Google puts its newest Gemini models behind paid tiers as Argon waits in the wings
From 9 October, Google will limit personal Gemini accounts without a paid plan to the (3.5) Flash-Lite model, according to an updated Google support document reported by 9to5Google on 3 October. Free users lose access to (3.6) Flash and (3.1) Pro.

The change lands on 9 October. Google's support document, updated this week, sets out three access levels. No plan gets Flash-Lite only. AI Plus at $4.99 a month gets Flash-Lite and Flash. Pro and Ultra tiers keep all three models. 9to5Google reported the document's contents on 3 October.
The cheapest paid plan loses Pro, and the free tier drops to one model. Google's own table shows the split plainly. Flash-Lite is available everywhere, Flash from AI Plus upward, and Pro only from AI Pro at $19.99 a month. AI Plus subscribers will get an email explaining when the change takes effect, according to 9to5Google.
Why it matters: anyone using Gemini without paying now gets the smallest model, and $5 subscribers lose the top one. Power users, a small slice of the base, mostly pay already. The day-to-day impact on most users is likely modest. What remains unclear is how Gemini 4 Argon fits. Google said this week it arrives first for AI Ultra subscribers. It has not said whether Argon counts as a Pro model or sits in a new tier above it, 9to5Google reported.
The Decoder, writing on 4 October, framed the move as a mirror of OpenAI's freemium model with ChatGPT. It noted that Google may be preparing for Argon to be costlier to run than current models. Google's support page carries the tier table itself.
Google is not standing still on features for the people who do pay. 9to5Google reported that AI Pro customers will gain access to a Deep Think option for "maximum parallel reasoning". Until now that option was limited to the $99.99 and $199.99 a month plans. The Gemini app also looks set to add low, medium and high effort levels for each available model. That would raise answer quality while consuming more of a user's allowance.
None of this settles the question the announcement raises. A $4.99 subscriber who joined for Pro now pays for Flash. Whether that customer stays depends on Argon's price and placement, and Google has not published either.
The rest of the week: image editing, voice data and a distillation fight
Away from subscriptions, the week's releases ran in three directions: image editing, model access policy, and security.
Ideogram shipped version 4.5 of its image model on 4 October. It pitches the model at multi-turn editing rather than one-shot generation. The company says the model eliminates artifact buildup across successive edits, the usual failure mode when an image is revised repeatedly. It runs at native 2K resolution in four quality modes across the Ideogram app, MCP and the API. Pricing is per image and flat. Prompt-only generation costs $0.03 at low quality, $0.06 at medium and $0.10 at high. High-quality work with reference images or a precise edit rises to $0.22. Ideogram says the price does not change with prompt length, reference count or output resolution, and no subscription is required.
Anthropic went the other way, asking Claude users to hand over data. BleepingComputer reported on 4 October that a prompt now appears when people use Claude's voice features. It asks permission to use voice recordings for model training. The setting is off by default and sits separately from the existing option to train on chats and coding sessions. Users can switch it off or delete the data later, per the report.
OpenAI used its own blog on 4 October to describe a coordinated campaign to extract protected reasoning from its models. It dates the activity to the first week of July. The operators did not break encryption or reach stored conversations, OpenAI says. Instead they manipulated model interactions. That included copying encrypted reasoning from one conversation and asking a model in another to decrypt and transcribe it. OpenAI says it saw high-volume spikes on 24 and 25 July of 16,000 requests using a relevant extraction pattern from over 4,000 users. It fully disrupted a cluster of more than 15,000 users by 28 July. It attributes a core cluster of the activity to individuals associated with Moonshot AI, the developer of Kimi. It says it is unclear whether all operators came from a single actor.
The company said it shared information through the Frontier Model Forum. Independent researchers also reported related cross-model and conversation-compaction vulnerabilities through responsible disclosure.
For developers, the more consequential shift may be where apps live. TechCrunch reported on 29 September that OpenAI's Dev Day announcements, taken together, turn ChatGPT into a place where software is discovered, launched and used. ChatGPT, which the company says has 1.2 billion weekly users, will suggest apps inside a conversation when one could help finish a task. An expanded plug-in architecture now supports extensions that run interactive panels inside the chat. "Sign in with ChatGPT" lets users carry their AI allowance into third-party apps, with 16 launch partners including Cognition's Devin, Notion and Vercel.
Two smaller threads rounded out the week. Jerry Gamblin, who serves on CVE Program working groups, ran Cloudflare's Clef models locally over all 27,489 CVEs published in the previous 60 days. He found 989 of 23,736 non-kernel descriptions never state a security impact, about one in 24. Including kernel records, which leave impact to the user by policy, the figure reaches 4,650 of 27,489. And on arXiv, Lewis Mitchell's paper on text entropy rate filtering reports a result. Filtering synthetic training text by entropy rate, with no model involved, gave +42% unique trigrams and -19% repetition in a six-generation QLoRA collapse experiment on Llama-3.1-8B.
For Gemini users, the calendar is short. The model limits take effect on 9 October, and Google has said nothing about where Argon lands.
Sources
9- 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
- 03Ideogram 4.5: The most precise edit modelEN
- 04Anthropic asks Claude users to share voice data for AI model trainingEN
- 05Disrupting a coordinated model-distillation campaignEN
- 06OpenAI's latest features take direct aim at the app store modelEN
- 07Decision Models vs. CVE DataEN
- 08No Model Required: Text Entropy Rate Filtering Mitigates Iterative Fine-Tuning CollapseEN
- 09Mellon Foundation Releases the First National Study of Community-Based ArchivesEN
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.