{"slug": "the-frontier-now-ships-twice-the-second-copy-is-not-for-sale", "title": "The frontier now ships twice. The second copy is not for sale.", "summary": "Anthropic, Google and OpenAI each shipped two versions of their best models between September 1 and 3, 2025 — a public tier priced at $10 per million input tokens and $50 per million output tokens for Claude Fable 5.1 and GPT-6 Astra, and a vetted tier with no published price, according to The Signal. Anthropic said its pair are \"the same model, but with different levels of safeguards,\" with Claude Mythos 5.1 limited to vetted US organizations via a two-business-day Cyber Verification Program, while Google's Gemini 3.8 Flash Cyber moves through the Fairwind Program for government agencies and critical infrastructure operators and OpenAI's GPT-6 Astra capability goes first to alpha testers through Daybreak Blue. OpenAI's new monitor can flag \"tasks in which an agent is running for an extended period\" and stop the task on the API, making access to the frontier a credential rather than a price for solo builders.", "body_md": "The Signal · Issue 6\n\n# The frontier now ships twice. The second copy is not for sale.\n\n## The short version\n\n- **Between September 1 and 3, Anthropic, Google and OpenAI each shipped their best model twice.** A public version you can buy today, and a vetted version with the same or fuller capability that you apply for. Anthropic says its pair are[**“the same model, but with different levels of safeguards.”**](https://www.anthropic.com/claude-fable-and-mythos-5-1)\n- **The public tier costs what the last tier cost.** Fable 5.1 and GPT-6 Astra both list at $10 in and $50 out per million tokens, twice Opus 5. Gemini 3.8 Flash stays cheap at $0.75 and $3.75 until the end of the year. Nobody raised the ceiling on what you can pay.\n- **The vetted tier has no price at all.** Mythos 5.1 goes to[**“a small, but growing, set of vetted organizations,”**](https://www.anthropic.com/claude/mythos) currently in the US. Gemini 3.8 Flash Cyber[**“is only available to trusted defenders”**](https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/) through a program built for governments and infrastructure operators. Astra’s sharp end goes first to alpha testers.\n- **What the gate asks for is identity.** OpenAI’s path for individuals is a government-issued ID and two hardware security keys. Anthropic ties approval to an organization ID. Google admits organizational categories. Three labs, one week, the same shape of door.\n- **The public tier is also the one that interrupts you.** OpenAI says its new monitor can flag[**“tasks in which an agent is running for an extended period”**](https://openai.com/index/path-to-astra/) and that on the API the task will stop. The long-running agent you run alone is exactly the pattern it watches.\n- **The argument: for a solo builder, the ceiling on the frontier stopped being a price this month and became a credential.** You can still buy the best public model. You cannot buy the one the benchmarks were run on. Plan the product around the door you can open.\n\n## What shipped, and what it actually means\n\nThe week of September 1 was one of the busiest frontier weeks of the year, and the coverage treated it as three launches. It was one launch, repeated.\n\nAnthropic went first. Claude Fable 5.1 is generally available on every platform at $10 per million input tokens and $50 per million output tokens. Claude Mythos 5.1 is the same weights with different safeguards, and it is available only through two trusted access programs, one for cyber defenders and one for the life sciences, both currently limited to US organizations. The Cyber Verification Program that will carry Mythos access is real and reasonably fast, with a two-business-day review, but approval is tied to an organization ID. Today it covers Opus and Sonnet; Mythos access through it is described as coming.\n\nGoogle went second. Gemini 3.8 Flash is public and, for now, inexpensive. Gemini 3.8 Flash Cyber, the version tuned to find and patch vulnerabilities, is not something you sign up for. It moves through the Fairwind Program, which names its audience plainly: government agencies and national cyber authorities, critical infrastructure operators, core technology platforms, Google Cloud customers and security partners. Participants must confine it to security staff and run multifactor authentication. There is no public API and no published price.