RTX 3070 Local AI Leads One User to Drop ChatGPT Plus, Not Cloud AI

Writing for XDA Developers, Nolen Jonker describes paying for Claude, ChatGPT, and Gemini simultaneously, then shifting enough everyday work onto an RTX 3070 to make the ChatGPT subscription difficult to justify. Claude retained a role in demanding creative and technical work; Gemini remained useful for research and studying. ChatGPT was left competing with local models for the tasks in between.

The important distinction is subscription redundancy, not model supremacy. Jonker’s account is a personal workflow assessment, not a controlled benchmark or proof that every 8GB PC can replace a premium AI service.

Three subscriptions, but only two distinct jobs​

Jonker began by asking each assistant to summarize how they used it. The resulting division of labor was revealing:

• Claude: Research, design, coding, Obsidian-vault work through Claude Code, and computer-administration tasks through Cowork.
• Gemini: Studying, Deep Research, Canvas, and expanded access to the notebook research product formerly called NotebookLM.
• ChatGPT: Experimenting with newer models, occasional assistance, and an integrated Codex development workflow that remained difficult to replace.

Those are Jonker’s reported uses and preferences, not a ranking of the services’ overall capabilities. Someone whose daily work revolves around ChatGPT’s tools could reach precisely the opposite conclusion.

There is also a naming distinction worth preserving: the roughly $20 ChatGPT subscription discussed here is ChatGPT Plus, not the separate Pro tier. OpenAI’s current documentation lists Plus at $20 monthly and explicitly states that API usage is billed independently. Paying for Plus does not cover ordinary API-key requests from a separate application.

That distinction became central to Jonker’s decision. When occasional GPT use was already happening through an API, the monthly app subscription needed to justify itself through something more than access to another chat box.

Cherry Studio made the overlap visible​

Jonker used Cherry Studio as a common workspace for cloud and local models, connecting cloud providers through API keys and local models through LM Studio’s running server. According to the account, its multi-model comparison feature made it possible to send the same prompt to several models and inspect their answers alongside one another.

The local responses often seemed adequate for everyday questions. Over time, Jonker increasingly started locally and escalated to a cloud model only when necessary.

That is a practical observation, but it needs a careful reading. Two answers that sound similarly polished are not necessarily equally accurate. Jonker’s experience demonstrates which responses met their needs; it does not establish equivalent reasoning, research quality, or reliability.

For readers conducting a similar comparison, the useful question is not simply, “Which answer sounds better?” It is, “Which answer completes my actual task correctly, with the least checking and repair?”

An assistant that produces a lovely paragraph and quietly changes an important fact is not saving time. It is assigning homework.

What actually ran on the 8GB GPU​

Jonker reports using an RTX 3070 with these local workloads:

These are the author’s configurations and observations, not WindowsForum-tested performance results.

Jonker puts the practical ceiling of this setup at approximately 30,000 tokens of context with a quantized 9B model. A larger Gemma 4 12B configuration felt sluggish and left little room for other applications or heavier tool use. The 27B models discussed in the account were not practical options for that card.

The useful hardware lesson is that a model’s download size is not the same thing as its complete running-memory requirement. LM Studio’s official documentation treats context length and GPU offloading as configurable inputs and provides a memory estimator that accounts for context length, flash attention, and whether a model is vision-enabled. A file fitting below the card’s nominal capacity is therefore not a sufficient sizing test.

Nor should Jonker’s experience be turned into a universal “8GB equals 30K context” rule. It describes one configuration, not a guaranteed capability for every model or runtime.

A practical check before loading a model​

For readers who already have LM Studio’s lms command available and a model downloaded, its documented CLI provides a useful preliminary check:

1. Run lms ls to identify the downloaded model’s key.
2. Substitute that key into a memory-estimation command:

lms load --estimate-only <model_key> --context-length 4096 --gpu max

1. Review the estimated GPU and total-memory requirements before attempting the load.

The --estimate-only option prints an estimate without loading the model, and the context and GPU options are reflected in that estimate. This is a sizing aid—not a guarantee of acceptable speed—but it is more informative than judging by file size alone.

API access changes the subscription calculation​

Jonker’s alternative was not to abandon OpenAI entirely. It was to retain occasional API access without maintaining a subscription that saw little use.

OpenAI’s published standard GPT-6.1 Sol prices support the rates cited in the account: $2 per million input tokens and $10 per million output tokens, with cached input priced separately at $0.10 per million tokens.

As an illustrative calculation, one million uncached input tokens plus 100,000 output tokens would cost $3 at those standard token rates, excluding additional chargeable features. That can make occasional access attractive. It does not establish that API billing is cheaper for a frequent user, particularly one who values bundled tools rather than text generation alone. OpenAI documents separate API billing and notes that charges can include applicable tool and storage costs.

The comparison should therefore be between complete workflows, not just subscription dollars and token prices.

The part local chat did not replace​

Jonker specifically misses the Codex workspace, describing a development experience with changes presented beside the conversation and browser annotations for requesting visual fixes.

OpenAI’s documentation independently confirms that Codex is a separate view in the ChatGPT desktop app and can work with local folders, repositories, terminals, and developer tools. It also distinguishes subscription-based Codex access from API-key billing. These are meaningful workflow capabilities, not merely different packaging around a prompt field.

That explains why the account’s caveat matters. Replacing routine conversation is one task. Recreating an integrated development environment is another.

Audit the work, not the launch announcements​

For readers considering the same exercise, Jonker’s experience suggests a sensible evaluation plan:

1. List recent completed tasks. Identify which service actually helped finish each one.
2. Mark unique dependencies. Include research tools, development interfaces, document workflows, and other capabilities you would genuinely miss.
3. Compare representative prompts. Check correctness and required rework, not just presentation.
4. Test the PC under normal conditions. An otherwise idle machine is not necessarily representative of a working day.
5. Measure occasional cloud spending. Compare the actual bill and missing features before deciding that pay-as-you-go is the better arrangement.

As an editorial recommendation, treat an assistant’s summary of your past usage as a starting point rather than a complete audit. Cross-check it against the conversations and work you can actually inspect.

The strongest conclusion here is refreshingly modest: an existing PC can change what an AI subscription is worth to its owner. Jonker’s RTX 3070 did not replace every cloud capability. It absorbed enough overlapping work to expose a recurring payment without a distinct purpose.

That is less dramatic than declaring a chatbot pointless—and considerably more useful when the next renewal arrives.