Gemini 4 Carbon: Benchmarks, Pricing, Access and How to Prepare Now

Source: Author

Google DeepMind released Gemini 4 Carbon on October 28 and called it its strongest model yet. By the end of launch week it sat at #1 on Arena's Text Arena and #1 on the Artificial Analysis Intelligence Index.

The catch is that most of us can't use it. Access goes first to vetted cyber defenders in Google's Fairwind program, while everyone else gets a price and a promise of "as soon as possible." Here is the short version of what Carbon is, which numbers hold up, and how to be ready when it opens.

What Carbon is

• The top model in the Gemini 4 family, one step above Gemini 4 Argon.
• Pitched on long-horizon work and cybersecurity. Google says its teams have seen it run for days on multi-step engineering problems.
• A language model that reads text, images, video and speech and writes text. The world model DeepMind announced the day before is Genie 4.
• Independent evaluators list a 1M-token context window, a 1M-token max output and closed weights.

The scoreboard, with the fine print

• Arena Text Arena: #1 at 1,528, eight points above Argon and 19 above Claude Opus 5.5.
• Arena Agent Arena: #2, behind Opus 5.5 Max.
• Artificial Analysis Intelligence Index v5: #1 with 59, up from Argon's 50.
• Vals AI Index: #2 at 70.36%, just behind Claude Sonnet 5.5 at 70.85%.

Gemini 4 Carbon against Claude Opus 5.5 and GPT-6 Astra on four headline numbers

The lead is narrow, and the details matter more than the headline. On the Coding Agent Index, Carbon scored 59 inside Google's Antigravity CLI against 62 for Claude Code with Opus 5.5. On AA-Omniscience, accuracy rose from 48% to 58%, but the hallucination rate rose from 4% to 11%. Vals also measured 2.3 times Argon's reasoning tokens. Published Terminal-Bench 4.0 results disagree, so treat that one as a tie.

What it costs

Carbon launched at Argon's introductory price of $2 per million input tokens and $10 per million output tokens, with cached input 95% off. Both later double to $4 and $20, and Google hasn't said when. Artificial Analysis put Carbon at $7.76 per Intelligence Index task, about 23% cheaper than Opus 5.5. On Agent Arena its median task cost was $1.80, 72% below Opus 5.5 Max.

Who gets it, and when

Fairwind members come first, with priority for governments, critical infrastructure operators and core technology platforms. More trusted partners follow, then developers, businesses and consumers, with no date. Argon reached the paid Gemini API 13 days after its September 30 announcement. If Carbon follows that path, wider access lands in November, though Google hasn't committed to that.

Rings of access: vetted defenders now, trusted partners next, everyone else waiting with no date

Six ways to be ready on day one

1. Build on gemini-4-argon now and keep the model name in one config value.
2. Write a 10 to 20 prompt eval from your real work, with a clear pass for each and cost per task.
3. Route by phase: use Carbon for range, then hand the best draft to the model that finishes best.
4. Put stable material first in your prompts so the cache discount does its job.
5. Run long jobs in the background, checkpoint progress and make every step safe to retry.
6. Check anything factual with grounding or a second model before you publish.

Worth knowing

On creative work Carbon reaches further than precision-first models, but it doesn't always finish, and UI is its weakest area. Its boldest security claims are internal and can't be checked yet. Argon isn't on the Gemini API free tier, so budget for paid usage from the start.

Read the full Gemini 4 Carbon breakdown on Medium →

The cyber and safety scores, how Carbon handles creative work against rival models, the full access path and a model-by-phase routing table.