Owen

Posted on Aug 9 • Originally published at ofox.ai

Qwen 3.8 Max: Price, API Access, and Open Weights (2026)

#ai #qwen #pricing #opensource

Qwen 3.8 Max is out. Alibaba shipped it on August 3, 2026, two weeks after the preview. This is the release where the numbers stopped being rumours.

Released: August 3, 2026 (preview announced July 19, 2026)
Price: $2.00 / 1M input, $6.00 / 1M output on QwenCloud
Context: 1M window (991K max input, 131K max output)
Size: 2.4T parameters, 95B active
Open weights: announced for next week, no license or date published yet
Access today: QwenCloud API, Qwen Studio, QwenWork
On ofox: yes, bailian/qwen3.8-max, OpenAI and Anthropic protocols

TL;DR

The new flagship is cheaper than the old one: $2/$6 against Qwen 3.7 Max's $2.50/$7.50.
Max output doubled to 131K tokens; the 1M context window carried over.
Alibaba will open-source a Max-class model for the first time, along with Qwen 3.8-27B, "next week" (announced, not dated).
The official benchmark table puts it ahead of Claude Opus 4.8 on most agentic rows and behind Fable 5 on SWE-bench Pro and HLE. It is a vendor table.
No Kimi K3 column in that table, which is the comparison most people wanted.
It is already on ofox as bailian/qwen3.8-max, over both the OpenAI and Anthropic protocols, so it drops into Claude Code without waiting for Alibaba to open a Claude-shaped endpoint.

What Is Qwen 3.8 Max?

Qwen 3.8 Max is Alibaba's top-end proprietary model, a 2.4-trillion-parameter multimodal system with 95 billion parameters active per token. It is built on the Qwen 3.5 architecture rather than a new one, and Alibaba positions it around long-horizon autonomous work rather than single-turn chat.

Inputs and outputs:

Feature	Details
Supported Input modalities	Text, image, video (QwenCloud); text and image on ofox, video not listed
Output modalities	Text
API features	Function calling, structured outputs, context cache, prefix completion
Rate limits (QwenCloud)	2M tokens/min, 15K requests/min

The launch post leans hard on multi-day autonomy. Its headline demonstration is an unattended coding run that produced 265 commits, 127 PRs and 151 issues over roughly 16 days in a public repository. Whether that generalises to your codebase is the one thing a launch post cannot tell you.

When Was Qwen 3.8 Max Released?

August 3, 2026. The preview came first, and the two dates get conflated in coverage:

July 19, 2026. Qwen3.8-Max-Preview announced at WAIC in Shanghai. Reachable only via Alibaba's Token Plan subscription, Qoder, and QoderWork. No pay-as-you-go API, no model card, no benchmark table.
August 3, 2026. General availability. QwenCloud API opens, prices published, full benchmark table published, open-weights timing announced.

Any piece dated between those two has stale "no public pricing" and "no benchmarks" claims, including anything citing the 10%-of-standard preview discount, which applied only to the Token Plan window.

Is Qwen 3.8 Max Open Source?

Not today. Alibaba says the weights ship next week, the week of August 10, 2026.

What was actually announced:

Qwen 3.8 Max weights will be released, described as the first time a Qwen-Max-class model gets open weights.
Qwen 3.8-27B goes open-weights in the same drop, which is the one most people can actually run.
No license named. No exact date. No model card or activated-parameter breakdown for the open release.

Three things to hold loosely until the drop lands. A vendor's "next week" is a plan, not a shipping date. The license matters more than the timing for commercial use. Qwen's open releases have historically shipped under Apache 2.0 (Qwen3-235B-A22B still carries apache-2.0 on Hugging Face), but that is precedent, not a commitment for a Max-class release. And a 2.4T model with 95B active is not a workstation self-host; the 27B is the one with a realistic hardware story.

How Much Does Qwen 3.8 Max Cost?

$2.00 per 1M input tokens and $6.00 per 1M output tokens on QwenCloud, as of the August 3, 2026 snapshot. The new flagship is 20% cheaper than the model it replaces.

Pricing	Qwen 3.8 Max (QwenCloud)	Qwen 3.7 Max (Model Studio, Singapore)
Input / 1M	$2.00	$2.50
Output / 1M	$6.00	$7.50
Implicit cache read / 1M	$0.25	n/a
Explicit cache write / 1M	$2.50	n/a
Explicit cache read / 1M	$0.17	n/a

Two notes on that table. The cache rows come from different price sheets and are not directly comparable. ofox's Qwen 3.7 Max page lists cache read at $0.50/M and cache write at $3.125/M, which is a gateway's listing rather than Alibaba's. And Model Studio's Singapore sheet was showing a limited-time 50% discount on 3.7 Max at the time of the check, so the effective gap can be smaller than list-to-list suggests.

