The subscription
Claude Pro is $17 a month billed annually, $20 monthly. Claude Max is "from $100" with "5x or 20x more usage than Pro."
Guide
The cheapest way to build a mobile app in 2026
It is not a tool. It is a pricing model: a flat-seat builder orchestrating the AI subscription you already pay for. Here is the arithmetic from the vendors’ own published prices, the history of why app builders started on credits, and the honest limits of the cheap route.
Every AI app builder is, underneath, the same machine: an LLM writes code, something compiles and runs it, errors go back to the LLM, repeat. The tools differ in polish and platform — but the largest line in what you pay is the same everywhere: inference, the cost of the model thinking.
So the cheapest way to build is not a brand; it is whichever arrangement gets you inference cheapest. And in 2026 that question has a clear answer with a slightly surprising shape: the cheapest inference an individual can buy is a consumer LLM subscription — Claude, ChatGPT, Gemini, Copilot — and the cheapest way to build an app is a tool that lets you bring that subscription with you, instead of selling you inference again by the credit.
The industry shorthand is BYOA — bring your own agent. This article is the case for it, with every number from a published pricing page, dated. Modaal is a BYOA tool and we sell it; the disclosure you expect from this site applies, and so does the discipline — including the parts of the argument that cut against us.
Why vibecoding tools started on credits — a fair history
When the first wave of AI app builders shipped, credits were not a trick. They were the only viable model.
A tool in 2023 or 2024 bought inference from the labs at metered API rates and resold it inside the product. Every prompt a user typed cost the vendor real, variable money — so every prompt had to be counted, and the credit was born: a unit that maps your usage to their API bill, plus margin. Rork's definition is the category's in miniature: "a credit is a unit of AI compute, roughly one request where Rork builds or edits your app." Lovable meters the same way. It was rational, it still is rational for the vendors — and it quietly makes the user pay retail-plus-markup for every attempt, including the failed ones.
What changed is not that credits got worse. It is that a second channel for the same inference appeared and was priced completely differently: the consumer subscription. The labs sell individuals flat monthly plans whose included usage, if you bought it at their own metered rates, would cost several times more. Once that channel existed, reselling API inference by the credit stopped being the only model — and stopped being the cheap one.
Credits are not evil. They are historical — the founding constraint of a category, outliving the constraint.
The arithmetic: what the flat plans are actually worth
Everything in this section is from claude.com, read 30 August 2026, using Claude as the worked example because it publishes both prices side by side. The shape is similar at the other labs; check their pages for the current figures.
The meter
The same vendor's API sells Sonnet 5 at $2 per million input tokens and $10 per million output tokens (Opus 5 at $5/$25; Fable 5 at $10/$50).
Now the part you do not have to take from us
Claude's own pricing page explains what happens when a plan's limit is hit: you can wait, upgrade, or "turn on usage credits" to continue "at standard API rates." Read that twice — the vendor itself prices overflow usage at API rates. The gap between the flat plan and metered usage is not our marketing framing; it is built into their own price list.
An illustration — assumptions stated, not a claim
An agentic coding session is token-hungry: the agent re-reads project files, generates code, reads build errors, regenerates. Suppose a day of real building processes 5 million input tokens and 500,000 output tokens — a deliberately rough figure. At Sonnet 5 API rates that is:
- 5M input tokens × $2 per million
- $10
- 500K output tokens × $10 per million
- $5
- One day of building — twenty such days, ~$300 a month
- $15 vs a $17–20 plan
Assume our token estimate is too high by threefold and the metered month still costs several times the subscription. The precise multiple is unknowable and does not matter; the direction is enormous, and it is the vendors' own two price lists that create it.
Whether the labs profit on the plans, break even, or invest in them is their business and we will not speculate. For you, the only relevant fact is the checkable one: the same building month costs a small flat fee on one channel and a large metered bill on the other.
BYOA: the model that rides the cheap channel
A bring-your-own-agent tool inverts the old architecture. Instead of reselling inference, it orchestrates an agent you already pay for — Claude Code on your Claude plan, Codex on your ChatGPT plan, Gemini CLI, Copilot and the rest — and charges only for what the tool itself adds: the workflow, the scaffolding, the build loop, the platform expertise.
That split is why the economics work. The lab bills you once, at the flat consumer rate it set. The tool bills you once, at a flat seat priced like software rather than like compute. Nobody in the chain is metering your attempts, which changes behaviour exactly where it matters: the fortieth debugging iteration costs what the first did, so you iterate like someone who owns the meter — because there isn't one.
In the native-app category, two tools verifiably work this way. Modaal (ours): you connect any of thirteen supported agents with your existing subscription, and the free plan is one active project with unlimited prompts — Pro at €9 a month billed annually adds both platforms and App Store distribution. Bitrig: native Swift from founders who co-created SwiftUI, with agents connecting over the Agent Client Protocol — Claude Code, Codex, Cursor and others — and even its free tier is BYOA, offering "10 daily requests using your Claude or ChatGPT subscription"; Pro is $25 a month on their website. The two tools differ in meter (Bitrig counts credits for its own features; Modaal counts nothing) and in scope (Bitrig is Apple-only by design; Modaal does iOS and Android) — the full side-by-side is at Modaal vs Bitrig.
And the purest BYOA of all costs nothing: Claude Code driving Xcode directly, through Apple's own MCP bridge, on your existing Claude plan. No builder at all — you are the orchestration. It is the right answer for terminal-comfortable developers, and its existence keeps every BYOA tool honest about what its seat fee buys. Apple shipping agent support inside Xcode 26 tells you where the platform itself thinks this is going.
