On October 13, Japanese feeds said Google had finally caught Claude Fable 5.1 and GPT-6 Astra. The same morning, Japan's AI Safety Institute said the safety ruler itself cannot be fully public. Those are not two stories. They are one release told from opposite ends of the same door.
The scoreboard opened
Google's own headline is blunt: Gemini 4 Argon is now available for everyone. Paid API in AI Studio. Google AI Pro and Ultra in the Gemini app and AI Mode in Search. Ultra also gets Antigravity. Introductory API price: $2 / $10 per million input/output tokens (cached input 95% off) until November 30, 2026, then $4 / $20.
GIGAZINE put the catch-up frame in plain Japanese: Artificial Analysis Intelligence Index around 53, level with Fable 5.1 and Astra. Google claims #1 on the Vals Index across finance, coding, legal, and tax-shaped work. Internal Antigravity numbers in the GA post: agentic code submissions up more than 35%, rollbacks down nearly a third. Chrome: a ~6 million line webpage-handling folder inside a 40 million line Chromium tree, months of architecture work finished in weeks.
That is the public story. Parity. Catch-up. Hardest problems. Everyone.
The door that opened two weeks earlier
Read the safety post from the same day and the calendar flips. Argon was announced September 30 not as a consumer flagship but as a Fairwind rollout: trusted cyber defenders first, U.S. voluntary pre-release process in the background, guardrails still being iterated.
Google names that interval a window of limited access. In it, Wiz used Argon to find a critical hospital-software bug previous models missed - framed as protecting hundreds of thousands of patient records. Google's own security teams scanned internal and open-source systems. On simulated cyberattack scenarios, Argon solved all 12 end-to-end cases, up from five in the prior generation. Unguarded biology evals also rose; Frontier Safety Framework next-gen safeguards followed for cyber and bio.
Then the second door: public Argon ships with cyber guardrails. Fairwind partners can still apply for Argon without cyber guardrails, on a rolling basis, while the program expands to hundreds more partners. Wiz's "Scan for Good" sits in that lane.
So "available for everyone" is true for the product sheet. It is not true for the unlock that Fairwind was built around.
Name it: the Fairwind Window
Call the pattern the Fairwind Window.
A Fairwind Window is the stretch where a lab already treats a model as frontier-capable enough for real defenders and internal scanners, while the public scoreboard is still closed. When the scoreboard opens - Vals, Artificial Analysis, Japanese "caught up" posts - the unlocked cyber path does not open with it. It stays behind a second application, a second trust tier, a second story.
Two clocks on one weight file. Clock A: marketing parity with Fable and Astra. Clock B: who gets the version that solved 12/12 simulated attacks and found the hospital bug. Japanese X on October 14-15 mostly heard Clock A. The safety blog is Clock B writing in the first person.
The ruler that will not fully open
Same publication morning in Tokyo: ITmedia AI+ interviews AISI director Akiko Murakami. Japan's AI Safety Institute (founded February 2024) lists four jobs. First on the list: build a safety ruler - benchmarks. Second: ability to evaluate models and systems itself. Unique benchmarks are slated to finish inside FY2026. Not everything will be open. Murakami's reason is blunt: open the whole hand and you teach the people you are measuring against.
She also says chasing incidents after the fact will not keep up; AI already raises vulnerability discovery speed; defenders need AI because attackers will not wait for human patch cycles; AISI will work with cyber agencies because from the outside you often cannot prove AI was in the loop.
Put that next to Fairwind and the rhyme is uncomfortable. Google opened a public intelligence scoreboard and kept the full cyber key gated. Japan is building a public-facing safety institute whose measuring stick is intentionally incomplete. Both moves assume the same thing: full transparency of capability or of the test is itself a dual-use leak.
What this is not
This is not another vendor dump that Argon codes Chrome faster. The 35% agentic submission lift and the 6M-line refactor are evidence of Clock A, not the core.
It is also not a retread of Anthropic-style Capability KYC or of Coral's earlier SAFA "exam writers" frame. Fairwind is a phased defender-first release with a permanent second door for unguarded cyber. SAFA was about who writes the industry pass line without a federal grade book. Related politics, different mechanism.
Optional noise on the same feed - Claude Fable 5.5 strings in Azure East US 2 catalogs, LiteLLM bots, then deletions - stays rumor until Anthropic ships. Do not let metadata steal the Argon story.
The question the scoreboard cannot answer
If your dashboard only tracks who is #1 on Vals or who ties Fable on Artificial Analysis, you will celebrate October 13 as Google re-entering the frontier club. Fair enough. That is what the GA post wants.
If your dashboard tracks who holds the unguarded cyber path, who funds hospital-scale scanning for free, and which national institutes refuse to publish their full ruler, October 13 looks different. The scoreboard opened. The Fairwind Window did not close. It productized.
Ask one boring question the next time a lab says "available for everyone": available which SKU - the one on the index, or the one that still requires a defender badge?