Prices on both sides of this table move. Pull the current number before you build a cost model on it.

What's the Context Window and Max Output?

1M tokens total, split as 991K max input and 131K max output, with a separate 262K reasoning ceiling.

Context window: 1M
Max input: 991K
Max output: 131K
Max reasoning tokens: 262K

The output ceiling is the real change. ofox lists Qwen 3.7 Max at 64K max output on the same 1M window, so 3.8 Max doubles what it can emit per response. That matters for the workloads Alibaba targets: a model that plans across 262K reasoning tokens and then has 64K to write the artefact is a model that truncates.

What Do the Official Benchmarks Say?

Alibaba published a full comparison table at launch, and it is a mixed picture rather than a sweep. Every number below is vendor-reported.

Selected rows, Qwen 3.8 Max against the frontier set:

Benchmark	Qwen 3.8 Max	Opus 4.8	Fable 5	GPT-5.6 Sol (max)	Qwen 3.7 Max
Terminal Bench 2.1	86.6	84.6	84.6	88.8	74.5
SWE-bench Pro	67.7	69.2	80.0	64.6	60.6
PaperBench	93.0	80.3	88.8	90.5	64.8
GPQA Diamond	92.6	92.0	92.6	94.1	92.4
HLE	43.6	45.7	53.3	47.2	41.4
IFBench	82.8	62.2	63.5	72.7	79.1
OSWorld-Verified	86.1	83.4	85.0	83.2	73.3
MMMU-Pro	82.3	75.6	81.2	83.0	79.0

Read it with the usual discount. The competitor Terminal Bench figures are lifted from Artificial Analysis and OpenAI's own posts rather than re-run by Qwen, and the SWE-bench Pro row used the Claude Code harness at temp 1.0, top_p 0.95, 256K context, a configuration Qwen chose. Several rows are Qwen's own in-house benchmarks (QwenSWEBench, QwenQoderBench, QwenReactBench), where the model that shares the name is the one being measured.

The gap in the table is louder than anything in it: there is no Kimi K3 column. Moonshot's open-weight flagship shipped days before the Qwen 3.8 preview and is the model Qwen 3.8 Max is most often asked about. Alibaba benchmarked against three closed Western models and its own predecessor instead.

How Do I Access Qwen 3.8 Max?

Through QwenCloud today. Three first-party routes exist, and one widely-reported fourth needs checking:

QwenCloud API. Model ID qwen3.8-max, the developer path. Keys from home.qwencloud.com/api-keys.
Qwen Studio. chat.qwen.ai, the chat interface.
QwenWork. The workplace agent platform, for non-developers.
Alibaba Cloud Model Studio. SCMP reported global API access on August 3. Model Studio's English model list and price sheet still topped out at qwen3.7-max on our check the same day, so the catalog pages lag the announcement even though the endpoint is shared.

That last point is worth unpacking, because the branding hides it. QwenCloud's own developer guides call the OpenAI-compatible endpoint at https://dashscope-intl.aliyuncs.com/compatible-mode/v1 and read the key from DASHSCOPE_API_KEY, the same DashScope international host Model Studio uses. QwenCloud is a front end on that infrastructure, not a separate API.

The call is OpenAI-shaped:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.getenv("DASHSCOPE_API_KEY"),
    base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
)

resp = client.chat.completions.create(
    model="qwen3.8-max",
    messages=[{"role": "user", "content": "Summarize this repo's test strategy."}],
    max_tokens=4096,
)
print(resp.choices[0].message.content)

The Route Alibaba Does Not Offer: Anthropic Protocol

Qwen ships an OpenAI-shaped API only, so Claude Code cannot call qwen3.8-max directly. Claude Code speaks the Anthropic Messages protocol and will not negotiate. A gateway closes that gap on day one.

bailian/qwen3.8-max went live on ofox the same day as the launch. A catalog call on August 3, 2026 returned 123 models with the new ID present, and the model page lists it at $2/M input, $6/M output, cache read $0.25/M, cache write $2.50/M, with a 1,131,072-token context length and a 131,072-token completion ceiling. Prices match QwenCloud list on all four rows.

Worth flagging, because it nearly caught us out. The catalog's supported_endpoints field lists only /v1/chat/completions and /v1/responses, which reads like the Anthropic route is closed. It is not: a live Messages call on August 3, 2026 came back well-formed, stop_reason: end_turn. Send one real request before you plan around that field.

