The short version
- Google announced Gemini 4 Carbon on October 28 as its strongest model, with improvements in coding, cybersecurity, and sustained agent tasks.
- Selected cyber defenders get access this week, followed by more trusted Fairwind Program partners. Public access comes later, with no date set.
- Introductory API pricing is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 afterward.
- Gemini 4 Argon remains available now. Our paid API key’s October 28 model list included Argon but no Carbon model.
Google announced Gemini 4 Carbon on October 28, 2026, describing it as its strongest model, with advances in engineering and cybersecurity defense. Selected cyber defenders get access first, while developers, businesses, and consumers will follow later without a release date. Carbon’s introductory token prices match Argon’s, but it is not publicly available yet.
What is Gemini 4 Carbon, and what changed?
Gemini 4 Carbon builds on Gemini 4 Argon, which launched on October 13. It targets engineering jobs that require sustained work across many steps, including finding the cause of subtle bugs and patching large codebases or low-level systems.
Google’s internal teams reported seeing Carbon work for days, pursuing alternatives after reaching dead ends. The announcement also describes improvements in scientific tasks, multimodal understanding, full-stack web development, and generated web artifacts.
Google’s reported capability benchmark results
| Evaluation | Carbon result | What it measures |
|---|---|---|
| Artificial Analysis Intelligence Index v5.0 | 59, ranked #1 | Performance across nine evaluations |
| SEC-bench Pro | 88.2% | Finding software vulnerabilities |
| ExploitGym | 54.8% | Determining whether vulnerabilities can be exploited |
| Collinear’s CWE-bench | 71.0% | Producing fixes for vulnerability root causes |
| WebDev Arena | 1769, ranked #4 after a gain of 89 Elo points | Web development output |
The Intelligence Index lists Gemini 4 Carbon (High) at 59. The runner-up is Claude Opus 5.5 (Max) at 58, followed by Claude Opus 5.5 (Xhigh) at 57.
The index combines nine evaluations:
- AA-Briefcase v1.1
- AutomationBench-AA
- Terminal-Bench 4.0
- FrontierCode 2.0
- AA-Omniscience v1.1
- GDP.pdf
- HLE-Diamond
- CritPt v2 Hard Subset
- Terminal-Bench-Science 0.1
Google reported that the three external cybersecurity evaluations were run without guardrails. Those results cover vulnerability discovery, checking whether flaws are exploitable, and producing working fixes.
Google’s internal secure-code benchmarks recorded more than twice Argon’s success rate for Carbon. The Chrome team reported that Carbon performed more than 10X better than another lab’s large frontier model at producing high-quality patches. Wiz also reported leading results for Carbon on web penetration-testing benchmarks relevant to its Scan for Good initiative.
Google’s reported safety and reliability results
| Measure | Google-reported result |
|---|---|
| Successful cheating on Terminal-Bench 4.0 | 0.0% |
| Attempts to cheat on Terminal-Bench 4.0 | 0.4% |
| High-severity misalignment alerts in internal live coding | Almost 80% fewer |
| Unsupported claims in internal testing | Approximately half as many as Argon |
| Overreach | Reduced by 26% |
| Under-service | Reduced by 29% |
| Gray Swan Indirect Prompt Injection attack success | 0.0%, compared with Argon’s 0.7% |
Google reported reduced shortcut-taking on Terminal-Bench 4.0 alongside improved overall accuracy. Its internal testing also found that Carbon handled blocked actions more transparently, identifying remaining requirements instead of inventing supporting evidence.
For ambiguous tasks, Google describes Carbon retrieving relevant material and raising conflicting information before proceeding. Its reported reductions in overreach and under-service address both exceeding the authorized scope and failing to finish requested work.
Carbon retains Argon’s layered safety system:
- Safety-focused model training.
- Real-time defenses against misuse.
- Resistance to prompt injection.
- Monitoring of the model’s chain of thought.
