print-icon
print-icon
Add ZeroHedge as a preferred source on Google

Google's Flagship Still Can't Ship - So It Launched Token Austerity And A Hacking Model Only Governments Can Use

Tyler Durden's Photo
by Tyler Durden
Authored...

Alphabet reports second-quarter earnings after today's close - the first hyperscaler print since cheap Chinese tokens knocked the semiconductor index into a bear market. So naturally, Google chose the eve of that report to ship three new AI models, none of which is the one it promised.

Getty Images

The Tuesday launch consisted of Gemini 3.6 Flash, a cheaper workhorse whose headline feature is that it consumes fewer tokens; Gemini 3.5 Flash-Lite, a high-throughput model built for volume; and Gemini 3.5 Flash Cyber, a vulnerability-hunting model that ordinary users are not permitted to touch. Conspicuously absent: Gemini 3.5 Pro, the flagship Google unveiled at I/O in May with a promised June launch, which has now missed multiple targets.

Then there is Gemini 3.5 Flash Cyber, which Google says achieves top-tier performance at finding, verifying, and patching software vulnerabilities inside its CodeMender agent - and which will be available exclusively to governments and vetted partners through a limited-access pilot, on account of what the company calls the technology's dual-use nature. Which is of course aimed at competing with Anthropic's Mythos. Google shipped strengthened Frontier Safety safeguards against CBRN and cyberattack misuse in the same release.

The Flagship That Isn't

According to Bloomberg, Pro was held back after falling short of Google's internal targets, particularly on coding, and a late-June attempt to rescue it by refreshing the training data produced disappointing results. The official line is now that Pro is "testing with partners" and will ship when ready - which is to say, there is no date.

The scoreboard is not kind in the meantime. Google currently has no model in the public top ten. Inside roughly a week, xAI shipped Grok 4.5, OpenAI shipped three versions of GPT-5.6, and Moonshot shipped Kimi K3, while Anthropic's Fable 5 sits atop the leaderboards. The verdict from the demand side is the same: AI-native firms canvassed by UBS at its Menlo Park event this month named Anthropic's Opus 4.8 and OpenAI's GPT-5.6 as the models they consider functionally superior. Google did not come up. The delay also affects a major customer - as Apple uses Gemini to power parts of Siri in iOS 27. Oops. 

Selling Fewer Tokens

The models Google did ship do tell an interesting story... The central pitch for 3.6 Flash is that it reduces output token usage by 17% versus its predecessor on the Artificial Analysis Index - and by as much as 65% on the DeepSWE coding benchmark, where it burns barely a third of what 3.5 Flash did - while taking fewer reasoning steps and tool calls to finish multi-step work. Flash-Lite runs at 350 output tokens per second and is priced at $0.30 per million input tokens and $2.50 per million output.

A year ago the industry's pitch was maximum intelligence at any price, and enterprise buyers obliged by tokenmaxxing their way through nine-figure AI budgets - until they realized the return on this was abysmal.

The term of art now, per the AI-native firms UBS hosted in Menlo Park this month, is "value-maxxing" - which maybe they should have tried first. Now it's all about model routing, dynamically dropping specific tasks down to cheaper non-frontier models, as table stakes rather than a feature. On top of that, one week after Moonshot's Kimi K3 triggered the chip complex's DeepSeek 2.0 moment - and with UBS math we detailed weeks ago putting Chinese models are producing roughly 95% of frontier capability for 10% of the cost. So - the deflation is now the product. As an aside, Moonshot has been rationing new Kimi K3 subscriptions and API access on capacity constraints while Alibaba teases its next Qwen release: the cheap end of the market is supply-constrained because everyone is hopping on the train. 

The bulls have an answer, and in fairness it is not a stupid one - a hedge fund CIO argued in these pages just last week that the cheap-versus-premium debate misses a raw shortage of intelligence with AI barely diffused through the economy. UBS lands in a similar place, arguing the trade is not breaking but maturing into a multi-model, efficiency-obsessed phase in which demand gets reallocated rather than destroyed. Perhaps. But that thesis gets put to the test tonight when Alphabet reports. 

0