Google Ships Three New Gemini Models — But the One Everyone Wanted Is Still Missing
Google's Gemini team introduced three new models Tuesday — 3.6 Flash, 3.5 Flash-Lite, and a specialized security variant called 3.5 Flash Cyber — in a release the company framed as a breakthrough in efficiency for AI agents. But the announcement arrived shadowed by an absence Google could not paper over: Gemini 3.5 Pro, the flagship model promised at May's Google I/O conference, remains stuck in "partner testing" more than a month past its original June target, with no public release date.
"Developers and customers building production AI agents need higher token efficiency, lower latency, and more reliable performance," wrote Tulsee Doshi, Senior Director of Product Management for Gemini, in the announcement.
The Numbers
Google's headline claims are substantial. 3.6 Flash, priced at $1.50 per million input tokens and $7.50 per million output tokens, is said to use 17% fewer output tokens than 3.5 Flash on the Artificial Analysis Index — and up to 65% fewer on the DeepSWE coding benchmark. It posted gains on MLE Bench (63.9% vs. 49.7%) and OSWorld-Verified (83.0% vs. 78.4%), and now includes computer use as a built-in tool.
3.5 Flash-Lite, at $0.30/$2.50 per million tokens, is pitched as the fastest model in its class — 350 output tokens per second — and reportedly outperforms even the older 3 Flash on SWE-Bench Pro and OSWorld-Verified. 3.5 Flash Cyber, meanwhile, is being restricted to a "limited-access pilot" for governments and trusted partners through Google's CodeMender vulnerability-patching agent, reflecting what the company called the "dual-use nature" of cybersecurity-focused AI.
Customers cited in the release — including Figma, Harvey, Hebbia, JetBrains, Ashler, Palo Alto Networks, and Ramp — offered favorable testimonials on cost and throughput.
A Colder Reception in Practice
Independent user assessments told a more complicated story. Praise centered narrowly on speed, cost-efficiency, and batch-mode reliability. But recurring complaints described a quality regression: the models were seen as "dumber" than 3.5 Flash in everyday use, prone to repetition, weak at creative writing, and unreliable on coding and HTML tasks. Users also flagged prompt-adherence failures — instances of the model describing an action rather than executing it — along with inconsistency and hallucination. The overall sentiment among many testers was that the release felt like a minor iteration rather than a meaningful leap, deepening a sense that Google was falling behind rivals.
Death in the Crib
The deeper story is what didn't ship. Gemini 3.5 Pro was announced in May with promises of major reasoning and coding improvements "next month." That slipped from June to July and now sits with no firm date, even as Google publicly redirects attention to Flash-tier releases and confirms it has begun pre-training for Gemini 4.
Reports citing sources inside Google and DeepMind describe a model that fell short of internal targets — particularly on coding and long-horizon agentic tasks, considered critical for enterprise adoption. A late-June effort to retrain on improved coding data reportedly produced disappointing results, prompting further delay and internal frustration among researchers who fear Google is losing ground in the frontier race. Alphabet shares fell roughly 4% following reports of the delay, and speculation spread that Google had quietly pivoted toward iterating on Flash models while regrouping — or leapfrogging — toward Gemini 4.
The Analysts' Verdict
In the house assessment, analysts at CTOL Digital Solutions argued the episode points to something structural. With Chinese labs — Kimi, GLM, and DeepSeek among them — now fielding open-weight models competitive with frontier systems like Fable 5 and GPT-5.6 Sol, they contend large language models are becoming commoditized. A company with Google's resources failing to ship a frontier-class Pro model, in their view, "reflects deeper issues on talents, management styles and probably the DNA."
The firm's broader thesis: Google's ecosystem advantage will hold for everyday consumer chat — where Flash-tier models suffice, provided Google gives them away free — but will erode wherever frontier capability actually matters. CTOL's conclusion urges investors to treat legacy technology companies, as distinct from platform companies, as prime short candidates in the period ahead. That is one analytical firm's interpretation, not a settled market consensus, and readers weighing investment decisions should treat it as one viewpoint among many rather than a recommendation.
For now, Google says 3.5 Pro remains in testing and will ship "as soon as it's ready."
not investment advice