Two lines of shell put it inside Claude Code:

export ANTHROPIC_BASE_URL=https://api.ofox.ai/anthropic
export ANTHROPIC_MODEL=bailian/qwen3.8-max

Or keep the Python above and change two arguments:

client = OpenAI(
    api_key=os.getenv("OFOX_API_KEY"),
    base_url="https://api.ofox.ai/v1",
)
resp = client.chat.completions.create(model="bailian/qwen3.8-max", ...)

Which route to take:

Consideration	Go direct to QwenCloud	Go through a gateway
Lowest possible per-token price	Yes, this is list	Matches list here, verify per model
Day-zero access to Qwen-only features	Yes	Lags by however long the gateway takes
Claude Code / Anthropic-protocol tools	Not possible today	Yes
One key across Qwen, Claude, GPT, Gemini	No	Yes
Failover when one vendor rate-limits you	You build it	Built in

The honest caveat: a gateway adds a hop you do not control, and its price sheet can lag a mid-cycle vendor change. Read the model page, not the marketing page, on ours as much as anyone else's.

Qwen 3.8 Max vs Qwen 3.7 Max

Cheaper, twice the output ceiling, and a large jump on agentic benchmarks.

Comparison	Qwen 3.8 Max	Qwen 3.7 Max
Released	2026-08-03	2026-05-19
Parameters	2.4T (95B active)	Not disclosed
Context / max output	1M / 131K	1M / 64K (ofox listing)
Input / output per 1M	$2.00 / $6.00	$2.50 / $7.50
Terminal Bench 2.1	86.6	74.5
SWE-bench Pro	67.7	60.6
On ofox	Yes, `bailian/qwen3.8-max`	Yes, `bailian/qwen3.7-max`

Pick 3.8 Max unless you have a reason not to. It is cheaper, writes twice as much per response, and both are reachable with the same key, so the migration is a model-ID swap rather than an integration. Stay on 3.7 Max only if you have pinned a version for reproducibility or your evals are calibrated against it. The price gap is 20%, not an order of magnitude, so cost alone will not force the move. The 3.7 Plus vs 3.7 Max benchmark comparison covers the cheaper sibling if cost is the deciding factor.

For the cross-vendor question this launch reopens, the Kimi K3 vs GPT-5.5 and Opus 4.8 comparison has the open-weight side of it, and how to use Kimi K3 covers access. Both are on ofox today, K3 as moonshotai/kimi-k3 and this one as bailian/qwen3.8-max, so comparing them is a model-ID swap rather than a procurement decision.

FAQ

When did Qwen 3.8 Max come out? August 3, 2026, as a general-availability release with published pricing and benchmarks. A preview build was announced July 19, 2026, limited to Alibaba's Token Plan, Qoder and QoderWork.

Is Qwen 3.8 Max free? No. It is priced per token on QwenCloud at $2/$6 per 1M. The July preview ran at a reported 10% of standard pricing for Token Plan subscribers; that window is over.

Can I download Qwen 3.8 Max weights? Not yet. Alibaba announced an open-weights release for the week of August 10, 2026, without a license or exact date. Qwen 3.8-27B is in the same announced drop and is the realistic self-host target.

Does Qwen 3.8 Max support vision? Yes for images, with a caveat on video. The QwenCloud model page lists text, image and video as input modalities, and Alibaba's multimodal table reports 82.3 on MMMU-Pro and 86.1 on OSWorld-Verified. Through ofox, an image sent on August 3, 2026 was described correctly and raised prompt tokens from 60 to 90, so image input is live there. Video is not in the ofox catalog entry, so do not assume it.

Can I use Qwen 3.8 Max with Claude Code? Not against QwenCloud directly, because Claude Code speaks the Anthropic Messages protocol and QwenCloud is OpenAI-shaped. Through a gateway that speaks both it works on day one, tested on August 3, 2026: point ANTHROPIC_BASE_URL at https://api.ofox.ai/anthropic and set ANTHROPIC_MODEL=bailian/qwen3.8-max.

What is the difference between Qwen 3.8 Max and Qwen 3.8-27B? Max is the 2.4T proprietary flagship available now via API. The 27B is a much smaller model announced for the open-weights drop next week. They share a generation, not a size class.

Is Qwen 3.8 Max better than Kimi K3? Alibaba did not publish that comparison. There is no K3 column in its benchmark table. The two are also not comparable on availability: K3 has open weights and broad gateway support today, 3.8 Max does not.