A failing test that points straight to a typo is easy to hand to a coding agent. A bug that appears only when a background job retries a payment takes a different kind of work. The agent has to follow the request across services, reproduce the failure and verify that the fix leaves other behavior intact. Google’s Gemini 4 Carbon, which began rolling out more broadly on November 10, 2026, is designed for that second kind of job.
Where Carbon fits alongside Argon
Google first gave a group of cybersecurity defenders access to Carbon through its Fairwind program in October. The November rollout brings Carbon to paying Gemini API developers and businesses. Google AI Ultra subscribers are also getting it in the Gemini app, AI Mode in Search and Google Antigravity. Google continues to position Gemini 4 Argon for everyday work, while Carbon is designed to spend more tokens investigating hard problems and verifying its work. Google’s launch announcement
For a development team, that makes Carbon worth testing on issues that stay unresolved after a first pass. Take the payment bug. An agent could trace where the retry originates, write a test that reproduces the duplicate charge, then check whether its patch also handles concurrent retries. A focused change with a clear acceptance test is a good starting point for Argon. Carbon has more room to help when the cause is unclear and several plausible fixes need to be ruled out.
What Antigravity’s /swarm adds
Google is also previewing /swarm in the Antigravity app and CLI. A coordinator can split an assignment among agents working in separate Git worktrees, with other agents tasked with testing and challenging the changes. Different models can fill different roles, so Carbon could investigate a tricky dependency while Argon handles well-defined edits. Google says parallel runs consume quota faster than a standard chat. Google Antigravity’s /swarm announcement
That setup is easiest to justify when work can actually proceed in parallel. In a library migration, one agent might update test fixtures while another checks how the new version handles serialization. If both jobs depend on the same unresolved API decision, settling that decision first can save two rounds of rework.
What the first coding benchmarks show
On CursorBench 4.0, which uses ambiguous coding tasks spanning multiple files, Cursor listed Carbon High at 60.4% and $9.22 per task, and Claude Fable 5.2 Max at 60.0% and $32.79 per task. Cursor calculates those costs from published token rates and measured usage. Cursor also cautions that small score differences may reflect test variation. These results compare two specific configurations and show a larger gap in estimated cost than in score. CursorBench 4.0 leaderboard and methodology
Google DeepMind’s model card reports a different ordering on FrontierSWE 2. Carbon scored 61.2%, ahead of Argon at 55.0%, while GPT-6 Astra scored 65.5%. Which model comes out ahead depends on the tasks, tools and evaluation setup. Google DeepMind’s Gemini 4 Carbon model card
API pricing and how to run a useful trial
Google announced the following rates for the public Carbon model. The introductory rate matches Argon’s. Google’s Gemini 4 Carbon pricing announcement
| Period | Per million input tokens | Per million output tokens |
|---|---|---|
| Through November 30, 2026 | $2 | $10 |
| From December 1, 2026 | $4 | $20 |
The amount of work billed can differ even at the same token rate. A longer investigation may use more tokens and still cost less overall if it eliminates retries or cuts review time. The reverse is possible when the task is straightforward and both models reach the same answer.
To compare them, take a few closed issues and give both models the same starting commit, tools and time limit. Keep the original fixes out of view. Check whether each patch passes the reproduction test and the broader test suite, then record API spend and the time a reviewer needs to approve it. If Argon reliably handles small fixes and Carbon resolves more of the issues that cross subsystem boundaries, that gives the team a concrete rule for when to switch models.