The honest limits of the cheap route
Three caveats, because "cheapest" earns trust only when it is bounded.
Flat is capped, not infinite. The consumer plans have usage limits — Claude's reset on "a rolling five-hour session window." A BYOA tool's "unlimited prompts" means the tool adds no meter; the agent underneath still has its plan's ceiling, and a truly heavy day can bump it. The difference from credits is what happens next: you wait a few hours or upgrade a plan you use for everything, rather than watching a tool-specific balance drain per attempt. But a cap is a cap, and pretending otherwise would be the kind of sentence this site does not write.
BYOA is cheapest if the subscription is already in your life. Our target reader pays for Claude or ChatGPT anyway — then the marginal inference cost of building is zero and the arithmetic above is overwhelming. If you pay for no AI plan and would not otherwise start one, BYOA means opening one, and a bundled-credit tool with one bill is simpler and possibly cheaper at very light usage. Below some floor of building activity, simplicity wins.
Credits sometimes buy more than inference. Rork's $200 Max tier is not a pure inference markup — it includes building on their cloud Macs and managed App Store submission, which is why it is the right answer for founders with no Mac at any inference price. When a credit price bundles real infrastructure, compare it to the infrastructure, not just to tokens. The unbeatable version of cheap — a Mac you own, a subscription you already pay, a free BYOA seat — requires the Mac.
The cheapest stack, by situation
Published prices only, 30 August 2026. Apple’s $99/year applies to every row the day you ship.
| Your situation | Cheapest route | What it costs |
|---|---|---|
| Mac owner, already paying for Claude/ChatGPT/etc. | BYOA — Modaal Free (or Bitrig’s free tier for Apple-only) | €0 added. The subscription you already pay becomes your build engine. |
| Same, but shipping seriously — both platforms, App Store distribution | BYOA — Modaal Pro | €9/month billed annually (€15 monthly). |
| Developer happy in a terminal | Claude Code + Xcode via Apple’s MCP bridge | Your existing Claude plan. The tooling itself is free. |
| No AI subscription, very light building | A bundled-credit tool’s entry tier | One bill, metered. Simplicity beats optimisation at low volume. |
| No Mac, native iOS required | Cloud builders — Rork Max | From $200/month; the credits bundle cloud Macs and managed submission, which is the part you are actually buying. |
Frequently asked questions
If you own a Mac and already pay for an AI subscription like Claude or ChatGPT: a bring-your-own-agent (BYOA) tool, because your existing plan supplies the inference and the tool adds no meter. Modaal’s free plan is a real native app — one project, unlimited prompts, code you own — at zero added cost, with Pro at €9/month billed annually. The only universal extra is Apple’s $99/year to ship. Without a Mac or an AI plan, the answer changes — see the situations table in this article.
The tool orchestrates an AI coding agent you already subscribe to — Claude Code, Codex, Gemini CLI, Copilot and others — instead of reselling inference by the credit. You authenticate with your existing plan; the tool charges a flat seat for its workflow, scaffolding and build-loop management, and adds no per-prompt or per-token fee. In the native-app category, Modaal and Bitrig verifiably work this way; Claude Code driving Xcode directly is the do-it-yourself version.
We can show you the gap without explaining the vendors’ economics, which are not public. Claude Pro is $17–20/month; the same vendor’s API sells Sonnet 5 at $2 per million input tokens and $10 per million output tokens, and their own pricing page says overflow usage beyond plan limits runs "at standard API rates." Agentic coding sessions process tokens by the million, so a month of daily building at metered rates costs several times the flat plan — the exact multiple depends on your usage, but the direction is written into their own two price lists.
They are the older model, kept honest by history: early tools bought inference at API rates and had to meter what they resold. Today a credit means paying metered rates plus margin for every attempt, including failed ones, while the flat consumer channel exists — which is why we call the model historical. The fair exceptions: bundled credits are simpler if you have no AI subscription, and some credit prices — Rork Max’s $200 — bundle real infrastructure such as cloud Mac builds and managed submission, which is what you are actually paying for.
The tool side is: Modaal adds no meter, cap or per-prompt charge on any plan. The agent underneath follows its own plan’s limits — Claude’s plans, for instance, reset usage on a rolling five-hour window. In practice that means a heavy day can pause you for a few hours on the plan you use for everything, rather than draining a tool-specific credit balance per attempt. Capped differently is still capped; we would rather say so than have you discover it.
Between the two we can verify in the native-app category: Modaal if you want iOS and Android, nothing metered, and the widest agent support (thirteen agents, free plan with unlimited prompts); Bitrig if you are Apple-only and value its founders’ SwiftUI pedigree — its free tier is genuinely usable and Pro is $25/month on their site. The full sourced comparison is at modaal.dev/alternatives/bitrig. Terminal-comfortable developers should also price the zero-cost option: Claude Code driving Xcode directly.
Start building for free
Build for as long as you want on the Free plan. Only upgrade when you're ready to launch.
Keep reading
- How much does it cost to create a mobile app?
The full bill, itemised and checkable.
- Modaal vs Bitrig
The two verifiable BYOA tools, side by side.
- Rork alternatives
The credit model at its most expensive, sourced.
- The 7 best software to develop iOS apps in 2026
Every route, priced from vendors’ own pages.