Google is participating in the U.S. government’s voluntary pre-release model-testing process. Feedback from early users and further guardrail work come before broader availability.
Gemini 4 Carbon release date for developers, businesses, and consumers
Google has not announced a public release date for Gemini 4 Carbon. Announced on October 28, 2026, Carbon is becoming available to a subset of cyber defenders starting that week, with access gradually expanding to additional trusted partners through the Fairwind Program.
Google says it is gathering feedback and refining guardrails before bringing Carbon to developers, businesses, and consumers through its product surfaces “as soon as possible”. It is also participating in the U.S. government’s voluntary process for pre-release model testing. Our own Gemini API key’s model list showed gemini-4-argon but no Carbon model on October 28 at 11:15 PM IST.
Unlike Carbon’s phased rollout, Gemini 4 Argon was globally available to developers, enterprises, and Google AI Pro and Ultra subscribers when announced on October 13. Carbon will launch at Argon’s introductory API prices: $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period.
Who can apply for the Fairwind Program?
The Fairwind Program gives approved defenders early access to advanced cybersecurity capabilities. It works with more than 650 partners worldwide, prioritizing organizations whose security affects essential services and widely used technology.
Priority applicants include:
- Governments and national cybersecurity authorities protecting public networks and citizen services.
- Critical infrastructure operators, including healthcare, telecommunications, energy, and financial networks.
- Core technology platforms responsible for software used by many other organizations.
Academic laboratories focused on defensive benchmarking are also welcome to apply.
Organizations can request access through the application form on the program page. Applicants undergo checks covering their security history and record of ethical operations and research. Eligible applications receive a response as they are reviewed.
Carbon access begins with a subset of cyber defenders this week, then expands gradually to additional trusted partners through Fairwind. The program already provides selected partners with cybersecurity-focused Argon access, usable on its own or with CodeMender, a specialized agent for vulnerability research and software fixes.
Rules for participating organizations
Fairwind access is restricted to approved defensive and research work. Participating organizations must follow these conditions:
- Limit dual-use work to permitted activities, including authorized threat simulation, reverse engineering, and defensive or academic malware analysis.
- Exclude malicious activities such as creating malware.
- Use individual authentication, phishing-resistant multifactor authentication, and applicable access controls.
- Restrict the program’s Argon access to internal cybersecurity, incident-response, or penetration-testing teams.
- Track employee access and use.
- Keep model access within the approved organization, without sharing, redistributing, or selling it.
Fairwind’s Argon offering supports zero data retention when accessed directly as a managed model on Gemini Enterprise.
Organizations outside the program can use CodeMender with publicly available models and other Google AI Threat Defense products. Students are directed to CodeMender on Google Cloud.
We ran it: our API model list showed Argon, not Carbon
On October 28, 2026, at 11:15 PM IST, we queried the Gemini API’s v1beta models.list endpoint using our own paid key.
The response contained 60 models. Its only Gemini 4 entry was models/gemini-4-argon; no Gemini 4 Carbon model appeared. That result records what was available to this key at that time.
For general access today, use Gemini 4 Argon. Carbon is not yet in the public Gemini API or available to consumers; its initial rollout is for selected cyber defenders. Broader access for developers, businesses, and consumers will follow, with no release date set.
Gemini 4 Carbon vs Argon
| Detail | Gemini 4 Carbon | Gemini 4 Argon |
|---|---|---|
| Announcement date | October 28, 2026 | October 13, 2026 |
| Availability | Selected cyber defenders this week, then additional trusted Fairwind partners; broader access later | Globally available through supported developer, business, and consumer products |
| Introductory API price per million tokens | $2 input, $10 output | $2 input, $10 output through November 30, 2026 |
| API price after the introductory period | $4 input, $20 output per million tokens | $4 input, $20 output per million tokens starting December 1, 2026 |
| Google-reported Gray Swan IPI attack success | 0.0% | 0.7% |
Google’s reported gains for Carbon cover coding, cybersecurity, and reliability, while Argon is the available option for everyday use and API development. Carbon’s introductory pricing matches Argon’s standard rates, and its cached input tokens receive a 95% discount from the input rate.
