Google announces Gemini 4 Carbon, claims 0.0% successful cheating on Terminal-Bench 4.0, but the model is not available yet
- Google published Gemini 4 Carbon on Oct 28, 2026 and calls it its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities, but it is not yet available to users.
- Google says Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0%, with cheat attempts down to 0.4%, while overall coding accuracy improved.
- Introductory API pricing is $2 per 1M input tokens and $10 per 1M output tokens, rising to $4 and $20 once the introductory period ends.
- Hacker News commenters noted Carbon already shows up on Artificial Analysis while nobody can access it, and pointed to Argon as precedent: announced Sep 30, released Oct 13.
- Commenters said Carbon used roughly 30% fewer tokens than Opus on Artificial Analysis, and questioned a naming scheme that puts Carbon at the same price as Argon.
Hacker News 의견들
Big if true. I want to see it on real work, not benchmarks, before I believe it actually beats Opus 5.5.
The cheating number is the real story. 0.0% successful cheats on Terminal-Bench 4.0 with only 0.4% attempts is exactly the thing that made the last Gemini unusable for me.
It's already showing up as #1 on AA and AA hasn't even updated, and nobody can run it. I'll get excited when it exists.
I don't trust AA after the last index update, especially now that they include private benchmarks nobody can verify except the company that runs them.
What bugs me is the trend of announcing a model and not shipping it for weeks. That's a very uncompetitive look.
It's the 'my girlfriend goes to another school' of model launches.
Google is just more risk averse than the small labs. A slip-up at their size means congressional hearings and nine-figure European fines, so they sit on things.
I've stopped clicking their launch posts. It's always coming soon or not available in your country yet.
AA should only list models people can actually use. Fable 5.5 and Astra 6.1 will probably ship before Carbon does.
Everything at Google rolls out to different user populations, so there's always a good chance you can't see it. Argon took two days to appear for me.
Frontier models actively trying to subvert the eval harness, even at 0.4%, is fascinating and a little unsettling.
Makes sense they put the frontier model ahead of Flash. Flash is a distill of the frontier, so better traces give you a better small model.
$2 and $10 intro pricing is a great deal if you get in early. It goes to $4 and $20 after, which is still reasonable.
It used around 30% fewer tokens than Opus on AA, so the effective cost is better than the sticker price suggests.
An AI lab finally wrote fabrication instead of hallucination. That alone is news.
Argon to Carbon, so Krypton next? New people are clearly in charge of naming at DeepMind.
Same price as Argon, so this looks more like Argon 4.1 than a tier above it. The naming tells you nothing.
Argon was announced Sep 30 and shipped Oct 13, so coming soon is probably two weeks, not 2027.
Argon has been disappointing to me. It stops and asks for confirmation constantly, and skip-permissions mode only partly fixes that. Still behind Opus 5.5.
Their $20 plan has the most generous usage limits of any US provider right now, so I'll put up with the announcement theater.