Google has announced Gemini 4 Carbon, but most people cannot try it yet. In this five-minute explainer, Paul J Lipsky focuses on what could matter to everyday agent users: longer-running work, cybersecurity capability, a staged release and the eventual API price. The important distinction is that he is reading Google's announcement, not showing a hands-on Carbon test. For a deeper numerical comparison, see our Carbon benchmark and pricing analysis.
Google reports a big step beyond Argon, and an independent benchmark puts Carbon first on its overall intelligence index. But longer autonomous runs, fewer unsupported claims and stronger defences are claims to test in a real workflow once access opens. Paul's two-week public-release guess is a forecast, not a Google date.
Watch Paul Lipsky's Carbon Explainer
Source boundary: Paul's commentary is an interpretation of published launch material. Google's announcement supplies its rollout, price and internal-test claims. Artificial Analysis independently ran its own evaluations. Neither Paul nor this article demonstrates an ordinary user's Carbon session.
Why Longer Agent Work Is the Interesting Part
"Long-horizon" work means a task that takes many decisions and tool calls, not just a longer chat answer. Paul pictures an agent that can persist through dead ends while debugging or patching software. Google says its internal teams have seen Carbon run for days on difficult engineering problems. It does not publish a guarantee that a consumer agent will run unattended for days, or that every attempt finishes correctly.
A useful agent must do more than continue. It has to keep track of the goal, notice failed tests, ask for missing information, respect permissions and hand back an inspectable result. A long run that confidently follows the wrong plan is not progress. That is why Paul's optimism is worth pairing with a concrete question: how much of the job is accepted after human review?
Read the Cyber and Safety Scores Carefully
Google reports Carbon at 88.2% on SEC-bench Pro for finding flaws, 54.8% on ExploitGym for validating exploitability and 71.0% on CWE-bench for patching. Its launch post says these capability evaluations were run without guardrails. They are benchmark results, not an instruction to scan systems without authorization or a prediction of a company's real-world security outcome.
Paul also highlights Google's reported 0% successful prompt-injection attacks on the evaluated Gray Swan set and 0% successful shortcut-taking on Terminal-Bench 4.0. The first figure does not mean "impossible to trick"; it means none of the tested attacks succeeded under those conditions. Google says attempted cheating fell to 0.4%, which is distinct from the zero successful-cheating figure. Its claim that unsupported statements were roughly halved relative to Argon comes from internal testing. See Google's own wording before treating any of these as a safety guarantee.
Independent testing also complicates a simple "wins everything" headline. Artificial Analysis gives Carbon (high) the top 59 on its Intelligence Index v5, yet ranks it second behind Claude Opus 5.5 on its Coding Agent Index. Paul's account of Google's comparison is useful as a tour of the launch chart, not a replacement for task-level testing.
Who Gets Carbon First?
| Group | Status on 29 October 2026 | What it means |
|---|---|---|
| Selected cyber defenders | First access, beginning this week | Google says they can use it to assess and help secure critical software within the staged program. |
| Additional trusted partners | Gradual Fairwind expansion | No universal sign-up or fixed date is promised in the announcement. |
| Developers, businesses and consumers | Later, date unannounced | Google intends broader access. Until it announces a release, choose an available model for work you need done now. |
Paul predicts that Carbon might follow Argon's roughly two-week gap from selected-defender announcement on 30 September to public release on 13 October. It is a reasonable analogy, but not a Carbon schedule. Google says only that wider access will come as soon as possible after feedback and guardrail work.
What the Announced Price Does and Does Not Tell You
Google lists an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input priced 95% lower. Its Carbon footnote says the rates later become $4/$20, without an end date for the introduction. This is an announced price for future access, not proof that the Carbon API is available to everyone today. The official price note is the place to recheck it when Carbon opens.
Paul is right to look beyond model scores, but the cost of a successful task includes retries, tools, review time and errors as well as tokens. For a benchmark-based cost comparison and its limitations, use our companion analysis.
What to Do While You Wait
Use the available Gemini 4 Argon, or your current agent, to write down one task you would actually give Carbon. Include the starting material, allowed tools, review checkpoints and a pass/fail test. When Carbon arrives, run that same task in comparable conditions. Track quality, elapsed time, human corrections and total cost. For security work, test only code and systems you own or are expressly authorized to assess.
That is the useful takeaway from Paul's excitement: Carbon may widen what an agent can handle, but the first real answer will come from a reviewed deliverable, not a launch-day leaderboard.
Video Chapters
What Carbon promises
| Time | Topic |
|---|---|
| 00:00 | Paul introduces Google's new model |
| 00:24 | Long-horizon work and multi-step agents |
| 01:07 | Finding, checking and fixing vulnerabilities |
| 01:23 | The benchmark comparison with Argon and rivals |
Safety and the release pace
| Time | Topic |
|---|---|
| 01:45 | Prompt-injection results and their limits |
| 02:15 | Unsupported claims and coding shortcuts |
| 02:57 | Why a second Gemini 4 model arrived quickly |
| 03:27 | Argon, Carbon and the naming question |
Access and cost
| Time | Topic |
|---|---|
| 03:59 | Selected cyber defenders first |
| 04:24 | Paul's forecast for wider access |
| 05:17 | Introductory and later token prices |
| 05:31 | What Paul wants to test when access opens |
Sources and Limits
Checked 29 October 2026. The Paul J Lipsky video and attached transcript support his commentary and two-week prediction. Google's Carbon announcement supplies the official rollout, benchmark, safety and price claims; the Argon release confirms the earlier public date. Artificial Analysis supplies independent index context. We did not use Carbon or reproduce Google's internal tests. Availability, leaderboards and price terms can change; this is a dated release read, not a hands-on verdict.