LLM API prices per 1M tokens in 2026: 10 models vs list

Comparing LLM prices across providers is annoying because everyone quotes a different unit — per token, per million tokens, per character, in a currency that isn't yours. This is the full genmux chat-model catalogue — the 10 text models the upstream vendors can actually serve today — one table, one unit: USD (and credits) per 1,000,000 tokens, input and output priced separately, checked against official vendor list prices where that comparison is honest to make.

Reading the numbers

We meter in credits; 1 credit = $0.0125 at the best-value pack (Scale: 12,000 credits for $150). The table below shows both, since your actual per-credit cost depends on which pack you buy. Our pricing rule: we sit 10% under the model owner's own list price wherever our supply cost allows, and never below it. Prices update if upstream costs move — treat this as a snapshot, checked 2026-09-06, not a permanent quote.

"Tools" is OpenAI-style function calling (tools / tool_calls), which genmux emulates on the seven flagged models and which is what OpenCode, Cline, Roo Code and Continue need — the setups are on the coding agents page. Every other text model id you might have seen mentioned elsewhere (GLM, Kimi, Qwen's Plus/Coder variants, the older Qwen2.5 and DeepSeek R1 distill) is off this table because the upstream vendor can't serve it today — no page, no price, until that changes.

Where we're actually cheaper than the model owner

Five models carry a confirmed official list price we can check against, and all five clear our 10%-under bar. DeepSeek's list price above is off-peak — DeepSeek doubles its own price during peak hours, so the real-world gap is closer to 55% cheaper on both input and output during DeepSeek's peak. Doubao's four models are priced above Volcengine's own list (our upstream cost there is higher) — the pitch for Doubao is one key, one balance, and payment options Volcengine's own billing doesn't offer, never a lower number. qwen3-6-flash's reference price is a proxy rather than a confirmed Alibaba list figure, so we don't make a saving claim for it either.

Checking the price of a real call

Every response carries the exact cost, so you don't have to trust the table for production decisions — read it off usage:

from openai import OpenAI

client = OpenAI(base_url="https://api.genmux.tech/v1", api_key="nb_live_...")

for model in ["deepseek-v4-flash", "qwen3-6-flash", "doubao-seed-2-lite"]:
    resp = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": "Summarize the plot of Crime and Punishment in two sentences."}],
        max_tokens=150,
    )
    u = resp.usage
    print(f"{model}: {u.total_tokens} tokens, {u.cost} credits ({u.cost_source})")

Running the same prompt across a few candidate models and comparing usage.cost is a faster way to pick a model for a specific workload than reading any pricing table, ours included — token counts for the same prompt vary by tokenizer, and that's what actually determines your bill.

Picking by task, not just price

• Cheapest general chat: qwen3-6-flash or deepseek-v4-flash for high-volume, latency-sensitive traffic.
• Best reasoning per dollar: deepseek-v4-pro or qwen3-max — both clear the "cheaper than official" bar and support thinking.
• Longest context: doubao-seed-2-pro, -code and -lite at 256,000 tokens, or deepseek-v4-pro/deepseek-v4-flash at exactly 1,000,000.
• Coding: doubao-seed-2-code for multi-file tasks with thinking; qwen3-coder-plus and qwen3-coder-turbo return once their upstream is fixed.

Getting started

Every model above is behind one API key and one prepaid balance, toppable with USDT today (card and SBP are being enabled). Live rates always match the pricing page; the full request schema is in the docs.