Google moves Gemini's weakest model to the free tier; paid users get the rest
From 9 October, Google will limit Gemini app users without a subscription to the Flash-Lite model, according to an updated support document reported by 9to5Google on 3 October. Free users lose access to Flash 3.6 and Pro 3.1, which they can use today.
The change is a tier shuffle, not a shutdown. Free accounts keep a Gemini app; it just runs the smallest model on the menu. AI Plus subscribers at $4.99 a month end up limited to Flash-Lite and Flash. The Pro tier disappears from their plan too.
9to5Google reported the support document on 3 October and said the changes take effect on 9 October. The Decoder, in its 4 October write-up, laid out the full ladder. No plan gets Flash-Lite only. AI Plus gets Flash-Lite and Flash. Pro requires AI Pro at $19.99 a month or AI Ultra at $99.99 and $199.99 a month. Both outlets agree on the model numbers and prices. The disagreement, if there is one, is about motive rather than facts.
What Pro and Ultra buyers get back
Nothing is being taken from the top tiers, and one thing is being added. Per 9to5Google, AI Pro customers will get the Deep Think option, described as "maximum parallel reasoning." Until now it sat behind the $99.99 and $199.99 plans. The same report says the Gemini app looks set to add "low," "medium" and "high" effort levels "for each available model" in October. Those replace the current single Extended thinking toggle. Heavier effort will eat more of a user's usage limit.
Where Gemini 4 Argon lands is an open question. 9to5Google noted that Google announced Argon first for AI Ultra subscribers. The company has not said whether it counts as a Pro model or a new tier above Pro. The Decoder speculated that the restructuring may be staging ground for Argon, which it says could be costlier to run. Neither outlet has a Google statement on Argon's pricing.
The Decoder's read on the practical damage: most casual Google AI users have no idea which model they are running, and the power users who do are probably already paying.
That is a reasonable bet, and it is also the part of this story that nobody has measured. Google has not published how its Gemini user base splits across free, Plus, Pro and Ultra. Neither outlet offers a number. The $4.99 tier is the interesting one. Those subscribers were paying for Pro access and now lose it without a price change, which 9to5Google says will be explained by email.
The security angle is elsewhere
If the query is AI in cyber defence, the Gemini tier news is mostly a pricing story. The more concrete security material in this cycle comes from OpenAI. It published an account of a coordinated model-distillation campaign on 4 October and dated the earliest activity to the first week of July.
OpenAI's write-up is careful about what did not happen. The operators, it says, "did not break our encryption, compromise a database, or gain direct access to stored user conversations." Instead they manipulated model interactions. Protected reasoning, the model's internal record of working through a task, could be reproduced in visible form at scale. One technique OpenAI describes: copying encrypted reasoning from one conversation and asking a model in another conversation to decrypt and transcribe it.
The numbers in the post are specific. OpenAI says it saw high-volume spikes on 24 and 25 July. They consisted of 16,000 requests using a relevant extraction pattern from more than 4,000 users. Related prompt-pattern activity across a cluster of more than 15,000 users was fully disrupted by 28 July. OpenAI attributes "a core cluster of the activity" to individuals associated with Moonshot AI, the developer of Kimi. It states it is unclear whether all operators came from a single actor. OpenAI says it shared information with industry partners through the Frontier Model Forum. It credits independent security researchers who reported related cross-model and conversation-compaction issues through responsible disclosure.
That last detail matters more than the attribution fight. A lab confirming that outside researchers found real attack paths in its own product is a different kind of disclosure from a vendor saying it caught bad actors. OpenAI says it confirmed the paths were real before publishing.
Anthropic's quieter change
Anthropic's contribution to the week is a privacy setting, not a defence product. BleepingComputer reported on 4 October that Claude's voice features now prompt users with "Allow us to use your voice data to improve our AI models." The prompt has Allow and Not now buttons and a toggle under Settings, Privacy. It says audio recordings and voice chat data help models understand and respond to speech, and that the data can be deleted from settings.
BleepingComputer reports the setting is disabled by default. It is separate from the existing option to train on chats and Claude Code sessions. A user can share voice and withhold text, or the reverse. That separation is the detail worth noting. Consent here is granular, which makes it easier to audit and easier to forget you gave it.
Classifiers as a security tool
The most directly useful security item here is not a product launch. Jerry Gamblin says he serves on the CVE Program's Quality, Automation and Consumer working groups. He notes that nothing in his post is a CVE Program position. He ran Cloudflare's Clef models over every CVE published between 4 August and 3 October 2026, 27,489 records. He asked two questions of each description: does it establish a security impact, and how clear is it for a defender.
His results: outside the Linux kernel, 989 of 23,736 descriptions, or 4.2%, never say what an attacker could do. The kernel is its own case because its published policy leaves impact to the user. Of its 3,753 records, 3,661 state none. That takes the all-CVE figure to 4,650 of 27,489, or 16.9%. He also reports that Clef Flash alone was too strict. The larger Clef 27B model overturned 1,198 of 2,187 flagged records, 54.8%, outside the kernel. Against 129 CVEs he labelled by hand, the two-step method caught 15 of the 16 he judged to state no impact. It never flagged one of his positives.
The models are Apache 2.0, fine-tuned from Qwen 3.8 and Qwen 3.5. They run locally through Ollama's decision endpoint with weights on Hugging Face. Gamblin says everything in his run executed on a laptop with no API key. Treat the 4.2% and 16.9% as one researcher's measurement with a disclosed method, not an official CVE Program statistic. But the method is published and the data is public, which is more than most security claims in a given week offer.
Smaller signals
Two other items are worth a line each. Ideogram released version 4.5 of its image model on 4 October. It pitches artifact-free multi-turn editing at native 2K resolution with fixed per-image pricing from $0.008 to $0.22 depending on quality mode. And a paper submitted to arXiv on 1 October by Lewis Mitchell reports a filtering result. Filtering training text by a non-parametric entropy rate estimator, with no model involved, produced 42% more unique trigrams, 30% more vocabulary and 19% less repetition than logprob-based filtering. That was in a six-generation QLoRA collapse experiment on Llama-3.1-8B, with p values below 0.001.
Neither is a cyber defence story. Both point at the same underlying problem the OpenAI disclosure raises. The pipelines that train and refine models are now attack surface, whether the input is a CVE description, a scraped corpus or another model's hidden reasoning.
Sources
8- 01Gemini app limiting what models free and AI Plus users can accessEN
- 02Google's new Gemini tiers cut free users to its weakest model and lock $5/month subscribers out of ProEN
- 03Disrupting a coordinated model-distillation campaignEN
- 04Anthropic asks Claude users to share voice data for AI model trainingEN
- 05Decision Models vs. CVE DataEN
- 06Ideogram 4.5: The most precise edit modelEN
- 07No Model Required: Text Entropy Rate Filtering Mitigates Iterative Fine-Tuning CollapseEN
- 08OpenAI's latest features take direct aim at the app store modelEN
All figures and quotations in this text come from the sources listed below.
Content prepared by the editorial team with AI assistance.
Comments
0- No comments yet — be the first.