New models in the price index

The rate table behind this site changes every week, usually by addition: a new identifier appears, gets a rate, and becomes callable within hours. This page is the running record of those additions — what arrived, what it costs, and which ones actually change the arithmetic. 839 models are tracked; 11 arrived in the latest wave on 2026-09-26.

This wave — 11 models added 2026-09-26

Read the last two columns rather than the first two. The input rate is the number vendors print first and the number that matters least: output is priced at a median of 4.0× input across all 216 models here, and on agent workloads output is most of what you actually buy.

What the rates tell you

• The cheap end moved, not the flagship end. The lowest agent-mix rate among this wave is mimo-v2.6-flash at $0.1225 per 1M blended tokens — $2.24 a month at 20M input and 6M output.
• Some arrivals are free upgrades. 4 of them land on exactly the same rate as a model already in the table (grok-4.7 matches grok-3-deepersearch; mimo-v2.6-flash matches mimo-v2.5; mimo-v2.6-pro matches mimo-v2.5-pro; qwen3.8-omni-flash matches qwen3.8-flash) — a version bump with no price movement, which is the cheapest kind of upgrade there is.
• One ratio always stands out. The steepest is gemini-3.8-flash-tts at 18.0× output to input — 4.5× the 4.0× median across all 216 models. On a workload that produces more tokens than it reads, that ratio is the whole bill.
• None of this wave has a published list rate yet. 11 of 11 carry no vendor-published price we can cross-check, so no comparison is implied for them. That is normal for a freshly added identifier — and it is exactly why this page prints rates instead of discounts.

Calling a model the day it appears

Compatible endpoints do not gate new identifiers behind a release; if the rate table has it, the endpoint serves it. The practical problem is discovery, because most tools will not populate a dropdown from /v1/models for you. List the endpoint's models first and copy the identifier exactly as it comes back.

# List every model the endpoint currently serves, newest additions included
curl -s https://aicomp.ai/v1/models -H "Authorization: Bearer $KEY" \
  | python3 -c "import sys,json;print('\n'.join(m['id'] for m in json.load(sys.stdin)['data']))"

Then call it like any other model. The only thing that changes is the string in model — request and response handling stay identical, which is the whole point of one compatible endpoint.

from openai import OpenAI

client = OpenAI(base_url="https://aicomp.ai/v1", api_key="sk-...")

# A model that only appeared in the rate table this week
resp = client.chat.completions.create(
    model="gpt-6-luna",
    messages=[
        {"role": "system", "content": "Answer in one sentence."},
        {"role": "user", "content": "What changed in this release?"},
    ],
    max_tokens=200,
)

# usage is the only place the real price shows up — log it, don't guess
print(resp.usage.prompt_tokens, resp.usage.completion_tokens)

The second half of that snippet is the part people skip. A new model's headline rate tells you the price of a token, not the price of your workload, and the two differ by whatever your input-to-output ratio turns out to be. Log the usage object for a week and you have the ratio; guess at it and you have a number that is wrong in the expensive direction.

import json, datetime

def log_call(model, usage, extra=None):
    """写一行 JSONL。新模型上线前几周，这是唯一可信的成本数据来源。"""
    rec = {
        "ts": datetime.datetime.utcnow().isoformat() + "Z",
        "model": model,
        "in": usage.prompt_tokens,
        "out": usage.completion_tokens,
        **(extra or {}),
    }
    with open("usage.jsonl", "a") as fh:
        fh.write(json.dumps(rec) + "\n")

# 用它算出你自己的 in:out 比例，再回到本站的计算器按该比例比价
# ——比对着价目表猜要准得多。

When a new model is the wrong call

Every wave we have tracked

First-seen dates are recorded automatically each time the rate configuration is read, so the list below grows on its own. The oldest row is the state of the catalogue when tracking began — everything above it arrived after.

How this compares to the rest of the index

Additions matter because they change the ranking, and the ranking changes when you stop sorting by input price. Across the 216 cleaned models in the full index, sorting by raw input price instead of a realistic agent mix can put a model 68 places out. A new entry at the cheap end is worth checking against that re-ranked view, not against the top of a price table.

Related

• Full rate index — all 839 models
• 216 models re-ranked by token mix
• Why output costs 4.0× input on median
• Why the same model has two prices
• Cost calculator — model your own token mix
• The 2026 price index — method, findings, dataset
• Browse models by vendor

Get API access