Pricing and Rate Limits — Real Numbers

The numbers move; the structure is stable

Pricing changes; rate limits change; the categories don't. Each provider has: input tokens, output tokens, cached input tokens, image tokens, audio tokens, and reasoning tokens with their own bills. Rate limits exist as RPM (requests per minute), TPM (tokens per minute), and per-model concurrency caps.

Cost intuition (check live numbers before quoting)

Top-tier (Opus / GPT-5.5 / Gemini 2.5 Pro) is roughly 5–15× cheaper on input than on output, and reasoning tokens are billed separately.
Mid-tier (Sonnet / GPT-5.5-mini / Gemini Flash) is typically 3–10× cheaper than top-tier for similar throughput.
Cached input is 5–25% of normal input.
Open-source per-call is amortized compute; the bill comes from utilization, not per-token rate.

Rate limit traps

Burst patterns hit RPM limits before you exhaust TPM. Smooth your traffic.
Long-context calls eat your TPM budget faster than throughput suggests.
Some providers gate top-tier models behind tier-up requirements.

Code

Cost monitor (sketch)·python

from collections import defaultdict

buckets = defaultdict(int)
for resp in stream_recent_responses():
    buckets[(resp.provider, resp.model, "input")] += resp.usage.input_tokens
    buckets[(resp.provider, resp.model, "output")] += resp.usage.output_tokens
    if resp.usage.cache_read_input_tokens:
        buckets[(resp.provider, resp.model, "cache_read")] += resp.usage.cache_read_input_tokens
    if resp.usage.thinking_tokens:
        buckets[(resp.provider, resp.model, "thinking")] += resp.usage.thinking_tokens

Pricing and Rate Limits — Real Numbers

The numbers move; the structure is stable

Cost intuition (check live numbers before quoting)

Rate limit traps

Code

External links

Exercise

Progress

Comments 0