How indie iOS devs should think about LLM costs in 2026

Per-token pricing on Claude, GPT, and Gemini fell again this spring. But the cost that actually matters for an indie iOS app is rarely the per-token line — it's the prompt-design decisions that compound across millions of calls.

Per-token pricing across the major LLM APIs has dropped roughly 5× over the past 12 months. Anthropic Claude Haiku, OpenAI GPT-4o-mini, and Google Gemini Flash all now sit in the sub-$0.50-per-million-input-token range. For an indie iOS app integrating an LLM, the temptation is to read those numbers and conclude the cost problem is solved.

The cost that actually matters

It isn't. The per-token rate is the smallest variable. The cost of LLM features in production scales with three other knobs that most indie devs under-tune: prompt length, cache hit rate, and fallback behavior. A 2,000-token system prompt that gets resent on every call (because you didn't enable prompt caching) costs the same per-month as 20,000-token prompts that ARE cached.

Anthropic and OpenAI both ship prompt caching now. Anthropic's cache reads at 10% the input-token rate; OpenAI's caches at 50%. Not using it is the single largest under-pulled cost lever in the LLM economy.

The second knob — model-routing

Most app integrations send every request to the same model. The smarter pattern is a router that sends 80% of traffic to the cheap-fast tier (Haiku, Gemini Flash) and only escalates the hard 20% to the expensive tier. For chat-style features this can cut bill-per-MAU by 4–8×.

The third knob — output guardrails

An LLM that's allowed to output up to 4,000 tokens will routinely produce 800–1,200 even when 200 would do. Lower the max_tokens. Force structured JSON output where possible. Output tokens are 2–4× more expensive than input.

What this means for app builders

If your AI feature's monthly bill is creeping up, audit prompt-caching, model-routing, and output limits before assuming you need to switch providers. The biggest savings in 2026 aren't in negotiating with the vendor — they're in the prompt-design decisions you control directly.