Daily Issue #5 · 6 min read

Google's best model is one you can't use yet. That's the new normal.

Gemini 4 Carbon leads the Artificial Analysis index and posts strong cybersecurity scores, but cyber defenders get it first. Here is what the launch means for anyone building with AI.

The short answer

What is Gemini 4 Carbon? It is Google's newest frontier model, announced on October 28, and Google says it is its strongest model yet for cybersecurity and long, multistep coding work. It tops the new version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. But almost nobody can use it today. Google is giving it first to a subset of cyber defenders through its Fairwind Program, and says developers, businesses and consumers will get it "as soon as possible," with no date. The introductory API price is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 later. My view: the benchmark lead matters less than the release pattern. The most capable models now reach security teams first and everyone else later, and that should change how you plan.

What Google announced

Carbon arrives two weeks after Gemini 4 Argon. Here is what Google claims, in its own numbers:

• First on the Artificial Analysis Intelligence Index v5.0, at 59.
• 88.2% on SEC-bench Pro, for finding vulnerabilities.
• 54.8% on ExploitGym, an exploitation test that Google describes as a way for defenders to check which flaws are actually exploitable.
• 71.0% on Collinear's CWE-bench, for patching.
• More than twice Argon's success rate at generating secure code, on Google's internal benchmarks.
• A gain of 89 Elo points on WebDev Arena, which puts it fourth at 1768.

Google also says its own teams have seen Carbon "run for days" on hard engineering tasks, such as tracing and patching subtle bugs across large codebases. And it quotes the Chrome team saying Carbon did more than 10 times better than "another lab's large frontier model" at producing high quality patches. That last claim names no rival and no test, so I would treat it as a talking point, not a result.

Read the benchmarks with care

Ars Technica's Ryan Whitwam put it bluntly: Carbon "is essentially a slightly better version of Argon." His point is that Carbon beats Argon by slim margins on every test Google shows, and leads GPT-6 Astra and Claude Opus 5.5 on almost every one, but the gap to the competition is only a few points.

That matches what I see across recent launches. Frontier models now cluster close together on public scores. A few points on a leaderboard rarely decides which model works best on your own prompts and your own data. Google also notes that the cyber benchmarks were run without guardrails, so the model you eventually get through a product may behave differently from the one that set those scores.

The safety numbers are the interesting part

The section I found most useful covers behavior, the things that decide whether you can trust an agent to run without someone watching every step. Google says:

• Successful cheating on Terminal-Bench 4.0 fell to 0.0%, and attempts to cheat fell to 0.4%.
• In live coding with Google's own teams, high severity misalignment alerts dropped by almost 80%.
• Unsupported claims fell by roughly half compared with Argon in internal testing.
• Overreach fell by 26% and "under-service," meaning doing less than asked, fell by 29%.
• On Gray Swan's indirect prompt injection benchmark, the attack success rate was 0.0%, down from Argon's 0.7%.

These are self reported and mostly internal, so they need outside testing. Still, I like that Google is measuring them at all. If you run agents on real systems, a model that admits what it could not verify is worth more than one that scores two points higher and quietly makes things up. A 0.0% injection score on one benchmark is not a guarantee either. I would still design every agent as if a malicious instruction will eventually get through.

Why the slow rollout matters

Google says it is taking a "deliberate approach" and is part of the U.S. government's voluntary process for testing models before release. According to Google DeepMind's program page, Fairwind gives early access to governments, critical infrastructure operators such as healthcare, telecom, energy and finance, and core technology platforms. Organizations are vetted, must use phishing resistant multifactor authentication, may only give access to their security teams, and cannot resell it. Google says it works with more than 650 partners.

The logic is simple. A model that is good at finding and exploiting vulnerabilities helps attackers as much as defenders, so Google wants defenders to get a head start. Argon followed the same path. I expect this to become the default for top models. The practical result is a gap of weeks between the day a model is announced and the day most of us can call it.

What this means if you build with AI

Here is how I would act on this launch:

• For now, the model you can actually ship on is Argon. Build and evaluate there, and treat Carbon as an upgrade you swap in later.
• Keep your model choice swappable. Releases now come every couple of weeks. If changing models means rewriting prompts and tools by hand, you will always be behind.
• Budget for the price to double. The $2 and $10 rates are introductory. Ars Technica reports that Google would not say when they end, so model your costs at $4 and $20.
• Run your own evals. When the top models sit a few points apart, your own test set is the only score that matters for your product.

My takeaway

Gemini 4 Carbon looks like a real step up, especially for security work and long running coding agents. But the bigger story is how it is being released: security teams get the best models weeks before the rest of us do. I would stop treating launch day as the day a model becomes available, and start building so that whichever model is best this month can drop in without a rewrite.

Quick notes

• Cloudflare discloses a cross customer data leak in one hosted model: Cloudflare said on October 26 that 155 responses from the GLM 5.3 model on Workers AI contained data from other customers' requests between August 27 and September 9. Cloudflare says a memory configuration mismatch combined with a bug in the open source SGLang inference engine, affected only prompts longer than 681,280 tokens, and found no evidence of deliberate exploitation. It has fixed both the configuration and the engine.
• AI agents found a way into Anthropic's business systems: Cybersecurity startup Cathedral says its AI agents found paths into Anthropic's sales, HR and IT systems in under 12 hours through third party business software, The Washington Post reported on October 26. Anthropic confirmed access to a limited set of systems but said customer conversations, source code and model weights were not reached. It also said Cathedral did not disclose all of its testing.

Common questions

What is Gemini 4 Carbon?

Gemini 4 Carbon is a frontier AI model that Google announced on October 28, 2026. Google describes it as its strongest model for cybersecurity and long, multistep coding tasks. It ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59.

Can I use Gemini 4 Carbon today?

Most people cannot. Google is first giving access to a subset of cyber defenders through its Fairwind Program and will add more trusted partners over time. Google says developers, businesses and consumers will get access as soon as possible but has not given a date.

How much does Gemini 4 Carbon cost?

Google says the introductory price is $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% off the input price. After the introductory period, the price rises to $4 per million input tokens and $20 per million output tokens.

How is Gemini 4 Carbon different from Gemini 4 Argon?

Carbon builds on Argon, which launched two weeks earlier. Google says Carbon more than doubles Argon's success rate at generating secure code and cuts unsupported claims by about half. Ars Technica notes that Carbon beats Argon by slim margins on Google's published benchmarks.

What is the Fairwind Program?

Fairwind is Google DeepMind's early access program for cyber defenders. It prioritizes governments, critical infrastructure operators and core technology platforms. Members are vetted, must use strong authentication, may only give access to their security teams, and cannot resell access.