BYOK vs platform-paid AI: the real cost of 1,000 features

The question that motivated this post

If you're building software with AI tools in 2026, you're picking from two pricing models:

• Platform-paid — pay $20–40/mo to Lovable / Bolt / Replit / Cursor; the platform absorbs the model cost
• Bring-your-own-key (BYOK) — pay the AI provider directly (Anthropic, OpenAI, OpenRouter), the platform layer is free or near-free

Every comparison post I've read on this is either written by a platform with a vested interest in flat-rate, or written by a BYOK partisan who picked numbers to make BYOK win. The actual answer is: it depends on how much you build, and the breakeven point is sharper than most people realize.

This post does the math for both directions, with real per-task token costs measured against real OpenRouter and Anthropic prices. By the end you should be able to look at your own usage and pick the right model for you, not the right model for whoever wrote your last comparison post.

What a "feature" actually costs in tokens

The first place this analysis usually goes wrong is the unit. People say "AI coded my app for $20" and don't mean anything specific. Let me be specific.

A "feature" in this post = one round-trip with the agent that ends in deployed code. Examples:

• "Add a /about page that links to the social icons in the footer"
• "Fix the 500 on /api/users — I think the joins are wrong"
• "Make the hero section shorter and tighten the CTA copy"
• "Add Stripe checkout for the basic plan"

These are real prompts from real VibeKit users in the past two weeks. They span trivial CSS tweaks all the way to multi-step integrations. Token usage spans roughly:

Cost figures use Anthropic's published Claude Sonnet 4.6 prices — $3/M input, $15/M output. (Cache reads are cheaper, ~$0.30/M; that's why the average input cost is lower than the worst case once an agent has a warm cache.) These are Anthropic-direct prices. OpenRouter passes through within ~5% on most providers.

Two important things about this table:

1. Most features are "small" or "medium." In any given week of building, 70% of my tasks are in the $0.04–$0.13 range. The big-ticket integrations are exceptions, not the norm.

2. The cost-per-feature distribution has a long tail. If you're shipping primarily product polish (icons, copy, layout fixes), your average is closer to $0.05. If you're shipping infrastructure work (auth flows, billing, third-party APIs), your average is $0.25+.

The takeaway: the "average" feature cost depends heavily on what you're building. Don't trust any single number you see online — including mine.

What 1,000 features cost in tokens

A useful benchmark: a side project from blank to "real product" is somewhere between 200–2,000 features. A serious solo product over a year is 1,000+. Let me run that through the cost model.

Assume a realistic mix: 60% small, 25% medium, 12% large, 3% heavy. (These match what I see on VibeKit; your distribution will differ.)

1,000 features at this mix:

• 600 small × $0.042 = $25.20
• 250 medium × $0.128 = $32.00
• 120 large × $0.360 = $43.20
• 30 heavy × $0.675 = $20.25
• Total: $120.65

That's the BYOK base cost for a year-ish of building a real product. Some of you are already nodding ("that tracks"), some are surprised it's that low. Both are reasonable reactions — it depends on what you've been led to believe AI coding costs.

What the platforms charge

Now compare to flat-rate platform pricing. Verified on each vendor's pricing page on 2026-05-11. These are list prices for the entry-level paid tier:

(Verified 2026-05-11. They change — Replit Pro doubled from ~$45 to $95 in the last 12 months.)

A few things jump out:

1. The "unlimited" plans aren't actually unlimited. Lovable Pro is capped at 100 monthly credits — not "unlimited" by any honest reading, despite the marketing. Bolt's 10M tokens/mo translates to roughly 50–80 medium features before you hit "out of credits." The marketing makes the cap easy to miss.

2. The credit / token / "Agent credit" abstractions are deliberately opaque. Lovable's "credits," Replit's Agent credits, v0's "credits" — none map cleanly to anything you can predict. The opacity is a feature, not a bug — it means you can't easily compare across platforms.

3. The annual price has bifurcated. Entry tiers cluster $240–$300/yr; heavy-use tiers have jumped to $1,000+ (Replit Pro $1,140, Cursor Pro+ $720, Cursor Ultra $2,400). If your annual BYOK spend exceeds the entry tier, the platform wins on raw cost — but only at the entry tier. Heavy users are paying laptop-leasing money for a single subscription.

The breakeven calculation

Annual BYOK spend = (monthly features) × 12 × (avg cost per feature)

Solving for "monthly features at which BYOK = $240/yr":

$240 / 12 / $0.12 = ~167 features per month

That's about 5–6 features per day, every day, all year. For most solo builders, that number is high — you're not shipping 5 real features per day on a side project. You're shipping 5 features per week on a good week.

Inverting: at 5 features per week (~22/mo) and average cost $0.12, annual BYOK is $32. The platform plan would cost you $240 — an 8x premium.

