Skip to content
World clockEU--:--UK--:--USA--:--CN--:--PLDEFRIT中文EN

portal about AI and technologyevents · analysis · interviews · technical background

Search
LIVE
›

Google puts its weakest Gemini model behind the free tier from 9 October

Google will limit free Gemini app users to Flash-Lite from 9 October, cutting off Flash and Pro, according to a support document updated this week and reported by 9to5Google.

AI & modelsAnalysisRachel NwosuPublished: 4 October 20268 min readSources 8
Google puts its weakest Gemini model behind the free tier from 9 October

Google will stop giving free Gemini app users access to its stronger models on 9 October. From that date, anyone without a Google AI subscription gets Flash-Lite only, according to a support document updated this week that 9to5Google reported on 3 October. The same document puts AI Plus subscribers, who pay $4.99 a month, on Flash-Lite and Flash, with 3.1 Pro removed.

That is a cut, not a tidy-up. The free tier currently includes 3.6 Flash and varying access to 3.1 Pro, and the change removes both. Only AI Pro at $19.99 a month and AI Ultra at $99.99 or $199.99 a month keep all three models, per a table published by THE DECODER.

For ordinary users the practical effect is a weaker default answer. For anyone building on the free tier it is a forced upgrade or a downgrade. Google has not said how the limits interact with the compute-based usage change it introduced in May. Nor has it said when Argon, the next Gemini model, arrives for anyone outside AI Ultra.

What the tiers now buy

The structure is simple enough to print on a card. No plan: Flash-Lite only. AI Plus: Flash-Lite and Flash. AI Pro and AI Ultra: Flash-Lite, Flash and Pro. 9to5Google reports that AI Plus subscribers will get an email explaining when their changes take effect. It also reports that AI Pro customers gain the Deep Think option for what the support document calls maximum parallel reasoning, previously limited to the two Ultra plans.

Google is also preparing low, medium and high effort levels for each available model, 9to5Google reported. Higher effort means more thorough answers and a faster draw on your usage limit. That detail matters more than the model names. Google is moving from selling access to selling effort, the same lever it already exposes in AI Studio and Antigravity.

THE DECODER argues the real-world impact is likely small. Most casual users do not know which model they are running, and power users are probably paying already. That is plausible. It is also the argument every company makes when it narrows a free tier.

Why it matters beyond Google

The arrangement mirrors OpenAI's freemium split for ChatGPT, where the paid tier gets the better models. Google is following rather than inventing. The open question is where Gemini 4 Argon lands. 9to5Google notes Google said Argon comes first to AI Ultra subscribers. It has not said whether Argon counts as a Pro model or sits in a new tier above it.

Ideogram, meanwhile, went the other way on price. Its 4.5 image model is live in the Ideogram app, MCP and the API. It charges a fixed price per image with no subscription required: $0.008 for a low-quality generation or precise edit, up to $0.10 for a high-quality prompt-only generation and $0.22 when references or an edit are involved. The company says the price does not change with prompt length, reference count or output resolution, and that the model runs at native 2K. Multi-turn editing without artifact buildup is the pitch.

Two pricing philosophies in one week. Google charges more for the better model. Ideogram charges by the picture and lets you keep editing it.

The app store moves into the chat box

The more consequential shift is where the software lives. At its Dev Day on 29 September, OpenAI laid out features that together turn ChatGPT into a place to find, open and use apps, TechCrunch reported. ChatGPT, which the company says has 1.2 billion weekly users, will suggest an app mid-conversation when it thinks one would help. Users can connect and run it without leaving the chat.

An expanded plug-in architecture now supports extensions, so developers can build interactive panels that run inside the conversation. Sign in with ChatGPT lets users carry their AI allowance into third-party apps. OpenAI named 16 launch partners, including Cognition's Devin, Notion, Vercel, T3, OpenClaw and Dactyl, and says more are coming. Apps can be shared as lightweight ChatGPT sites, where colleagues sign in with their own credentials and permissions.

TechCrunch frames this as a challenge to the app store model, and the framing holds: discovery, launch and payment all move inside one interface. OpenAI has also reworked how developers submit plug-ins for review, its version of App Review. What none of this settles is whether users want their tools inside a chat window. Nor whether developers will accept another gatekeeper taking a cut of attention.

The model layer is getting crowded and cheap

Underneath the consumer tiers, the model market keeps fragmenting. Ollama 0.35.1 added Cloudflare's Clef and Clef Flash, as Jerry Gamblin documented. These are decision models that return choices, probabilities and scores rather than prose. Clef is a 27B model fine-tuned from Qwen 3.8 and Clef Flash a 9B model fine-tuned from Qwen 3.5, both Apache 2.0, with weights on Hugging Face and a copy in the Ollama library. Gamblin ran Clef Flash over all 27,489 CVEs published in the previous 60 days on a laptop, at a median of under a second each, with no API key and no bill.

