Illustration by tuput
English
Google announced Gemini 4 Carbon on 28 October 2026, put it in the hands of a subset of cyber defenders first, and opened it to all paying developers and enterprises 13 days later. Anthropic and OpenAI describe tiered access for their own newest models, each with its own rules.
Google DeepMind announced Gemini 4 Carbon on 28 October 2026 and said it would make the model available that week to “a subset of cyber defenders” through its Fairwind Program. On 10 November, 13 days later, it opened Carbon to all paying developers and enterprises through the Gemini API, and to Google AI Ultra subscribers in the Gemini app, at an introductory price of $2 per million input tokens and $10 per million output tokens.
The order was deliberate, and Google is not the only lab that works this way. Anthropic and OpenAI have both published tiered access for their newest models in the past ten weeks, split by time, by safeguards or by verified customer. This article sets out what each lab published and the reason each gives. It also notes which numbers are the labs’ own. Everything is as of 12 November 2026.
What Google released on each date
Google’s announcement post called Carbon “our strongest model with state-of-the-art cybersecurity and long horizon capabilities”. It said Google was taking part in the US government’s voluntary process for pre-release model testing, and that it would “gradually phase access” to more trusted partners through Fairwind before bringing the model to developers, businesses and consumers.
The 10 November post says Carbon is available through the API in Google AI Studio and Gemini for business for all paying developers and enterprises. It is rolling out to Google AI Ultra subscribers in Google Antigravity (Google’s coding tool), the Gemini app and AI Mode in Search. A footnote says the introductory price ends on 30 November 2026, after which the price is $4 per million input tokens and $20 per million output tokens.
One part stays behind the door. The same post sends defenders to Fairwind for “Carbon for cyber”, on a rolling basis, and describes the misuse safeguards built into the model. In simulated attack environments, Google says, Carbon blocked 99% of attack trajectories from its red-teaming agent, and its per-attempt breach rate was 0.8%, against 3.1% for Gemini 4 Argon. Those are Google’s own tests.
Argon followed the same path. SecurityWeek reported on 1 October that Google had announced it the day before, first to a select group of defenders through Fairwind, and that trusted defenders and Google’s own teams would get it “without cyber guardrails”. Google’s post of 13 October then opened Argon to developers, enterprises and paid subscribers, again 13 days after the first announcement.
Fairwind, the program that gets the first look
Google launched Fairwind on 2 September 2026 as a limited-access program for governments and trusted partners. At launch it paired a model called Gemini 3.8 Flash Cyber with CodeMender, Google’s tool for finding, checking and fixing software flaws, and Google said more than 650 partners were taking part.
The post lists three groups Google was reaching first: governments and national cyber authorities, operators of critical infrastructure such as healthcare, telecommunications, energy and financial networks, and core technology platforms. Members agree to limit access to staff in their own cybersecurity, incident response or penetration-testing teams, and to use protections such as multi-factor authentication.
Google’s stated reason is that early access gives defenders an “adaptation window” to harden their systems before attackers can use new capabilities. TechRepublic reported that Google has not said how many Fairwind members can use Carbon, that membership does not guarantee Carbon access, and that organisations ask through their Google Cloud account teams.
Why a lab holds a model back from some users
Anthropic’s post on its Cyber Verification Program gives the plainest version. It says cybersecurity is “inherently dual use”: the same capabilities that let a security team find and fix a flaw can help someone else exploit it. For that reason Anthropic’s generally available models carry safeguards that block most cyber work, and verified defenders can apply for fewer blocks.
OpenAI’s system card for GPT-6 Astra puts it as a rule of procedure: “As cyber capabilities increase, our approach is to expand defensive access and safeguards together.” In each case the lab publishes a capability, and a way for checked users to reach more of it, in the same breath.
Anthropic: one model, two doors
Anthropic launched Claude Opus 5.5 on 22 September 2026 with cyber safeguards under which, the company says, most cybersecurity tasks are re-routed to an older model, Opus 4.8, while routine bug fixing in a developer’s own code goes through.
On 6 October it merged Project Glasswing, an earlier trusted-access scheme for its Claude Mythos models, into an expanded Cyber Verification Program with three tiers. Defense Access covers work such as incident response, malware analysis and vulnerability validation. Red Team Access adds authorised penetration testing and is open to organisations only. Specialized Access, with the fewest blocks, is reserved for verified organisations that test systems such as power grids, telecom networks and interbank transfer infrastructure, and Anthropic says it reviews each one in depth with the US government. (The scheme was also covered in our 7 October roundup.)
Anthropic tested its tiers on its own benchmark, CyScenarioBench, using Opus 5.5. With no program access every task was blocked at the first prompt. In the Defense tier 46 of 50 trials were blocked at some point. In the Red Team tier nothing was blocked and 34 of 50 tasks were completed, which Anthropic says equals the model’s 67.6% with no safeguards.
