There is no cheapest model — only a cheapest model for your token mix

Every price table you have read ranks models by input price, because that is the column vendors print first. Real invoices do not look like that. Output tokens cost a median of 4.0× the input rate across 209 models, and on agent workloads output is most of what you buy. Re-rank the same 209 models on three realistic token mixes and the top ten barely overlap.

curl https://aicomp.ai/v1/models \
  -H "Authorization: Bearer sk-your-gateway-key"

Three mixes, three different winners

Below is the same 209-model index ranked three times: once for a chat workload (long context in, short answer out), once for coding agents (short prompt, very long output), once for batch generation (almost pure output). The blended figure is what 1M tokens actually costs you at that ratio.

1. Chat and RAG — 3 parts input to 1 part output

Median across all 209 models at this mix: $0.9 per 1M tokens.

2. Coding agents and tool loops — 1 part input to 3 parts output

Median across all 209 models at this mix: $1.425 per 1M tokens.

3. Batch generation — 1 part input to 9 parts output

Median across all 209 models at this mix: $1.65 per 1M tokens.

Median blended rate moves from $0.9 on the chat mix to $1.425 on the agent mix and $1.65 on the batch mix — the same 209 models, only the ratio changed. And only 7 of the top ten survive the move from chat to batch.

Where “cheapest” flips outright

These models move the most between the two extremes. Every one of them is a model with a flat or near-flat output rate — cheap-looking on an input-ranked table, genuinely cheaper once output dominates.

Largest move between the two mixes: 42 places out of 209.

The input-price trap, quantified

Compare the rank you get by sorting on input price alone against the rank you get by sorting on the real agent mix. The gap is the size of the mistake:

Largest gap: 66 places. At the top of the index that is the difference between %s and %s.

If you pick the vendor first: every vendor's floor

Most teams do not choose from a flat list of 209 models — they have a vendor already and want to know where the bottom of that range is. This is it, with the width of each range next to it.

Measure your own mix before you read any of this

Ten lines of code is all it takes, and it replaces every assumption above with your own number:

# 先量出你自己的 output 占比，再去看榜单。
# 这一步花 10 分钟，能省掉一次错误的选型。
import anthropic, collections

client = anthropic.Anthropic()
stats = collections.Counter()

for _ in range(200):          # 采样 200 次真实调用就够稳
    r = client.messages.create(
        model="claude-sonnet-5", max_tokens=1024,
        messages=[{"role": "user", "content": next_prompt()}],
    )
    stats["in"]  += r.usage.input_tokens
    stats["out"] += r.usage.output_tokens

total = stats["in"] + stats["out"]
print("output share: %.1f%%" % (100 * stats["out"] / total))
# 把这两个数代回本页的 blended 公式，而不是照抄别人的榜单。

Then blend the two candidate rates at your ratio:

blended = (input_rate * in_share + output_rate * out_share)

That one line is the whole method. If you skip it, you are comparing someone else's workload to your invoice.

FAQ

Related

• Why output tokens dominate your bill (4.0× median, up to 12×)
• Full rate index — all 209 models, filterable
• Cost calculator — blend any model at your own ratio
• What a cache hit does to the input side
• Why the same model shows two prices
• Browse models by vendor

Get API access