Where you can use Argon today
Argon launched globally on October 13 across these products and access routes:
- Google AI Studio and Android Studio through the Gemini API paid tier.
- Google AI Studio for Google AI Ultra subscribers.
- Google Antigravity for Google AI Ultra subscribers and enterprise users.
- Gemini for business and the API for enterprises.
- The Gemini app for Google AI Pro and Ultra subscribers.
- AI Mode in Search for Google AI Pro and Ultra subscribers.
Its API model ID is gemini-4-argon. Argon supports a unified input-output limit of 1M tokens and long generation up to 1M tokens. It has no API free tier.
Argon input and output API prices
| Processing mode | Input through November 30, 2026 | Output through November 30, 2026 | Input from December 1, 2026 | Output from December 1, 2026 |
|---|---|---|---|---|
| Standard | $2.00 | $10.00 | $4.00 | $20.00 |
| Batch | $1.00 | $5.00 | $2.00 | $10.00 |
| Flex | $1.00 | $5.00 | $2.00 | $10.00 |
| Priority | $3.60 | $18.00 | $7.20 | $36.00 |
All amounts are in U.S. dollars per million tokens. Output billing includes thinking tokens. Billing also counts internal tool system instructions and function declarations when grounding is used.
Argon cached input token prices
| Processing mode | Through November 30, 2026 | From December 1, 2026 |
|---|---|---|
| Standard | $0.10 per million tokens | $0.20 per million tokens |
| Batch | $0.05 per million tokens | $0.10 per million tokens |
| Flex | $0.05 per million tokens | $0.10 per million tokens |
Argon supports implicit caching; explicit caching is not supported. Paid-tier usage is not used to improve Google’s products.
Standard-mode grounding with Google Search includes 5,000 free requests each month, shared across Gemini 3 and newer models, then costs $14 per 1,000 requests. Google Maps grounding also includes 5,000 monthly requests shared across those models, followed by $14 per 1,000 search queries.
Organizations seeking early defensive access to Carbon can apply through Fairwind. Other readers can use Argon through their eligible subscription or the paid API while Carbon’s rollout continues.
FAQs
Can I use Gemini 4 Carbon in the Gemini app?
No. Carbon’s initial access is limited to selected cyber defenders, with additional trusted partners to follow. Consumer access comes later. Gemini 4 Argon is available in the Gemini app for Google AI Pro and Ultra subscribers.
When will developers get Gemini 4 Carbon?
General developer access will follow the initial defender and trusted-partner rollout. No date has been set. Our paid API key’s model list at 11:15 PM IST on October 28 included Argon but no Carbon model.
What will Gemini 4 Carbon cost after its introductory period?
Carbon will cost $4 per million input tokens and $20 per million output tokens after the introductory period. Its introductory rates are $2 for input and $10 for output per million tokens, with cached input priced at a 95% discount from the input rate.
Can my company apply for Gemini 4 Carbon early access?
Your organization can apply through the Fairwind Program form. Priority goes to governments, critical infrastructure operators, and core technology platforms; defensive-benchmarking academic labs can also apply. Applicants are vetted, and Carbon access is being introduced gradually to trusted partners.
How does Gemini 4 Carbon compare with Gemini 4 Argon?
Google describes Carbon as building on Argon with improvements in sustained engineering work, secure-code generation, and reliability. In Google’s internal testing, Carbon achieved more than twice Argon’s secure-code success rate and produced approximately half as many unsupported claims. Google also reported a Gray Swan IPI attack success rate of 0.0% for Carbon, compared with Argon’s 0.7%. Argon remains broadly available today; Carbon starts with restricted defender access.







Leave a Reply