Google has announced Gemini 4 Carbon, a model focused on cybersecurity and long-horizon engineering work, but access is initially restricted to selected cyber defenders. The launch pairs high-stakes security claims with a staged rollout and published API prices, giving developers something concrete to weigh while public access remains out of reach.
Google DeepMind Watch analysis
What happened
Google says Carbon can work through engineering tasks for days, diagnosing and patching bugs across large codebases. It is opening access to a subset of cyber defenders first, then plans to add trusted partners through its Fairwind programme. Google has not given a date for general access.
For now, Google lists introductory API pricing at $2 per million input tokens and $10 per million output tokens. After the introductory period, those prices rise to $4 and $20 respectively. Cached input tokens cost 95% less than the standard input rate during the introductory period.
Google also reports scores of 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. The company says Carbon produced secure code at more than twice Argon’s success rate in its internal benchmarks. These are Google’s reported results, and the external cybersecurity benchmarks were evaluated without guardrails, a detail worth keeping in view when considering what the scores mean.
Why it matters
The initial audience is telling: Google is making a security-focused model available to defenders before opening it to developers and businesses generally. That gives a limited group an early chance to test the tool, while keeping wider access behind a staged rollout and further work on guardrails.
The prices also offer an early budgeting reference, not a settled long-term rate. The gap between introductory and standard pricing is substantial, so anyone assessing a future deployment will want to know when the offer ends and what the model can reliably do in their own environment.
Our read
Carbon is a consequential launch because it brings cybersecurity capabilities, restricted access and real API prices into the same announcement. The benchmark figures are worth noting, not treating as a verdict: impressive percentages do not tell you how the model performs on your systems or under your safeguards. Google has supplied enough to start that conversation, but not enough to settle it. Security teams, in particular, should distinguish a model’s reported test performance from evidence that it is safe to hand a live vulnerability workflow.
What to watch
- When Google expands the Fairwind programme and opens access to other developers.
- When introductory pricing ends and the standard rates take effect.
- How independent testing compares with Google’s reported benchmarks, especially with safeguards enabled.
Discussion spark: Should a cybersecurity model with strong company-reported benchmark scores be opened to developers sooner, or is a defenders-first rollout the more responsible test?
Sources and evidence
- Source update (31 October 2026, 14:34 UTC)
not affiliated with, endorsed by, or operated by Google or Google DeepMind