Gemini 4 Carbon scores 65.2% on terminal bench, undercuts rivals at $2/$10 per million tokens

Google DeepMind's Gemini 4 Carbon scored 65.2% on terminal bench 4.0, beating Opus 5.5 and GPT6 Astra, which both landed around 59%. The model, now live, is the most capable release in the Gemini 4 lineup, which also includes the earlier and faster Gemini 4 Argon. According to Sam Witteveen, its longer task times and higher token generation make it best suited to autonomous workflows rather than interactive sessions.

On the automation bench for knowledge work, Carbon reached approximately 54%, against low-40s scores for comparable models. On the artificial analysis intelligence index, it initially ranked first with 60 points before Anthropic's Fable 5.2 overtook it with 65 points.

What does the performance cost?

The benchmark gains carry measurable trade-offs. Carbon requires 26 minutes per task on the artificial analysis intelligence index, twice the median of 13 minutes. It generates approximately 205 million tokens per task, nearly three times the median of 72 million tokens.

Those figures translate directly into spend. Carbon costs just under $7.80 per task on the index, versus $26.50 for Fable 5.2. That is roughly a third of the price for a five-point lower score.

Google's introductory pricing is $2 per million input tokens and $10 per million output tokens, matching Argon's launch pricing. The standard rate is $4 in and $20 out. The offer runs through November 30, 2026. Most cited benchmarks were created before Fable 5.2 shipped, which may affect direct comparisons.

What can it build?

Google identifies four focus areas for the model:

• Multi-day agentic engineering: deeper reasoning, surgical code edits, verification loops and cyber defense.
• High-precision knowledge work: finance, legal, healthcare and exhaustive retrieval, with a claim of zero fabricated citations.
• Frontier discovery: physics, quantum computing, genomics and drug discovery.
• Full-stack front-end and 3D web development: realistic physics, interactive game mechanics and self-verifying interfaces.

In testing, Carbon produced a native Mac Pomodoro app written entirely in Swift, including a compiled, ready-to-run version. The app added unrequested features such as a menu bar controller and project tracking. Xcode was installed in the test environment, though it remains unclear whether the model used it for signing.

The model also built a mashup combining features from Bear, a note-taking app, and Obsidian, a thinking-linking app. For 3D work, it rendered an interactive model of Marina Bay Sands in Singapore, including surrounding landmarks such as the Apple store sphere and the Louis Vuitton store pyramid, with minimal additional prompting. Observations indicate Carbon self-verifies and refines outputs iteratively during such tasks.

What are the API limits?

The API ships with preset values for temperature, top P and top K, which cannot be modified. Users can adjust the reasoning effort level, but the fixed sampling parameters may limit customization for some use cases.

Gemini 4 Argon remains the faster option, but testing found it struggles with tool calling and tends to overthink tasks instead of using search effectively. Both models are the same size; Carbon is newer and more capable.

What does this change for software businesses?

Developers at small and medium businesses may use tools like Carbon to replace high-cost SaaS subscriptions. Its ability to research and replicate app features could encourage more users to create and open-source their own software. Models that research apps via videos, screenshots and descriptions further lower the barrier to custom application development.

Software vendors may respond by shifting toward minimal front-end interfaces with intelligence and complexity hosted in the cloud, making cloning harder. Small software houses are expected to stay relevant by shipping polished apps that address user needs directly.

Speculation around the pricing suggests Google may extend the introductory rates or introduce a newer model at similar cost. For teams that prioritize cost over peak performance, Carbon currently offers the strongest price-to-capability ratio in its class.