\n\nOpenAI went last and said the most. GPT-6 Astra is the first model the company designates at the Critical cybersecurity threshold of its Preparedness Framework, meaning it can find unknown flaws and build exploits across hardened systems without a person guiding each step. The public Astra refuses that work. The capability goes to a small group of alpha testers first, then widens through Daybreak Blue. And in the small print of its own capability post, OpenAI says the results it published reflect Daybreak Blue access rather than the default configuration. The benchmark model and the store model are not the same product.\n\n## The price did not move. The gate did.\n\nLook at the stickers. Fable 5.1 and Astra both cost $10 in and $50 out. That is exactly double Opus 5, and it is the same number at two companies. Gemini 3.8 Flash costs a fraction of either, with the introductory rate ending December 31. Read as a market, the public tier priced itself in an afternoon and nobody blinked.\n\nNow look at what the second tier costs, and notice there is no number. Mythos has no checkout page at all, because the purchase is not the gate. The gate is an application scoped to an organization, in the US, reviewed by a person. Flash Cyber has no price because it has no checkout. Astra’s advanced tier has a queue.\n\nThat is the change worth naming. For two years the frontier was a budget question: pay more, get more. As of this month the top of the range is a status question. What you are asked for is a government ID, two hardware keys, an organization ID, or a category you belong to. None of those come from a credit card.\n\n## What the door asks for, lab by lab\n\nThe three gates are not identical, and the differences matter if you are one person.\n\nOpenAI’s is the most open to individuals. Daybreak’s own page says you verify with a government-issued ID, enable Advanced Account Security with two hardware security keys, and meet the program’s eligibility requirements. A solo builder doing authorized security work can, in principle, walk through that door alone. Astra’s most advanced configuration still starts with alpha testers, so the door exists and is not yet open.\n\nAnthropic’s is organizational. The verification program takes individual applicants, but approval attaches to an organization ID, and the Mythos page limits current access to US organizations. A one-person company is an organization; a personal account is not.\n\nGoogle’s is categorical. Fairwind admits kinds of institutions. There is no individual path described, and the program’s stated purpose is to give infrastructure operators an adaptation window before attackers catch up.\n\n## The part that hits the solo builder first\n\nIt would be easy to file this under “cyber models, not my problem.” Two details argue otherwise.\n\nThe first is the monitor. OpenAI states that its production safeguards for Astra can flag legitimate work as potential misuse, including work “that does not appear directly related to cybersecurity” and tasks where an agent runs for a long time. In ChatGPT or Codex you get asked to review. On the API, the task stops. The API is where a solo builder runs the overnight agent, the batch job, the thing that works while you sleep. The public tier is the tier that watches that pattern hardest, because it has to.\n\nThe second is what “same weights” implies. The version doing the impressive work in the launch posts is the one with fewer refusals. When a vendor benchmarks the vetted tier and sells the public one, the gap between the demo and the product is now a matter of who you are. That is a different kind of gap from the ones this publication usually prices, and it is not closed by upgrading your plan.\n\n## What to do this week\n\nDecide which door your product depends on, and say it out loud in your plan.\n\nIf you are building anything security-adjacent, a scanner, an audit tool, a patching agent, the model that does the real work is behind a gate. Apply now, before you need it. OpenAI’s individual path costs a hardware key and an afternoon; Anthropic’s wants an organization, which a registered one-person company is; Google’s is not for you yet. Do not price a product on a tier you have not been admitted to.\n\nIf you are building anything that runs long unattended jobs on the API, budget for interruption. Add resume logic, checkpoint state, and log the stops, because OpenAI says they will happen and says they will not always be about security.\n\nAnd read the launch benchmarks with the footnote in view. When a result says it was measured with special access, it is describing the model you are not being sold. The one you can buy is the one to test yourself, on your own task, at the sticker price, which is how this publication has always suggested you buy anything.\n\nOne email, when there's something worth sending\n\n## Get the research in your inbox.\n\nNo fixed schedule, no filler. You get an email when we've tested something, run the numbers, or found a tool worth your time.\n\nFree. Double opt-in, unsubscribe in one click.