On 8 October Anthropic added its Cyber Mission, which includes a Critical Infrastructure Defense Program with 11 founding partners, among them Accenture, CrowdStrike, Dragos, Palo Alto Networks and Rockwell Automation, and a free, opt-in OSS Scanner for open-source projects. Anthropic says OSS Scanner reports are sent without human review and that it expects a true-positive rate above 90%.
The newest step came on 9 November, with Claude Fable 5.2 and Claude Mythos 5.2. Anthropic says they are the same underlying model. Fable has strong cybersecurity and biology safeguards and is available to all users, at $8 per million input tokens and $40 per million output tokens. Organisations can apply for Mythos through the Cyber Verification Program and a Life Sciences program. Anthropic lists the UK AI Security Institute, Frontier Design and METR among the outside testers before release. The institute had earlier published an incident report on its own cyber test runs.
OpenAI: Daybreak Blue and Daybreak Red
OpenAI published the GPT-6 Astra system card on 3 September 2026. It calls Astra the first model to reach the Critical level of cybersecurity capability under OpenAI’s Preparedness Framework. The card describes a phased approach through its Daybreak program, “beginning with a limited set of organizations and full production cyber safeguards”. CNBC reported two days earlier that OpenAI had said access to its cybersecurity capabilities would be more limited. 9to5Google, relaying OpenAI, reported that paid ChatGPT plans would follow “in the coming days”; OpenAI’s own launch post could not be read for this article, so that detail rests on 9to5Google.
On 5 November OpenAI introduced GPT-6 Cyber, which it classes as Critical as well. It is initially available through the API and Codex to a select group of approved Daybreak Red customers, with wider access to Red customers promised “soon”. OpenAI describes Daybreak Blue as for verified defenders who want to validate vulnerabilities and fixes with fewer unnecessary refusals, and Daybreak Red as for deeper exploit development and authorised testing. It also announced a $1 billion grant program, Daybreak for Frontline Defenders, to subsidise access for resource-constrained teams such as government bodies and essential-service providers.
What the numbers can and cannot support
Google reports Carbon at 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploiting them) and 71.0% on CWE-bench (patching them), measured, it says, without guardrails. It also says Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 at 59. Artificial Analysis’s own page, opened on 12 November, lists the high-effort setting at 60 and describes the model as slow per task, at 28.1 minutes against a median of 13.0 for comparable models.
OpenAI reports 48.5% on ExploitGym for GPT-6 Cyber, up from 32.1% for GPT-5.6 Cyber. Anthropic’s launch table gives 76.3% on CyScenarioBench for Mythos 5.2, and footnotes that each lab’s scores on that test rest on different tasks.
Google’s and OpenAI’s posts both use the name ExploitGym, but neither says the settings match, so the figures should not be lined up against each other. None of the pages opened for this article reports an independent replication of the cyber scores. Our piece on benchmark audits covers why leaderboard numbers need that caution.
What government testing does and does not say
Google’s posts say it takes part in “the U.S. government’s voluntary process” and do not name the agency. TechRepublic notes that taking part is not government certification or approval. The National Institute of Standards and Technology describes a Center for Advancing Innovation and Standards for Super Intelligence that will establish voluntary agreements with developers and lead unclassified evaluations, focused on risks such as cybersecurity, biosecurity and chemical weapons. That page does not mention Carbon or any other named model.
As of 12 November 2026, Fairwind applications are open on a rolling basis, GPT-6 Cyber’s wider Daybreak Red rollout has no announced date, and Google’s introductory Carbon price of $2 and $10 per million tokens runs until 30 November.
Sources & further reading
- Google: Gemini 4 Carbon, our next frontier model (28 October 2026)
- Google: Gemini 4 Carbon is now available (10 November 2026)
- Google: Gemini 4 Argon is now available (13 October 2026)
- Google: Fairwind Program, cyber defense tools for trusted partners (2 September 2026)
- Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2 (9 November 2026)
- Anthropic: Expanding the Cyber Verification Program (6 October 2026)
- Anthropic: Introducing the Anthropic Cyber Mission (8 October 2026)
- Anthropic: Introducing Claude Opus 5.5 (22 September 2026)
- OpenAI: GPT-6 Astra System Card, Trusted Access for Cyber (published 3 September 2026)
- OpenAI: GPT-6 Cyber, advancing the frontier of aligned cybersecurity agents (5 November 2026)
- CNBC: OpenAI says Astra AI model is its first that crosses 'Critical' cybersecurity capability (1 September 2026)
- 9to5Google: OpenAI launches GPT-6 Astra (3 September 2026)
- SecurityWeek: Google launches Gemini 4 Argon with guardrail-free access for vetted defenders (1 October 2026)
- TechRepublic: Google unveils Gemini 4 Carbon, who gets access first
- NIST: Center for Advancing Innovation and Standards for Super Intelligence (CAISSI)
- Artificial Analysis: Gemini 4 Carbon (High), intelligence, performance and price analysis
Researched and written with the help of AI tools and edited for accuracy. Provided for general information and discussion only, not professional advice. See our editorial standards and disclaimer. Spotted an error? Tell us.
Enjoyed this? Get the next one.
One good read at a time, straight to your inbox. No spam, unsubscribe anytime.