BYOK vs. SaaS AI: what you actually pay, what you actually own

On May 13, 2025, a federal magistrate judge ordered OpenAI to retain every ChatGPT log going forward — including conversations users had explicitly deleted. The order ran for five months. If you used ChatGPT Free, Plus, Pro, Team, or the standard API between mid-May and September 26, 2025, your chats — including the ones you thought were gone — sit on OpenAI’s servers right now. On November 12, 2025, the court ordered roughly 20 million of those logs handed over to plaintiffs.

That’s the kind of thing that doesn’t happen with software you ran on your own machine, with your own keys. SaaS AI is the default because it’s the easiest thing to ship, not because it’s the right shape for the work. This post is an opinionated breakdown of the three things you actually buy when you click subscribe: the price, the data, and the durability — and why bring-your-own-key wins on all three at any usage tier above “light.”

The math

Start with the cost question because it’s the cleanest. Pin one number: a typical “chat” is about 5,000 input tokens (system prompt + a couple of turns of context) and 1,000 output tokens (a substantive answer). At Claude Sonnet 4.6’s API rates ($3 input / $15 output per million tokens) that costs three cents. Cheaper at GPT-5-mini (~$0.003); cheaper still on local Llama 3.3 70B (the electricity).

Two things to notice. First, at the lightest tier the SaaS markup is absurd — paying $20 for ChatGPT Plus when you’d run $3 of API traffic is a 6× premium for convenience and brand. Second, the markup multiple doesn’t shrink as you scale. A power user running ChatGPT Pro + Cursor + Midjourney + ElevenLabs + Suno + Runway is paying somewhere around $270 a month for what costs about $83 in direct API spend. Anthropic’s prompt caching cuts that further by about 90% on cached prefixes; batch APIs run another 50% off. The honest BYOK power-user number is closer to $40 a month with caching turned on.

Note the asymmetry. The SaaS subscription is billed whether you use it or not; the API is billed per request. A regular user travelling for a week or hitting a slow month at work wastes the SaaS premium and saves on BYOK by exactly what they didn’t use. This matters more than the headline markup, because most people’s month-to-month usage is much spikier than they admit.

What you actually own

SaaS terms-of-service give you the answer in lawyerly form. The operative version, after a year of policy turbulence:

• Conversation history.Lives on the provider’s servers, indexed for retrieval and (subject to your opt-out) eligible for training. This is not theoretical risk; the NYT-driven preservation order made every ChatGPT user’s deleted chats discoverable for five months. Enterprise and zero-data-retention API customers were excluded — the consumer plans you and your friends use were not.
• Generated content. OpenAI assigns you the output but other users may receive identical outputs and you may not use them to train competing models. Midjourney requires the Pro or Mega plan for any business with $1M+ in revenue to use images commercially. And the U.S. Copyright Office (Jan 2025) has been clear: AI-generated content without sufficient human creative input isn’t copyrightable at all. You can sell it. You can’t stop someone else from copying it.
• Custom configurations. Custom GPTs are not included in OpenAI’s data export. Claude Projects can’t be exported as a bundle. Cursor Rules are the lone exception — they live as .cursor/rules/*.mdc files inside your repo, version-controllable, portable. That tells you which company is actually thinking about lock-in and which two are quietly betting on it.
• Memory.Each provider’s memory is siloed. Anthropic literally shipped a tool to extract memories out of ChatGPT to help users migrate to Claude — implicitly conceding that the memory feature is a moat, not a benefit.

With BYOK, every one of these flips. Conversations live on your disk in plain files you can grep. Outputs are saved as files you own without a multi-tier commercial-use clause. Custom configurations are local files. There is no “memory” service to extract from, because the prompts and context you assemble are just text on your machine. You’re not arguing about ownership; you have it by construction.

The graveyard

The third question is the one nobody asks at signup: what happens when the company dies, gets bought, or quietly stops investing in the product? It’s not a hypothetical. By late 2025, 73% of AI startups funded in 2023 had pivoted or shut down. The pattern of failure has gotten specific:

The acquihires are the worst kind of failure mode because they don’t look like failure on the way down. The company gets a headline windfall (Character.AI was a $2.7B deal; Inflection, $650M; Adept, ~$25M plus investor make-whole), the founders end up at Microsoft / Google / Amazon, and the product limps along until subscribers churn out. You don’t get a shutdown date. You get a slow drift into irrelevance.

The pricing pivots

Even when a SaaS doesn’t die, the deal you signed up for can change overnight. The 2025 pricing-pivot list is long, and almost every one moved the cost up:

The pattern: a flat-rate subscription that looked like a product turns into a usage-metered service that looks like an API — without giving you the API. You pay the markup, accept the rate limits, and lose the predictability that was the whole reason to subscribe in the first place.

BYOK’s tradeoffs

Every honest critique of BYOK boils down to one of these five things, and they’re all real:

• Key management.A regular user keeps three to five API keys (OpenAI, Anthropic, Google, ElevenLabs, Replicate or fal). Each provider rotates independently. If a laptop is stolen, there’s no central revoke; you go to five dashboards.
• No unified billing. Five providers means five invoices, five tax forms, five FX conversions. SaaS giants put one line on your statement.
• Memory is siloed.ChatGPT memory and Claude memory don’t talk to each other. Until your wrapper app does the unification (CSuite’s project folder is one approach), you re-state context per provider.
• Premium features lag. Sora video, ChatGPT Deep Research, Operator agent, Atlas browser features — these ship to subscribers first; the API surface lags by months.
• Tooling gap.Raw APIs need a wrapper. Until that wrapper exists you’re picking between Postman scripts and someone else’s SaaS UI.

Three of these are real engineering work; two are real product gaps. None of them remove the cost or ownership advantages — but they do explain why most people stay on subscriptions even when the math clearly says otherwise.

The rational default

CSuite is, frankly, what you get when you take the BYOK argument seriously and then solve the five problems above. One desktop app, one model picker, one project folder on your disk. Cloud providers run with your own keys; the local runtimes (Ollama + HuggingFace) cover the long tail and the “run nothing through the cloud” case. Your conversations, generations, and edits land in regular files on disk. There is no “memory service” to be extracted, because there’s no service. There’s no retention order to be subject to, because there’s no log to retain. There’s no shutdown date, because the version you paid for keeps working.

We didn’t set out to build an anti-SaaS app. We built the tool we wanted ourselves and noticed the result was BYOK by default. The more time we spent looking at the cost math, the lock-in pattern, and the shutdown graveyard, the more this looked like the rational default — not a quirky alternative.

The cleanest version of the argument is this: if you’re a casual ChatGPT user, the $20/month is an acceptable convenience tax. If you use AI for real work — coding, writing, content production, research — the SaaS stack is taking 3× more than it should, and the durability of that stack is far weaker than you think. BYOK plus a local-first wrapper isn’t the alternative. It’s the version where the math, the data, and the durability all line up.