\n\nGot a signal we missed? [Send it in](/cdn-cgi/l/email-protection#bac9cedbc8dcc8d3dfd4defad5d1dbd4dfd6dbd4de94d9d5d785c9cfd8d0dfd9ce87e9d3ddd4dbd6)\n\n## Sources & how we researched this\n\n1. [Anthropic, \"Claude Fable 5.1 and Claude Mythos 5.1\" (Sep 1, 2026). Primary. States the two are \"the same model, but with different levels of safeguards,\" that Mythos is available only through trusted access programs, and gives the Fable API price of $10 in and $50 out per million tokens.](https://www.anthropic.com/claude-fable-and-mythos-5-1)\n2. [Anthropic, Claude Mythos page (read Sep 13, 2026). Primary for the access language: \"a small, but growing, set of vetted organizations,\" currently US organizations, via the Life Sciences and Cyber Verification Programs.](https://www.anthropic.com/claude/mythos)\n3. [Claude Help Center, \"Real-time cyber safeguards on Claude Opus and Sonnet\" (read Sep 13, 2026). Primary for how the Cyber Verification Program works today: identity verification, approval tied to an organization ID, a two-business-day review, and coverage limited to Opus and Sonnet.](https://support.claude.com/en/articles/14604842-real-time-cyber-safeguards-on-claude-opus-and-sonnet)\n4. [Google, \"Introducing Gemini 3.8 Flash and 3.8 Flash Cyber\" (Sep 2, 2026). Primary. Public 3.8 Flash pricing ($0.75 in, $3.75 out per million until Dec 31, then $1.50 and $7.50) and the line that Flash Cyber \"is only available to trusted defenders.\"](https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/)\n5. [Google, \"The Fairwind Program\" (Sep 2, 2026). Primary for who qualifies: government agencies, critical infrastructure operators, core technology platforms, Google Cloud customers and cybersecurity partners, with access restricted to security staff and multifactor authentication required.](https://blog.google/innovation-and-ai/technology/safety-security/fairwind-program/)\n6. [OpenAI, \"Path to Astra: critical capabilities and frontier safeguards\" (Sep 1, 2026). Primary. The Critical threshold definition, the alpha-tester and Daybreak Blue access path, the note that published results reflect Daybreak Blue access rather than the default configuration, and the warning that the misalignment monitor can stop long-running agent tasks on the API.](https://openai.com/index/path-to-astra/)\n7. [OpenAI, Daybreak identity verification page (read Sep 13, 2026). Primary for the individual path: a government-issued ID, Advanced Account Security with two hardware security keys, and the program eligibility requirements; organizations request access separately.](https://chatgpt.com/cyber)\n8. [OpenAI API pricing (read Sep 13, 2026). Primary for the GPT-6 Astra rate of $10 in and $50 out per million tokens; the page marks promotional pricing and its footnote names Sol, so treat the Astra figure as the current list rate.](https://openai.com/api/pricing/)\n9. [Claude pricing page (read Sep 13, 2026). Primary for the comparison rates: Opus 5 at $5 and $25, Sonnet 5 at $2 and $10, Haiku 4.5 at $1 and $5 per million tokens.](https://claude.com/pricing)\n10. [TechCrunch, \"OpenAI's Astra model is on the way\" (Sep 1, 2026). Secondary; carries OpenAI's line that \"access to its most advanced cybersecurity capabilities will be more limited.\"](https://techcrunch.com/2026/09/01/open-ais-astra-model-is-on-the-way-and-very-good-at-breaking-into-computer-systems/)\n11. [SecurityWeek, \"OpenAI widens access to cybersecurity model after Anthropic's Mythos reveal\" (Apr 16, 2026). Secondary; background on the earlier round of gated cyber models and the two labs' differing access philosophies.](https://www.securityweek.com/openai-widens-access-to-cybersecurity-model-after-anthropics-mythos-reveal/)\n\nThe Signal is Okane Land's weekly read on the AI news that changes a decision you make as a solo builder, verified to primary sources.", "url": "https://wpnews.pro/news/the-frontier-now-ships-twice-the-second-copy-is-not-for-sale", "canonical_source": "https://okaneland.com/signal/the-frontier-twice/", "published_at": "2026-09-13 00:00:00+00:00", "updated_at": "2026-09-13 12:41:00.665686+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-safety", "ai-policy", "ai-products"], "entities": ["Anthropic", "Google", "OpenAI", "Claude Fable 5.1", "Claude Mythos 5.1", "Gemini 3.8 Flash", "GPT-6 Astra", "Daybreak Blue"], "alternates": {"html": "https://wpnews.pro/news/the-frontier-now-ships-twice-the-second-copy-is-not-for-sale", "markdown": "https://wpnews.pro/news/the-frontier-now-ships-twice-the-second-copy-is-not-for-sale.md", "text": "https://wpnews.pro/news/the-frontier-now-ships-twice-the-second-copy-is-not-for-sale.txt", "jsonld": "https://wpnews.pro/news/the-frontier-now-ships-twice-the-second-copy-is-not-for-sale.jsonld"}}