Gemini 4 Carbon Tops Three Benchmarks: When to Use It vs. Cheaper Models

Introduction

Google has fully opened Gemini 4 Carbon, pushing model competition into a more practical dimension: not only winning in programming, search, and automation benchmarks, but also driving down the cost of complex tasks. Carbon ranks first in CursorBench 4.0 (AI coding), Perplexity WANDR (information search and deep research), and Zapier AutomationBench (enterprise software automation). It also integrates Antigravity's /swarm feature, allowing a single prompt to orchestrate a team of agents that handle architecture, implementation, review, and testing in parallel.

However, Carbon is not optimal for every task. Complex, open-ended, long-reasoning work suits Carbon; daily coding and knowledge Q&A are better served by the lighter Argon. Mature usage means matching model capability, task complexity, and budget — not simply throwing the strongest model at everything.

Three Benchmarks, Carbon Takes First in All

Carbon leads in three evaluations that reflect real developer work:

CursorBench 4.0 — AI programming and code tasks: Carbon ranks first.

Perplexity WANDR — Information search and deep research: Carbon ranks first.

Zapier AutomationBench — Enterprise software automation: Carbon ranks first.

These three directions cover the core daily activities of developers: writing code, finding information, operating business software. Benchmarks are shifting from "answering ability" to "can the model finish a job," moving closer to real agent workflows.

Benchmark leadership does not guarantee superiority in every scenario. Anthropic's concurrently released Fable 5.2 scores higher on some comprehensive and real software engineering tasks. Different models may excel at code, search, vision, long-horizon tasks, or cost control. Enterprises should route based on actual tasks rather than relying solely on aggregate leaderboards.

/swarm: An Agent Team Working in Parallel

Google added the /swarm function to Antigravity. Previously, a single agent handled understanding, writing, testing, and modifying code; now a single sentence can dispatch a group of agents to work in parallel.

A typical division of labor:

Carbon: architecture design and complex problem analysis
Fast models: routine code, tests, and documentation
Review model: identifies security, performance, and maintainability issues
Execution model: runs builds, tests, and integration checks

The key to multi-agent collaboration lies in task boundaries. Each agent needs independent context, clear inputs/outputs, permission limits, and acceptance criteria; a coordination node then merges results to avoid conflicts when multiple agents modify the same file.

Google demonstrated several cases: reconstructing the Eiffel Tower construction process from 1887–1889 historical photos and engineering drawings into an interactive 3D exhibit; and using 206 engineering drawings from 7 municipal planning PDFs, executed over 10 overnight runs, to rebuild a 3D model of an office area containing ~900,000 triangles. DeepMind also attempted reconstructing the Apollo lunar module from historical photos and technical drawings. For such tasks, the agent's value is not just generating a model but simultaneously handling data retrieval, geometric reasoning, code generation, and result verification.

Cybersecurity: Finding Vulnerabilities and Enforcing Boundaries

In Google's internal vulnerability scanning system Project Fortify, Carbon achieved 100% recall and 95% triage precision. Collaborating with security firm Wiz, it discovered a zero-day vulnerability in under two hours. Google's vulnerability research team also used Carbon to prove that two open-source software flaws were exploitable, helping maintainers reassess severity and release patches.

As models become better at finding vulnerabilities, they require stricter safety guardrails. In simulated attack tests, Carbon blocked 99% of attack paths; single-breach probability dropped from Argon's 3.1% to 0.8%. External institutions tested continuously for 132 hours with over 13,000 requests and found no new universal jailbreak methods.

The strongest cybersecurity version is not openly released to all users but managed through the controlled Fairwind program. For enterprises, a security agent must at minimum have target allowlists, network isolation, short-lived credentials, action approval, and full audit trails — the model must not autonomously decide scan scope or exploitation targets.

Research Agent: From Literature Search to Experiment Verification

Carbon shows dense applicability in research scenarios.

Cryptography: Google researchers built the AlphaProof Code framework, having Carbon generate millions of lines of Lean code to formally verify the security of Signal protocol's underlying code. Work that previously required expert teams for years was compressed to weeks.

Protein research: Carbon searched massive unlabeled sequence archives for homologs, increasing candidate structures for prediction by 40× on average. In lab testing, it detected insufficient automated mixing, overlapping chromatography peaks, and instrument carryover from previous samples — anomalies humans often miss.

These cases show that research agents add value not merely by generating summaries, but by placing data, programs, experimental records, and verification standards into a single workflow. Researchers still need to confirm data sources, method validity, and final conclusions; agents cannot replace peer review and experimental replication.

Carbon's Price and Real Cost

Carbon's current API pricing matches Argon: $2 per million input tokens, $10 per million output tokens. Fable 5.2 costs $8 input / $40 output — Carbon is roughly one-quarter the price. However, this promotional price lasts only until November 30, after which it rises to $4 input / $20 output.

Low per-token price does not guarantee low per-task total cost. Google acknowledges Carbon consumes more tokens on complex tasks because it performs deeper reasoning and multi-round verification.

Google suggests calculating real cost with this formula:

Total task cost = input tokens
  + output tokens
  + tool calls
  + retries and context recovery
  + build and test resources
  + human review and rework

If Carbon completes a complex task in one shot while a cheaper model needs multiple rework cycles, Carbon may be more economical; conversely, using Carbon for simple tasks wastes budget.

How Developers Should Choose Models

Google provides clear model segmentation: daily programming and knowledge work use Argon; complex, open-ended, long-reasoning tasks upgrade to Carbon.

Enterprises can design a three-layer model routing:

High-frequency simple tasks → low-cost models.

Complex code, research, and multi-agent orchestration → Carbon.

Production releases, sensitive data, and external side effects → mandatory human approval.

Carbon can also act as Advisor or Orchestrator: it decomposes complex tasks, cheaper models execute individual steps, and Carbon reviews results. This leverages strong planning while controlling large-scale invocation costs.

Domestic developers needing to connect tools like Claude Code to compatible APIs can refer to Code80 (https://code.ai80.vip/home) for endpoint configuration; model routing, keys, data retention, budget, and production auditing should still be managed by the project side.

Frequently Asked Questions

Is Gemini 4 Carbon the strongest for all tasks?

No. It suits complex programming, deep research, enterprise automation, and long-horizon agent tasks; ordinary Q&A and daily coding can use lower-cost models.

What problem does /swarm solve?

It enables multiple agents to work in parallel, assigning architecture, implementation, review, and testing to different models, but requires task isolation, state management, conflict handling, and unified acceptance.

Will Carbon stay at $2 per million input tokens?

No. That price lasts until November 30; afterward input becomes $4 and output $20 per million tokens.

Why doesn't low per-token price equal low total cost?

Complex tasks may consume more tokens, tool calls, and retries. Input, output, execution resources, human rework, and test costs must all be counted.

Is Carbon suitable for direct operation of enterprise systems?

It should be used in controlled environments. Target systems, networks, credentials, file permissions, and external actions must pass through policy gateways and human approval.

How can domestic developers access compatible APIs?

They can learn integration via Code80 while independently managing model permissions, keys, data retention, call budgets, and audit logs.