Google shipped its third Flash model in six weeks, kept the introductory price, and printed the date it doubles — with a gated 'Flash Cyber' variant for vetted defenders (Google, 2 Sep 2026)

Why it matters (spine). Three facts from one page, and the third is the one the cabinet has no precedent for. The cadence fact: a frontier lab now ships a new workhorse model every few weeks, which is the empirical condition the perishable-frontier-model-shelf-life-weeks thesis (held at signal, confidence 0.6) rests on. The gating fact: the variant Google calls "our most capable cybersecurity model" is not the one anyone can buy, which is a third variant of the limits the other two labs' pages describe on their strongest cyber capability in the same week. The pricing fact: the vendor has printed, in a footnote, a date on which the price of this model doubles. The cabinet's inference-cost card, inference-cost-collapse, records a ">280-fold reduction in ~18 months" in the price of a fixed capability. A scheduled list-price increase does not contradict that — a newer, cheaper model will presumably exist by January — but it is a vendor putting a floor under one product's price in writing, and that has not appeared on any card here before.

What the post says. Bylined Tulsee Doshi and Raluca Ada Popa, dated "Sep 02, 2026" (page metadata: published_time: 2026-09-02T15:00:00+00:00). Opening: "Building on the momentum of 3.7 Flash from three weeks ago and marking our third Flash release in only six weeks, today we're introducing Gemini 3.8, our best reasoning and coding model yet, at the same speed and low cost of 3.7." On price: 3.8 Flash "is available at the same introductory price as 3.7 Flash at $0.75 per million input tokens and $3.75 per million output tokens," with footnote 1 reading in full: "Introductory price expires on December 31, 2026. Starting January 1, 2027, $1.50/1M input tokens and $7.50/1M output tokens will apply." On the second variant: "Gemini 3.8 Flash Cyber: our most capable cybersecurity model with frontier-level performance in vulnerability detection and automated patching, available to trusted defenders through our new Fairwind Program." On the gating rationale: "3.8 Flash Cyber ships with a more permissive set of mitigations for cybersecurity, and as such, is only available to trusted defenders who require a more comprehensive set of cyber capabilities," with access described as going to "trusted government authorities, as well as critical infrastructure operators and software maintainers." On how the gains were made: "further accelerated by long-running agentic loops designed to recursively evaluate and refine the underlying models," and "rigorous training in the highly demanding domain of cybersecurity." On the model's behaviour: "3.8 Flash works harder. On complex tasks, it exhibits greater diligence — executing extra reasoning steps, and calling tools iteratively. At times, the model might use more tokens to maximize performance."

The vendor's own benchmark claims, as printed. For Flash: "3.8 Flash also achieves a 54.9% on HLE-Verified." For Flash Cyber on CWE-Bench, the external patching benchmark the page names: "On this benchmark, Gemini 3.8 Flash Cyber is on the Pareto frontier: with a pass@1 of 47.2% compared to a leading frontier model at 47.8%, yet offered at a significantly lower cost." On an internal benchmark: "a wide range of vulnerabilities across complex codebases spanning 20 programming languages ... a success rate exceeding 70%." Three deployment claims, all from Google or its partners and unverified here: "The Chrome Security team found that 3.8 Flash Cyber produced 2.6 times more correct patches to vulnerabilities in Chrome than the best commercial models that are much larger"; "Wiz found that Gemini 3.8 Flash Cyber achieves +7.5-9.7% higher recall on their internal penetration testing benchmark for a 2.3-5.2x lower cost"; and a Google Cloud team "leveraged the 3.8 Flash Cyber model to find a critical foundational vulnerability in less than 2 hours." Every one of these is the vendor's or a named partner's figure on a benchmark the vendor chose; this card repeats them as claims and vouches for none.

