Gemini 4 Carbon opens up to all builders at Argon pricing

Summary: Google opened Gemini 4 Carbon to developers, enterprises, scientists, and consumers on November 10, 2026, two weeks after giving cyber defenders early access through its Fairwind Program. The headline fact for builders: Carbon launches at the same introductory price as Gemini 4 Argon — $2 per million input tokens and $10 per million output tokens — despite being positioned as the stronger, more token-hungry model. That combination changes the cost math for long agentic runs, and it’s worth understanding how before you rewire your stack.

What actually changed

Carbon is Google DeepMind’s newest frontier model, announced by Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect. When I covered the initial Carbon announcement on October 28, access was limited to a subset of cyber defenders. The November 10 availability post widens that in three channels:

The paid tier of the API in Google AI Studio and Gemini for business, for all paying developers and enterprises. Google’s separate developer post adds Android Studio and points to the Gemini API Developer Guide for setup.
Google AI Ultra subscribers get it in Google Antigravity, the Gemini app, and AI Mode in Search.
The scientist-access program that offered 10,000 researchers six months of free Argon in Antigravity now includes Carbon, and will keep providing the latest frontier models going forward.

One completeness note: the text of the availability post in my material cuts off mid-sentence in the scientist-program section, so details about that program’s ongoing onboarding are incomplete here. I have not drawn any conclusion from where the text stops.

How Carbon actually works: more tokens, longer loops

The mechanism behind Carbon’s positioning is straightforward. Google says Argon remains “the ideal choice” for most software engineering and knowledge work, while Carbon targets tasks that stress a model’s limits — complex multi-step coding and debugging, open-ended scientific research — and gets there by using more tokens than Argon to explore, reason deeply, and run extended verification loops. In the October post, Google’s teams described Carbon runs persisting for days through dead ends on tasks like root-causing subtle bugs across large codebases.

The companion feature here is /swarm, previewed in Google Antigravity. It turns a single prompt into a coordinated team of agents you can track in real time. The suggested pattern is assigning Carbon to the deepest-reasoning roles — architectural design, coordination, auditing — while pairing it with faster, lighter models across parallel isolated workspaces and adversarial verifiers that test every change. Google’s showcase examples range from a 906K-triangle Three.js reconstruction of an office building built from seven planning PDFs across ten overnight loops, to a multi-agent pipeline that reconstructed a playable 1780 fortepiano from CAD drawings and X-ray radiographs.

The pricing trade-off, spelled out

Carbon prices at $2 per million input tokens and $10 per million output tokens — identical to Argon’s introductory price — with cached input tokens at 95% off the input rate. That per-token parity is the interesting part, because Carbon’s own token consumption is higher.

Google’s benchmark framing tries to capture the net effect: Carbon ranks #1 on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench at more than 2X lower cost per task than the next best external model. On the new Multimodal Knowledge Agent benchmark built with Artificial Analysis, Carbon ranks #2 but at more than 6X lower average cost per task than the top model.

Read those as conditional claims. Cost-per-task on a benchmark with a fixed workload tells you less about your workload, where Carbon’s willingness to burn tokens on verification loops could either save you retries or inflate your bill. If your tasks are ordinary feature work, Argon probably still wins on total spend. Also keep in mind these benchmark numbers are Google-reported or partner-reported; independent replication at scale isn’t something these sources establish.

Safety numbers worth your attention

The October post included alignment metrics that matter if you’re running autonomous agents, and they carry through to the November launch:

Successful cheating on Terminal-Bench 4.0 dropped to 0.0%, with cheat attempts at 0.4% — while overall accuracy improved.
0.0% attack success rate on Gray Swan’s Indirect Prompt Injection Benchmark, up from Argon’s 0.7%.
Overreach reduced by 26% and under-service by 29%, meaning more reliable multi-step execution that stays in scope.
Roughly 2x fewer unsupported claims than Argon, with methodological caveats written into deliverables instead of fabricated around missing evidence.

On the defender side, Google’s Project Fortify saw 100% recall on scanning and 95% precision on triage on its internal benchmark, and Wiz’s Scan for Good team reported uncovering 0-day vulnerabilities in two major SaaS providers in under two hours. Google’s CVR team also used Carbon to prove two open-source bugs exploitable, with details withheld until responsible disclosure completes.

What to check before you switch

A practical sequence:

Run your most token-hungry, verification-heavy workflow on Carbon and the same workflow on Argon. Compare total token spend and task success, not leaderboard rank.
Price the caching path. If your agents replay large fixed contexts, the 95% discount on cached input tokens may dominate the comparison more than model choice.
If you use Antigravity, try /swarm on one multi-step project before committing a whole pipeline, and watch whether the adversarial verifier loop catches real regressions or just adds latency.
Treat the safety numbers as a starting point for your own evals — especially prompt injection resistance, which is cheap to test against your actual tool surface.

The honest limitation: almost every performance figure here traces back to Google or its partners, and the announcement material doesn’t establish independent verification. The price, the access channels, and the /swarm preview are concrete and verifiable today — start with those.

Sources

Original post

View the original post on X