The findings are blunt. Outside the Linux kernel, 989 of 23,736 CVE descriptions, 4.2%, never say what an attacker could do. The kernel's published policy leaves impact to the user, so 3,661 of its 3,753 records state none either. That takes the all-CVE figure to 4,650 of 27,489, or 16.9%. A second pass by the larger Clef model overturned 1,198 of 2,187 flags, 54.8%, showing how much a cheap classifier gets wrong before review. Gamblin discloses that he serves on CVE Program working groups and notes the work is not a CVE Program position.

That is the shape of the new stack: small open models, run locally, doing classification work that used to need a vendor API. Cloudflare also hosts Clef on Workers AI, but the run described here happened on one laptop.

Where the multimodal money is going

Video is the next front. OpenAI's Dev Day demo showed a meeting app inside ChatGPT pulling upcoming meetings from a calendar, taking notes with AI and returning a summary of follow-up items. Figma and Adobe appeared as examples of design apps that could be worked on inside the chat. Each of those is a multimodal surface: text in, images and audio out, or the reverse.

The open-source side is moving too. Explainroo, a GitHub project from vincentsch, turns a script into a narrated MP4 using Kokoro for the voice, Whisper for word timing, Chrome for rendering and ffmpeg for the final file. Its release notes list version 0.2 on 13 October adding five more languages and video clips. Version 0.3 the same day added Japanese and Mandarin narration, 0.4 on 15 October added a camera that follows the narration, and 0.4.1 on 20 October kept arrow and chart labels whole. The project is free and open source, and the voice, timing and rendering run on the user's computer.

Those dates sit ahead of today, which is worth flagging plainly: the repository's changelog lists releases for the coming weeks. Treat the roadmap as a plan, not a shipped record.

Data, and who agrees to give it

Training these systems needs material, and the terms are shifting. Anthropic has started asking Claude users to voluntarily share voice conversations to improve its models, BleepingComputer reported. The prompt reads: Allow us to use your voice data to improve our AI models. The setting is off by default, sits under Settings then Privacy, can be switched off later, and users can delete the voice data.

Voice training is separate from the existing option to train on chats and coding sessions. A user can allow one and refuse the other. BleepingComputer reported the prompt as entirely optional. That distinction will matter as regulators look at consent for audio, which carries more identifying information than typed text.

On the security side, OpenAI said it disrupted a coordinated campaign to extract protected reasoning from its models, with the earliest activity in the first week of July. The operators did not break encryption or reach stored conversations. Instead they manipulated model interactions, including copying encrypted reasoning from one conversation and asking a model in another to decrypt and transcribe it. OpenAI says activity began on 1 July and spiked on 24 and 25 July with 16,000 requests from over 4,000 users. It involved a cluster of more than 15,000 users that OpenAI fully disrupted by 28 July. OpenAI attributes a core cluster to individuals associated with Moonshot AI, the developer of Kimi.

Three threads run through the week. Access is being metered by tier. Distribution is moving into the chat window. And the models doing the work are getting smaller, cheaper and, in some cases, open. Google's 9 October change is the one users will notice first, and the one they can do least about.

Comments 0

Sources

8
  1. 01Gemini app limiting what models free and AI Plus users can accessEN
  2. 02Google's new Gemini tiers cut free users to its weakest model and lock $5/month subscribers out of ProEN
  3. 03Ideogram 4.5: The most precise edit modelEN
  4. 04OpenAI's latest features take direct aim at the app store modelEN
  5. 05Decision Models vs. CVE DataEN
  6. 06Explainer videos and product demos made by your AI agent. Free and open sourceEN
  7. 07Anthropic asks Claude users to share voice data for AI model trainingEN
  8. 08Disrupting a coordinated model-distillation campaignEN

All figures and quotations in this text come from the sources listed below.

Content prepared by the editorial team with AI assistance.

Rachel Nwosu

Rachel Nwosu

AI, models and technology

Rachel Nwosu covers AI, models and technology for FLASH24, working from public model documentation, benchmark releases and repository histories rather than press summaries, and she skips announcements that arrive without reproducible numbers. She checks training-data claims against dataset cards and reruns reported metrics where code is available. She spends much of her week interviewing researchers and engineers, tracking model launch calendars, and comparing vendor benchmarks with independent evaluations. Outside the desk she runs 3D printers, restores old computers, and tests how models learn from internet junk. She does not publish benchmark figures she cannot trace to a source.

Newsroom →

Comments

0
  1. No comments yet — be the first.

Write a comment

Comments are public. We do not publish abuse, spam or advertising.