Evals: find out which model does your team's work best

Join the waitlist

Qwen3.8 2.4T A95B

  • Tool calling
  • JSON schema
  • Reasoning
  • Streaming
  • Open weights
  • 2 providers

Model id

qwen/qwen3.8-2.4t-a95b
Context window
256K
tokens
Input price
$2.00
per 1M tokens
Output price
$6.00
per 1M tokens
Max output
Provider
no narrower limit

Provider list price

Pricing

USD / 1M tokens

Input$2.00 / M tokens
Output$6.00 / M tokens
Cache read$0.25 / M tokens
Context window256K tokens
RoutingFree

Provider list price. BYOK bills your provider; managed credits draw down at these rates. No markup or request fee. Every request is metered per engineer and joined to shipped work.

Registry 2026-11-09.1 · effective 9/17/2026 · Private rate card. Source withheld.

Providers

Qwen3.8 2.4T A95B runs on 2 providers. Automatic uses the cheapest one that is working, or you can choose. You pay the same list price whichever one serves it.

Provider for this request

No provider needed. TokenSpend picks the cheapest working one.

Automatic
curl https://api.tokenspend.dev/v1/chat/completions \
  -H "Authorization: Bearer $TOKENSPEND_API_KEY" \
  -H "content-type: application/json" \
  -d '{"model":"qwen/qwen3.8-2.4t-a95b","messages":[{"role":"user","content":"hello"}]}'

Tool calls and JSON schema only go to providers that have passed our tests for them. Only Novita has passed for tools so far, so tool requests have no fallback.

Registry contract

Capabilities

Protocols
3
Chat · Responses · Messages · streaming
Tools
2
Chat · Messages
JSON Schema
2
Chat · Responses
Vision
No
Not supported
Reasoning
1
Chat
Output limit
Provider
No narrower TokenSpend limit

canonical qwen/qwen3.8-2.4t-a95b · revision qwen/qwen3.8-2.4t-a95b · aliases qwen3.8-2.4t-a95b

First request

Run it

Python · TokenSpend SDK
from tokenspend import TokenSpend

client = TokenSpend()
resp = client.chat.completions.create(
    model="qwen/qwen3.8-2.4t-a95b",
    messages=[{"role": "user", "content": "hello"}],
)
TypeScript · TokenSpend SDK
import { TokenSpend } from "@tokenspend/sdk";

const client = new TokenSpend();
const resp = await client.chat.completions.create({
  model: "qwen/qwen3.8-2.4t-a95b",
  messages: [{ role: "user", content: "hello" }]
});
curl
curl https://api.tokenspend.dev/v1/chat/completions \
  -H "content-type: application/json" \
  -H "Authorization: Bearer $TOKENSPEND_API_KEY" \
  -d '{"model":"qwen/qwen3.8-2.4t-a95b","messages":[{"role":"user","content":"hello"}]}'
Claude Code integration
export ANTHROPIC_BASE_URL=https://api.tokenspend.dev
export ANTHROPIC_API_KEY=$TOKENSPEND_API_KEY
export ANTHROPIC_MODEL=qwen/qwen3.8-2.4t-a95b

One request shape reaches every model. Claude Code connects as an integration with two environment variables. TokenSpend never stores prompts or responses.

Questions

Which provider will run my Qwen3.8 2.4T A95B requests?

Novita and Together can all run it. By default TokenSpend sends each request to the cheapest one that is working and moves to the next if that one errors, so you do not have to pick. To pin a provider, send it in the request as the provider option; the code on this page shows how. Every provider bills the same list price.

What happens if a provider goes down mid-request?

A failed attempt, such as a timeout, a rate limit or a server error, is retried once on the next provider before any output reaches you. A stream that has already started is not restarted on another provider. If you pinned one provider, requests stay on it and fail with a clear error instead of moving.

What does Qwen3.8 2.4T A95B cost?

$2.00 per 1M input tokens and $6.00 per 1M output tokens. TokenSpend adds no markup and no per-request fee. If you pay with managed credits, buying them carries a 5% fee.

What can I do with Qwen3.8 2.4T A95B through TokenSpend?

Chat and streaming, plus tool calling (Chat, Messages), JSON schema output (Chat, Responses) and reasoning controls (Chat). It does not support image input here. Tool calling and JSON schema are enabled per provider, only after that provider passes our tests, so a provider chosen by hand may lack them.

How long can my prompt be?

256K tokens of context, prompt and reply together.

Does prompt caching make Qwen3.8 2.4T A95B cheaper?

Yes, when the provider caches a repeated prompt prefix. Cached input is billed at $0.25 per 1M tokens instead of $2.00.

How current is this pricing?

It follows rate card 2026-11-09.1, effective 9/17/2026. Source: Private rate card.