Gemini 4 Carbon is an announced model, not a model most readers can try yet. Google says it is starting with selected cybersecurity defenders, then trusted partners. Artificial Analysis puts Carbon (high) first on its Intelligence Index v5, but that does not settle which model is best for every coding, research or business task. In his video, Matthew Berman walks through the announcement and scorecards; he explicitly says he has not used Carbon himself. That makes this a launch analysis, not a hands-on review.
Carbon looks like a substantial advance over Gemini 4 Argon on published evaluations, especially reasoning and cybersecurity. Independent testing also shows stronger coding but not a clean sweep. Google has not given a public access date. Wait for a real workflow test before switching models or budgeting around the launch discount.
Watch Matthew Berman's Breakdown
What the video establishes: Berman reads Google's launch post and third-party charts and offers his interpretation. The Factory segment is sponsored. Its product and savings claims are advertising, not evidence about Carbon, and are not used in the comparisons below.
What Google Released, and Who Can Use It
Google announced Carbon on 28 October 2026, two weeks after Argon. It describes longer-running engineering work, better coding and stronger cyber-defence capabilities. Starting this week, access goes to a subset of cyber defenders, then expands gradually to trusted partners through the Fairwind program. Developers, businesses and consumers are promised later, but Google gives no firm date. A product page saying "Try Gemini 4" does not mean the Carbon variant is already in every Gemini surface.
This also changes the status of earlier Carbon coverage: before the announcement, its name and performance circulated as a leak. The official release post now confirms the model and access plan. A specific training recipe, public API model ID and broadly available download are not established by that post. Do not confuse Carbon with the already accessible Argon model.
Read the Benchmark Lead in Context
Artificial Analysis's 28 October evaluation gives Carbon (high) 59 on Intelligence Index v5, its highest score at publication. That is an independent aggregate, not a universal win. Its Coding Agent Index places Carbon second to Claude Opus 5.5; its knowledge-work Briefcase score improves on Argon but trails several rivals. Google also publishes a broader model comparison, whose individual tests use different harnesses and conditions.
| Measure | Carbon result | How to read it |
|---|---|---|
| AA Intelligence Index v5 | 59, first at launch | Composite reasoning/knowledge/science result; not a claim of first place in every domain. |
| AA Coding Agent Index | 59, second to Opus 5.5 at 62 | Measured with Antigravity CLI for Carbon and Claude Code for Opus; harnesses are part of the comparison. |
| AA Briefcase | 1609 Elo, up 126 from Argon | Strong improvement in knowledge work, but not the highest-scoring model. |
| SEC-bench Pro / ExploitGym / CWE-bench | 88.2% / 54.8% / 71.0% | Cyber scores reported in Google's launch post, evaluated without production guardrails. They do not imply permission to test others' systems. |
The video's strongest conclusion, that Google is back in the frontier conversation, is supported by the published aggregate and the size of the jump from Argon. The stronger conclusion, that Carbon is the best model for your work, still needs accessible, repeatable task tests.
Token Price Is Not the Same as Task Cost
Google lists an introductory API rate of $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. Its footnote says the rate becomes $4/$20 afterward; it does not specify when that period ends. These are listed rates for a future broader rollout, not proof that a public API endpoint is live today. See the pricing note in Google's announcement.
On its own test suite, Artificial Analysis calculates $7.76 per Intelligence Index task at the standard rate, versus $10.04 for Opus 5.5 and $6.25 for GPT-6 Astra. It records about 123,000 output tokens per Carbon task, against 173,000 for Opus in that setting. The output-token qualifier matters: these are not total input-plus-output tokens, and a different prompt, harness, cache hit rate or success criterion changes the bill. Berman's "intelligence density" idea is a useful way to think, but measured cost per acceptable result is the decision metric.
Cybersecurity Strength and Safety Are Different Claims
Google says Carbon's cyber scores are strong enough to warrant a staged release. It also reports internal reductions in unsupported claims and high-severity misalignment alerts. In its safety readout, successful shortcut-taking on Terminal-Bench 4.0 is 0% and attempts are 0.4%. Berman says 4% at 10:32; the official article says 0.4%. Google also reports 0% successful attacks on the evaluated Gray Swan indirect prompt-injection set. Those results describe the tested conditions, not immunity to new attacks or a guarantee of safe autonomous behavior.
Cyber benchmarks compare capability; deployment requires scope and authorization. The initial defender-only rollout is an access decision, not evidence that the public should run unattended security agents. Anyone evaluating a future release should use their own code or written authorization, keep tests within an allowlist and independently verify findings before acting.
What to Test When Access Opens
| Your question | Fair first test | Record |
|---|---|---|
| Does it improve coding? | Run the same bounded issue in the same repository with the same tests and review process used for your current model. | Accepted fix, regressions, human repair time, elapsed time and total cost. |
| Does long-horizon work hold up? | Give both models one multi-step but reversible task with checkpoints and a known completion criterion. | What each agent actually completed, where it stopped and how many iterations needed supervision. |
| Is the price worthwhile? | Compare the total cost of a successful result using real cache settings, not the advertised token rate alone. | Input/output/cache tokens, retries, tool charges and reviewer time. |
Keep the harness and evaluation as consistent as possible, state what changed, and inspect the deliverable. Until public access exists, these are an evaluation plan, not a tutorial claiming we tested Carbon.
Compare It With Argon
For a hands-on baseline, Berman's earlier Gemini 4 Argon review shows a model he could actually use, including 3D and coding examples. It is not a Carbon demo. Watching it beside this launch analysis makes the difference between visible output and a pre-access scorecard clear. Google's Argon announcement supplies the release context.
Video Chapters
The launch and access
| Time | Topic |
|---|---|
| 00:00 | Carbon announced; not publicly released |
| 01:10 | Artificial Analysis Index v5 |
| 02:28 | Why cyber defenders get access first |
| 04:27 | Introductory price and cached input |
Benchmark and safety readout
| Time | Topic |
|---|---|
| 07:07 | Coding, computer use and cyber benchmarks |
| 10:32 | Google's alignment and safety results |
| 11:38 | Output tokens per task and efficiency |
The model-market verdict
| Time | Topic |
|---|---|
| 12:32 | Quality versus cost on the frontier |
| 14:32 | Berman's conclusion and what remains untested |
Sources and Provenance
Checked 29 October 2026. The Matthew Berman video and supplied transcript support his analysis and his statement that he had not personally tested Carbon. Google's launch article and model comparison support rollout, list pricing and Google-reported cyber/safety figures. Artificial Analysis supplies its independently run aggregate, coding, knowledge-work and cost-per-task results; its model page can change as tests update. We did not receive Carbon access, reproduce these evaluations, validate Google's internal security tests or run an app comparison. Figures are a dated snapshot, not a standing ranking.