Google Opens Gemini 4 Carbon: $7.76 per Solved Task and 0% Breaches in an Injection Test

Two weeks ago it looked as though Google had pulled ahead again. Then Anthropic released Fable 5.2, and Carbon, barely out in general availability, already looks less fresh. The model cycle is spinning so fast that a release goes stale before it leaves beta. But for a developer, the headline race matters less than a practical question: what to plug in, and how much will it cost.

Not a replacement for Argon, but a tool for hard tasks

Google says outright that Argon remains the better fit for most development and knowledge-work tasks. Carbon is for when you have hit the ceiling of a model's intelligence: complex multi-step debugging, open-ended scientific research, long loops of verifying results. Both models cost the same per token, but Carbon reasons deeper and spends longer checking itself, so it generates more tokens. On the same task, the final bill will almost certainly be higher than with Argon.

Now the numbers. In Artificial Analysis testing, Carbon costs $7.76 per solved task at the standard price ($4 per million input tokens and $20 per million output tokens). That is cheaper than the older budget Flash 3.8, and the model performs better. For comparison: Opus 5.5 Max — $10.04, GPT-6 Astra Max — $6.25, Fable 5.2 at maximum — $26.50. At the High level, Fable 5.2 costs about $9 per task, but is still ahead by two points (62 vs. 60). Google's introductory price is currently half the standard one: $2 and $10.

On CursorBench 4.0, Carbon took first place with a score just under 60%. It also leads on Perplexity's WANDR and Zapier's AutomationBench, at less than half the price of its closest rival, Fable 5.2. On the new MKA benchmark for dense documents with charts, blueprints, and tables, Carbon is second, but more than six times cheaper than the leader.

Agent swarms: impressive, but debatable

Alongside the model, Google showed a /swarm mode in Antigravity, its answer to Claude Code. A single prompt turns into a team of agents whose work you can watch in real time. Carbon takes the roles that require deep thinking: architecture, coordination, auditing. Fast, lightweight models work in parallel in isolated environments, and some of them play the role of skeptic-verifiers that try to break the result.

Imagine refactoring a service. Carbon splits the task into pieces, small models edit modules simultaneously, and others run tests and hunt for bugs. In its demo, Google showed a 3D game about a pelican riding a bicycle around the Mountain View campus, a reconstruction of the building of the Eiffel Tower, and a playable 18th-century piano. Whether such swarms pay off is still debated in the industry: OpenAI and Anthropic sound cautious, and researchers are skeptical. The main gain is speed through parallelism, but orchestration is complex.

Security: big numbers with caveats

During early access through the Fairwind program, Carbon found real zero-day vulnerabilities at two large SaaS providers in under two hours (as part of the Wiz Scan for Good program). The Google Cloud Security team used it to write working exploits for two known bugs in a popular open-source project. In the Project Fortify system, the share of vulnerabilities found rose by more than 25% compared with Argon. But the lead on SEC-bench Pro, ExploitGym, and CWE-bench v1 applies only to the Fairwind version, and Google does not disclose specific figures.

According to Google, on its own honesty test of 375 tasks Carbon passes 92% (Argon — 70%) and in 98% of cases refuses to invent missing experimental data. In attack simulations, it blocked 99% of paths from adaptive red-team agents, and the chance of a breakthrough per attempt fell from 3.1% to 0.8%. External testers from Trajectory Labs found no universal jailbreak over 132 hours and more than 13,000 queries. On Gray Swan IPI, successful prompt injections — 0%. That does not mean the problem is solved: the test may simply have become saturated. But for agents with access to email, a browser, and code, the shift is noticeable.

What to do in practice

Keep Argon as your default model and bring in Carbon selectively: for complex debugging, code auditing, and research tasks. Compare not the price per token but the price per solved task on your own data: the benchmarks are published by the vendor itself, and the extra tokens from deep reasoning can easily eat up all the savings. A Flash version of Gemini 4 is expected soon, so cheap use cases can wait. Agent swarms are still worth trying only on isolated tasks with a clear success criterion. And do not drop your defenses: sandboxes and least privilege for agents are needed even with a zero result in a single test.