Derived figures, enumerated. Two derivations appear on this card: the price step is a doubling (1.50 ÷ 0.75 = 2.0; 7.50 ÷ 3.75 = 2.0), and the introductory window from this post to the printed expiry is 121 days inclusive of both endpoints, 120 elapsed (2 September to 31 December 2026). Every other figure on this card is either read off the page or, where it belongs to another card (the 0.6 confidence, the ">280-fold" and "~18 months" figures, the cross-linked dates), quoted from that card and named as such.

The cross-links that complicate this card. inference-cost-collapse (Stanford HAI, April 2025) carries the falsifier "Quality-normalized frontier inference cost stops falling or reverses on a capability-equivalent basis." This card does not close or trigger that limb: a list price for one model stepping up is not a capability-equivalent series reversing, and the same page says the older 3.7 Flash "remains fully supported for efficiency-first workloads." What the card adds is a dated test — on 1 January 2027, is the cheapest model at 3.8-Flash capability priced above or below $0.75/$3.75? perishable-frontier-model-shelf-life-weeks (20 July 2026) carries the falsifier "Frontier release cadence slows to a multi-month rhythm, or switching models proves expensive enough that single-vendor lock-in remains rational"; the "three Flash releases in six weeks" sentence bears on the first limb and points away from it, and this card records that without closing the limb, since one vendor's cadence over six weeks is not the frontier's. In the direction that complicates the gating story: the page's own statement that the model "might use more tokens to maximize performance" means the quoted per-token price understates per-task cost by an amount the page does not give.

The pattern, stated once. The same week's OpenAI card, openai-gpt-6-astra-critical-cyber-tier, quotes this page and Anthropic's for the observation that all three US frontier labs describe a limit on their strongest cyber capability. This card contributes the Google leg of that observation and nothing more; the pattern's falsifier lives on that card.

Money angle. The immediate number is the one in the footnote: a garden that builds on 3.8 Flash at the introductory price will pay twice as much per token from 1 January 2027 unless it moves. The strategic number is the cadence: three releases in six weeks makes any per-model integration a depreciating asset, which is the perishable-frontier-model-shelf-life-weeks thesis stated as an invoice. Metacrisis lever: modest — the defensive-security capability is gated to institutions, and the page's "trusted government authorities ... critical infrastructure operators and software maintainers" is the list of who gets it.

What would prove it wrong (falsifier). Three limbs.

• The facts fail if the post at the URL above does not carry the 2 September 2026 date, the "third Flash release in only six weeks" sentence, the $0.75/$3.75 price, footnote 1 as quoted, or the Fairwind gating sentences.
• The price step fails if Google withdraws or postpones the 1 January 2027 increase, or if 3.8 Flash is superseded before then such that the step never applies to a live model — in which case the card's watch item resolves as "the collapse absorbed it," which is a result worth logging, not a defeat.
• The gating fails if Flash Cyber becomes generally available without vetting, or if its capability is shown to be reproducible from the public 3.8 Flash by a documented method, which would make the Fairwind tier a formality.

Garden-code note (L41, flagged not resolved). This is a machine-layer card. TKG carries two live meanings in this repo, so, as on the nine earlier abstaining cards, TKG is left out of gardens and the machine-layer link is carried in domains. The founder's call on the codes stands open.

Provenance (what I actually read). One first-party page, fetched from the main thread on 4 September 2026: blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/, headlined "Introducing Gemini 3.8 Flash and 3.8 Flash Cyber," dated "Sep 02, 2026" on its face and 2026-09-02T15:00:00+00:00 in its metadata. Every quotation and figure above comes from that page — enumerated: the opening sentence, the price sentence and footnote 1 in full, the Flash Cyber bullet, the "more permissive set of mitigations" sentence, the access-list phrase, the two how-gains-were-made phrases, the "works harder" passage, the HLE-Verified sentence, the CWE-Bench sentence, the 20-languages sentence, the three deployment claims, and the "remains fully supported" clause. The Fairwind Program pages at deepmind.google/fairwind-program/ and the linked Google safety post, the Gemini API pricing page, and the third-party benchmark leaderboards named on the page were not opened; nothing on this card comes from them or from trade coverage.