Carbon Just Opened to Ultra Subscribers — Here Is What Changed

Google announced Gemini 4 Carbon on October 28, 2026, but held it back: access was restricted to vetted organisations through the Fairwind programme, specifically cyber-security defenders. As of November 10, that began to change. 9to5Google reported that Carbon is now rolling out in the Gemini app for AI Ultra subscribers, with Argon remaining the default for lower tiers. Third-party benchmark tracker BenchLM added Carbon to its index on the same date, suggesting broader access is arriving. Google had not publicly listed a Carbon model ID in its API documentation as of late October; teams should verify availability in Vertex AI or Google AI Studio before building against it in production.

The pace is striking. Gemini 4 Argon only reached general availability on October 13 — Carbon’s announcement followed 15 days later. Google describes Carbon as the same safety foundation as Argon, plus additional alignment training, with particular emphasis on security workflows and long-horizon agentic tasks (source: VentureBeat, The Decoder).

One timing note: Argon’s introductory pricing ends November 30, per VentureBeat. Teams on Argon have three weeks to decide whether Carbon — which carries the same introductory rates — is the right next step.

Benchmarks: Google’s Table vs Independent Scores

Google published a 17-benchmark table comparing Carbon against Argon, GPT-6 Astra, and Claude Fable 5.1. Carbon leads 15 of 17 rows. This table is vendor-produced: Google chose the benchmarks, ran the evaluations, and selected which competitors to include. The table does not include Claude Fable 5.2, only Fable 5.1. Read it accordingly.

The independent picture is more nuanced. BenchLM’s composite ranking (214 models, data as of November 10, 2026) scores Carbon at 87.91 — second place, behind Claude Fable 5.2 at 91.41. The BenchLM page for Carbon was partially truncated during research; treat sub-category figures as directional, not definitive.

Key benchmark rows for engineering teams, pulling from Google’s table (vendor-run) and BenchLM for the independent composite:

Benchmark scores from Google’s vendor-run table (via VentureBeat, CellCog, The Decoder). Google’s table compares against Fable 5.1; BenchLM’s #1 position belongs to Fable 5.2, which is not included in Google’s comparison.

Three headline findings for a technical audience. First, Carbon leads on agentic coding (Terminal-Bench 4.0: +8.1pp over Argon) and long-horizon reasoning (Agents’ Last Exam: +8.4pp over Argon). Second, Carbon’s Gray Swan indirect prompt-injection attack success rate is 0.0%, down from Argon’s 0.7% — a meaningful improvement for teams running multi-agent pipelines where prompt injection is a live threat (source: Google red-team data, via VentureBeat and The Decoder). Third — and this matters most if your use case is coding — GPT-6 Astra leads FrontierSWE 2 by 4.3 percentage points (65.5% vs 61.2%), and BenchLM’s independent composite ranks Claude Fable 5.2 comfortably ahead of Carbon in the overall standings.

Cost per Task: Carbon vs Argon vs Astra

Carbon’s introductory pricing matches Argon exactly: $2.00 per million input tokens, $10.00 per million output tokens, and $0.10 per million cached input tokens — a 95% discount on the input rate for cached context. The post-introductory rate is $4.00/$20.00; Google has not disclosed when the intro period ends (source: VentureBeat, CellCog, Smartscope). Argon’s intro pricing ends November 30.

GPT-6 Astra’s standard API rate is $10.00 per million input tokens and $50.00 per million output tokens — five times Carbon’s introductory input price — with a long-context surcharge for inputs exceeding 272,000 tokens ($20.00 input, $75.00 output) and Fast mode at 2× the standard rate (source: eesel.ai, citing OpenAI API documentation).

Artificial Analysis, an independent benchmark evaluator, computed cost per evaluation task across four models using its standard protocol. These figures, as reported by The Decoder:

Source: Artificial Analysis, an independent benchmark and pricing evaluator, as reported by The Decoder (November 2026). Task definition follows Artificial Analysis’s standard evaluation protocol; task composition may differ from your workload.

The counterintuitive result: despite Astra’s 5× higher raw-token price, it delivers lower per-task cost on Artificial Analysis’s evaluation ($6.25 vs $7.76), because Astra produces more concise outputs on the same tasks. If your workloads involve long system prompts you can cache — agentic pipelines with stable context, for example — Carbon’s 95%-off cached input rate narrows or reverses that delta. Do not assume per-task cost scales linearly from raw-token pricing.

For Swiss and EU teams

Google has not disclosed which regions will gain access to Carbon or on what timeline. The Fairwind programme had no stated regional targets, and the November 10 Gemini Ultra expansion has not specified EU availability separately (source: Smartscope). For teams that route through OpenAI for EU data-residency reasons: GPT-6 Astra endpoints on EU data-residency carry a 10% surcharge for models launched after March 5, 2026, and Fast mode is unavailable on EU data-residency requests (source: eesel.ai, citing OpenAI API documentation). Until Google confirms Carbon’s regional availability and publishes a model ID in its API documentation, EU-based teams should not migrate production agentic workloads from Argon.

Verdict: Use It / Test It / Skip It

Use it now if you are running security-focused agentic workloads and you are already a Gemini Ultra subscriber. Carbon’s ExploitGym score (54.8%) is 12.4 percentage points ahead of GPT-6 Astra (42.4%) and 20.3 points ahead of Claude Opus 5.5 (34.5%). The 0.0% Gray Swan indirect prompt-injection rate matters for multi-agent pipelines where unvetted tool outputs feed back into the agent’s context. The access path is now open if you are on Ultra.

Test it if you are on Gemini 4 Argon and your workloads are more agentic than narrowly coding-focused. Terminal-Bench 4.0 and Agents’ Last Exam are both meaningfully higher, the per-task cost is essentially flat, and the prompt-injection safety improvement is real. Run your own eval suite — do not use Google’s vendor table as the trigger. See our guide on evaluating a new frontier model before committing production traffic.

Skip it for now if FrontierSWE 2 performance is your primary criterion. GPT-6 Astra leads that benchmark by 4.3 percentage points, and BenchLM’s independent composite still places Claude Fable 5.2 clearly ahead of Carbon. Carbon’s API availability via Vertex AI is also not yet publicly confirmed as of this writing — wait for the model ID to appear in Google’s official API documentation before building against it in production.

One-line verdict: Carbon is a meaningful step up from Argon for agents and security at essentially the same cost — but it is not yet the clear frontier leader that Google’s vendor table implies.

Further Reading

• VentureBeat: Google unveils Gemini 4 Carbon — The most detailed reproduction of Google’s benchmark table, with caveats on early access restrictions and pricing.
• The Decoder: Gemini 4 Carbon two weeks after Argon — Covers the Artificial Analysis cost-per-task figures and the AA Intelligence Index context in depth.
• How to Evaluate a New Frontier Model Before Switching — Our evergreen guide on building your own eval suite so vendor benchmarks do not drive your stack decisions.

Your turn: Has your team started testing Carbon, or are you waiting for a confirmed API model ID before switching from Argon? Reply to our newsletter or send us a note — we feature the best answers in the Friday Scorecard.

How this article was made: AI researched and wrote this article under the Adrian persona, using the sources linked above, and it was published automatically without a human edit. Editorial guidelines are set by Adi. Spotted an error? Tell us and we will correct it.