Where the break flips:

The takeaway: for solo builders shipping at solo-builder pace, BYOK wins by 4–8x. For teams or full-time AI-coding-as-job users, flat-rate wins because token costs scale linearly with usage and platform plans don't.

This is exactly the inverse of how every platform's marketing copy frames it. Their copy is optimized for the heavy user; their margin comes from light users overpaying.

What gets left out of the math

Three things complicate the simple "$120 vs $240" comparison.

1. Subscription bundles you already pay for. If you have Claude Pro ($20/mo) or ChatGPT Plus ($20/mo), and the agent platform supports OAuth into those subscriptions, your AI cost might be zero on top of what you're already paying. VibeKit supports both — your Claude Pro / Max subscription works directly. In this case BYOK isn't $120/yr — it's $0/yr for AI, and you only pay the existing $20/mo Claude subscription which you'd be paying anyway.

This single factor flips the math for anyone already on a Pro subscription. The breakeven moves from "5 features/day" to "you'd literally never hit a flat-rate plan's value."

2. Hidden platform costs. Most flat-rate platforms also charge for hosting, databases, custom domains, and "team features." A $20/mo plan often expands to $50–80/mo by the time you need anything real. BYOK platforms tend to charge for these separately and predictably (or include them in the free tier).

3. Cost of "platform lock-in." Hard to put a dollar figure on, but real. If your code lives only inside a platform's sandbox, leaving means rewriting everything. If your code lives in a real GitHub repo (which BYOK platforms tend to give you), leaving costs nothing.

When platform-paid actually wins

I want to be honest about the cases where flat-rate is the right answer, because those exist:

1. You ship constantly and don't want to think about it. If you're a full-time AI-coding user shipping 50+ features per week, the cognitive cost of watching token burn outweighs the savings. A flat $20/mo and "stop thinking about cost" is genuinely worth it.

2. Your team is bigger than 1. Per-seat token economics get expensive fast when you have 5 people each running their own agents. Platform plans with per-seat pricing make this predictable in a way that BYOK doesn't, especially when you have to reconcile shared OpenRouter accounts at end of month.

3. Your client wants a single invoice. If you're billing this through a customer or employer, "I expensed Lovable Pro for $25" is cleaner accounting than "I expensed $32.47 of Anthropic API tokens this month."

4. You don't want to manage a third-party account. Some users just don't want to maintain an Anthropic console account, watch the credit balance, and rotate keys. That's reasonable; flat-rate trades money for not-thinking-about-it.

The first one is the strongest. If you're shipping all day every day with AI as your daily driver, flat-rate is genuinely better. The "BYOK is always cheaper" claim is wrong at high usage.

What I optimized for in VibeKit

I'm biased — I built VibeKit on the BYOK premise. But the bias is rooted in this exact math: most users (solo builders, side projects, weekend hackers) ship 10–50 features per month, not 200+. For that population, flat-rate is a 3–8x markup on what they actually use. We don't take a cut of AI spend on the free tier — your Anthropic / OpenAI bill goes straight to them, no markup, no middleman.

Where VibeKit does charge is on the optional paid tier for users who want bundled billing or specific managed features (always-on apps, higher session caps). Those are real costs we incur and pass through, not the AI itself. The AI itself is between you and the model provider, and we don't insert ourselves.

You can read more on /byok-ai-agent about the implementation. The pricing page at /pricing has the actual numbers.

How to decide for yourself

Three questions:

1. How many features per week do you realistically ship? Be honest. If you don't know, count the merge commits in your last 4 weeks.
2. Do you already have a Claude Pro or ChatGPT Plus subscription? If yes, BYOK via OAuth is essentially free.
3. How much does mental overhead cost you? If watching a token meter stresses you out, that's a real cost — flat-rate buys you not-thinking-about-it.

If you ship under 20 features/week and either already have a Pro AI subscription or are willing to spend ~$10/mo on tokens, BYOK wins by a wide margin. If you ship 50+/week and are billing through a company, flat-rate is fine and saves you a few hours of reconciliation.

If you're somewhere in the middle (20–50/week, no AI subscription), the answer is "either is fine" — pick based on which platform you actually like building on, because the cost difference at that volume is in the noise.

The deeper point: you can do this math. The opacity of platform pricing is intentional, and it works because most people don't bother running the numbers. If you're spending money on a category, even small money, run the math. Sometimes the cheap-looking option is the expensive one.

Coming next: a piece on what "AI agent" actually means, and which popular tools that call themselves agents actually qualify under any rigorous definition. (Spoiler: fewer than the marketing suggests.)

Use VibeKit free with BYOK if you want to test this math against your own usage. Or follow @609.sol on X for new posts.