instead of running other models’ numbers itself. The benchmark tests whether an AI can finish technical tasks by typing commands, and Carbon improved by eight points over Argon on it.
Defenders Get It First, Everyone Else Waits
Google is rolling Carbon out to a small group of cybersecurity defenders first. There is no public release date for developers, businesses, or consumers. That timing lets Google claim leaderboard dominance and put early benchmarks in front of the market before the model is widely available. The recent PHP2Shell incident is a reminder of how much disruption a new AI model can cause across the internet.
Carbon uses the same safety system as Argon, and Google says its alignment, meaning how reliably it behaves as intended, has improved. The company’s internal testing points to these results:
- Carbon never cheated on Terminal-Bench 4.0, though it tried to only 0.4 percent of the time
- High-severity misalignment alerts fell by nearly 80 percent during internal coding work
- It blocked every attack on Gray Swan’s prompt injection benchmark
Priced Below Opus and Astra Before It Ships
When Carbon reaches the API, Google says it will cost $2 per million input tokens and $10 per million output tokens, with cached input tokens 95 percent off the input price. That is significantly cheaper than Claude Opus 5.5 and GPT-6 Astra. Google has not said when it will be publicly available.
Hashlytics Take
Right now Carbon is a set of claims. The benchmark lead is self-reported, the safety numbers come from internal testing, and nobody outside a small defender group can try the model. The Terminal-Bench 4.0 scores from Artificial Analysis are the one result with outside verification, and that is where the eight-point gain over Argon deserves the most weight. The pricing is the part that could hurt rivals, but a price for a model you cannot buy is a positioning move. Until it reaches the API and independent testers get hands on it, the 15-of-17 headline tells us what Google wants the comparison to look like, not how Carbon performs in real work.
Follow Hashlytics on Bluesky, Facebook, LinkedIn , Telegram and X to Get Instant Updates
