Google moves Gemini's free tier to Flash-Lite as model access becomes the paywall
From 9 October, Google accounts without an AI subscription will be limited to Gemini 3.5 Flash-Lite, losing access to 3.6 Flash and 3.1 Pro, according to an updated Google support document reported by 9to5Google on 3 October.
From 9 October, Google accounts without an AI subscription will be limited to Gemini 3.5 Flash-Lite, losing access to 3.6 Flash and 3.1 Pro, according to an updated Google support document reported by 9to5Google on 3 October.
AI Plus subscribers at $4.99 a month keep Flash-Lite and Flash, but lose Pro. The Decoder reported on 4 October that the free tier had previously offered 3.6 Flash and varying access to 3.1 Pro. This is a cut, not a reorganisation. Only AI Pro at $19.99 a month and AI Ultra at $99.99 or $199.99 a month keep all three models, per the same support page. Google is also pushing Deep Think, currently limited to the Ultra plans, down to AI Pro subscribers, 9to5Google reported. AI Plus users will get an email explaining when the change takes effect.
It is a pricing story dressed as a product story. Google is not shipping a new flagship here. It is moving the frontier behind a card on file, while its consumer base is still learning which model names mean anything.
What the benchmark chatter misses
The week's headlines have been crowded with model comparisons: Gemini 4 Argon against GPT-6 Astra against Claude Fable 5.5. Leaked benchmarks and speculative spec sheets are circulating on aggregator sites. None of that is a release. Neither report gives benchmark numbers for any unreleased model. 9to5Google notes it is not clear where Gemini 4 Argon will sit, whether as a Pro model or in a tier above. Google has said Argon will come first to AI Ultra subscribers.
The Decoder suggests the tiering may be preparation for exactly that: a more expensive model to run, and therefore a higher wall to climb before casual users reach it. That is a reasonable read, but Google has not confirmed it. The support document does not mention Argon at all.
In practice, the number of people who notice may be small. Most casual users do not know which model answers their prompt. Power users are likely already paying. The measurable effect lands on the $4.99 cohort, who were sold Pro access and are now being told to upgrade or lose it.
Elsewhere, the release cycle keeps turning
Ideogram pushed out version 4.5 of its image model on 4 October. It claims multi-turn editing without the artifact buildup that normally accumulates across passes. The company's own pricing page lists per-image rates from $0.03 at low quality to $0.10 for prompt-only generation at high quality. It also lists $0.22 for generate-with-references or precise edit at high quality, with no subscription required. Those are vendor figures and no independent evaluation is public.
On the open-weights side, researcher Jerry Gamblin published an evaluation on 4 October. It ran Cloudflare's Clef and Clef Flash decision models locally over all 27,489 CVEs published in the previous 60 days. His finding: outside the Linux kernel, 989 of 23,736 descriptions, 4.2%, never state a security impact. He discloses that he sits on CVE Program working groups and that the result is not a CVE Program position.
A separate arXiv paper submitted on 1 October tests a filter for model collapse during repeated fine-tuning on synthetic data. In a six-generation QLoRA run on Llama-3.1-8B, the author reports a result. Entropy-rate filtering based on raw text produced 42% more unique trigrams and 19% less repetition than an established logprob-based baseline, which showed no significant benefit. It is one paper, one model family, and it has not been replicated.
Anthropic, meanwhile, has begun prompting Claude users to let it use voice recordings and voice chat data for model training, BleepingComputer reported on 4 October. The setting is off by default. It sits separately from the chat-training option, and users can switch it off or delete the data later.
The part that actually matters
Two things are happening at once. Labs are releasing smaller, cheaper, more specialised models at a steady clip. The biggest consumer platform is narrowing what a free account can reach. Ideogram's flat per-image pricing and Cloudflare's Apache 2.0 weights both point toward cheaper access at the edges.
Google is going the other way, and it can, because distribution beats benchmarks in a market where most users never choose a model at all. Whether Gemini 4 Argon justifies the new wall is a question the support document does not answer. Until Google publishes numbers, the tier table is the only concrete release in this story.
Sources
12- 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
- 04Decision Models vs. CVE DataEN
- 05No Model Required: Text Entropy Rate Filtering Mitigates Iterative Fine-Tuning CollapseEN
- 06Anthropic asks Claude users to share voice data for AI model trainingEN
- 07OpenAI's latest features take direct aim at the app store modelEN
- 08Language Model "Shape"EN
- 09China's space plane appears to have released a mystery object in orbitEN
- 10Mellon Foundation Releases the First National Study of Community-Based ArchivesEN
- 11What happens when an AI model is put in a "pain" state?EN
- 12FF7 Revelation director says he "did his best" to fight for a full disc releaseEN
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.