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

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

Search
LIVE
›

Google Moves Gemini's Cheapest Tier to Its Weakest Model as Argon Looms

Google will cut free Gemini users to the Flash-Lite model and pull Pro from the $4.99 AI Plus plan, with the change taking effect on 9 October, according to an updated Google support document seen by 9to5Google.

AI & modelsAnalysisRachel NwosuPublished: 4 October 20265 min readSources 9
Google Moves Gemini's Cheapest Tier to Its Weakest Model as Argon Looms

The change lands on 9 October. From that date, anyone using the Gemini app without a paid Google AI plan gets Flash-Lite only. The 3.6 Flash and 3.1 Pro models disappear from the free tier, 9to5Google reported on 3 October.

What changes for the reader is simple: the model behind the answer gets smaller unless you pay. AI Plus subscribers at $4.99 a month keep Flash-Lite and Flash but lose Pro. Only AI Pro at $19.99 and AI Ultra at $99.99 or $199.99 keep all three, according to a table published by The Decoder. Google has not said how many users sit on each tier, so the size of the demotion is unknown. The tiers now map almost exactly onto OpenAI's ChatGPT freemium split, where the good models sit behind the subscription wall. For buyers, the practical question is whether the cheaper plan still buys a model worth paying for.

A backward step for the free tier

The cut is a regression, not a reshuffle. Until now the free tier offered access to 3.6 Flash with varying access to 3.1 Pro, The Decoder reported. Now the same users get the smallest model and nothing else.

Google frames the move through a support page rather than an announcement, and 9to5Google notes that Plus subscribers will get an email explaining when the limits bite. The Decoder's read is that casual users will barely notice, since few know which model answers them. Power users, it argues, already pay. That may be true. It still leaves the cheapest paying customers with less than they had before. Neither company has said how much the shift is worth to Google, and no user numbers have been published for any tier.

AI Pro does pick up something in return. 9to5Google reports that $19.99 customers get access to a Deep Think option for maximum parallel reasoning, previously limited to the $99.99 and $199.99 plans. The Gemini app also looks set to add low, medium and high effort levels for each model, which would raise answer depth at the cost of usage limits.

Where the next model fits is the open question. Google said earlier this week that Gemini 4 Argon comes first to AI Ultra subscribers, and it has not specified whether Argon counts as a Pro model or sits in a new tier above it, according to 9to5Google. The Decoder suggests the tiering may be groundwork for Argon, which could be more expensive to run.

Reasoning gets its own price list

Google is not alone in selling access to thinking time. Ideogram pushed out version 4.5 of its image model on 4 October, with four quality modes and fixed per-image pricing from $0.008 to $0.22. Its pitch is that repeated edits no longer pile up artifacts, so multi-turn editing works at native 2K resolution. It is a different product, but the pattern is the same: capability is metered, and the meter is the product.

The clearest test of whether reasoning models earn their keep is happening in security data. Jerry Gamblin ran Cloudflare's Clef and Clef Flash, two decision models that return typed answers instead of prose, against all 27,489 CVEs published between 4 August and 3 October 2026. Outside the Linux kernel, 989 of 23,736 descriptions, or 4.2%, never say what an attacker could do, he wrote on 4 October.

The kernel is its own case: its published policy leaves impact to the user, so 3,661 of its 3,753 records don't state one either.

That takes the all-CVE figure to 4,650 of 27,489, or 16.9%. Clef Flash, a 9B model, answered at a median of under a second per CVE on a laptop with no API key. The 27B Clef re-read the flags and overturned 1,198 of 2,187, or 54.8%, which is a useful reminder that a small model's judgment needs a second pass.

Benchmarks meet their limits

Reasoning quality is also being probed from the model side. OpenAI published an account on 4 October of a coordinated campaign to extract protected reasoning from its models, with the earliest activity in the first week of July. Operators copied encrypted reasoning from one conversation and asked a model in another to decrypt and transcribe it, the company says. OpenAI attributes a core cluster to individuals associated with Moonshot AI, the developer of Kimi, and says it disrupted more than 15,000 users by 28 July.

The lesson for anyone benchmarking reasoning is that the scratchpad is an asset worth stealing. Extracted reasoning could train another model without carrying over the original's safeguards, OpenAI argues, which is why the company shared details through the Frontier Model Forum.

Research keeps pointing at the same fragility. A paper by Lewis Mitchell on arXiv, submitted on 1 October, tested filtering synthetic training text by entropy rate rather than model log-probabilities. In a six-generation QLoRA collapse experiment on Llama-3.1-8B, logprob-based filtering gave no significant text-diversity benefit on any metric (p > 0.23). The entropy filter delivered 42% more unique trigrams, 30% more vocabulary and 19% less repetition, all at p < 0.001.

Then there is the shape problem. Alex L. Zhang argued on 4 October that language models have kept the same text-in, text-out form since ChatGPT, and that most agent work is spent building a harness around it. He points to Jev, a model from Typesafe AI whose output is constrained to a probability between 0 and 1, making it fast on a narrow class of questions. A smaller model that answers one question cheaply is a benchmark result in itself.

Meanwhile Anthropic is asking Claude users to hand over voice recordings for training. The prompt appears in the voice features and is off by default, BleepingComputer reported on 4 October, with a separate toggle in Settings, Privacy. It is optional, deletable, and entirely separate from the chat-training setting.

None of this settles which model reasons best. It does show where the money and the risk are going: behind the paywall, into the scratchpad, and into the data that trains the next one.

Comments 0

Sources

9
  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. 04Decision Models vs. CVE DataEN
  5. 05Disrupting a coordinated model-distillation campaignEN
  6. 06No Model Required: Text Entropy Rate Filtering Mitigates Iterative Fine-Tuning CollapseEN
  7. 07Language Model "Shape"EN
  8. 08Anthropic asks Claude users to share voice data for AI model trainingEN
  9. 09OpenAI'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.

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.