Qwen 3.8 Max
Alibaba's 2.4T-parameter MoE model, 95B active, aimed at agentic coding and long-horizon tool use at $2/$6 per million tokens.
Updated 2026-08-10
Yes, within limits. Qwen 3.8 Max runs a free tier with paid plans above it.
Listed pricing: $2 / $6 per 1M tokens (input/output); open-weights release announced, not yet confirmed.
Overview
Qwen 3.8 Max is Alibaba's frontier language model, announced on 3 August 2026 with the API live the same day. The architecture is a sparse mixture-of-experts: 2.4 trillion total parameters with roughly 95 billion active per token. Alibaba positions it around agentic work rather than chat, meaning long-horizon tool use, multi-step coding, and native multimodal input handled by the same checkpoint instead of a separate vision variant. The headline benchmark in the announcement is TerminalBench at 86.6, which is a vendor number.
The figure that actually settles the case is on the pricing page: $2 per million input tokens and $6 per million output tokens. Set that against the other frontier-tier options in this directory and the gap is wide. OpenAI's GPT-5.6 Sol lists $5/$30, and Moonshot's Kimi K3 lists $3/$15. Output tokens are where agentic loops burn budget, since a coding agent generates reasoning and tool calls across dozens of turns per task, so a 5x spread on the output side compounds fast at production volume. Anyone running a high-throughput agent fleet should price the swap before dismissing this as a second-tier option.
The honest take: treat the price as a fact and everything else as a claim awaiting evidence. A week after launch there is still no independent replication of TerminalBench 86.6, and the open-weights release Alibaba dated to the following week has not been confirmed as a download in any source we can check. The launch-day framing also oversold that commitment: open weights at this scale would not be unprecedented, because Kimi K3 already ships downloadable weights at 2.8T with a 1M-token context. Scale cuts against the self-hosting pitch anyway, since a 2.4T model at 95B active is inference-provider territory, not something a small team stands up on a whim. Two practical caveats before adoption: inference is China-hosted, which regulated buyers need to resolve first, and pricing here was last read on 3 August 2026, so confirm current rates on the vendor page. Note also that this is the model, not Alibaba's consumer Qwen App; coverage conflates the two constantly.
Is Qwen 3.8 Max free?
Yes, within limits. Qwen 3.8 Max runs a free tier with paid plans above it.
What the free tier covers: None announced for the API. If the open weights ship as described, self-hosting would carry no licence cost, but a 2.4T-parameter model means serious compute regardless of what the licence says.
Listed pricing: $2 / $6 per 1M tokens (input/output); open-weights release announced, not yet confirmed.
Pricing verified . That is the date the plans were last re-checked against the vendor's own pages, not today's date. Confirm on the official site before you pay.
What the free tier leaves out
Read straight off the plan list below. Vendors move features between tiers, so check the current split before you pay.
- API (pay-per-token) $2 per 1M input tokens / $6 per 1M output tokens Full model access including agentic tool use and multimodal input. No separate capability tiers were announced at launch. Rates were last verified on 3 August 2026 — check Alibaba Cloud Model Studio for current pricing and any volume or cached-input discounts.
- Open weights Announced for the week following the 3 August 2026 launch; release not confirmed Self-hosting rights under whatever licence Alibaba publishes. Licence terms, exact release date and hardware requirements were never specified in the launch announcement and remain unpublished as far as we can verify.
Qwen 3.8 Max pricing
| Plan | Price | What's included |
|---|---|---|
| API (pay-per-token) | $2 per 1M input tokens / $6 per 1M output tokens | Full model access including agentic tool use and multimodal input. No separate capability tiers were announced at launch. Rates were last verified on 3 August 2026 — check Alibaba Cloud Model Studio for current pricing and any volume or cached-input discounts. |
| Open weights | Announced for the week following the 3 August 2026 launch; release not confirmed | Self-hosting rights under whatever licence Alibaba publishes. Licence terms, exact release date and hardware requirements were never specified in the launch announcement and remain unpublished as far as we can verify. |
Full model access including agentic tool use and multimodal input. No separate capability tiers were announced at launch. Rates were last verified on 3 August 2026 — check Alibaba Cloud Model Studio for current pricing and any volume or cached-input discounts.
Self-hosting rights under whatever licence Alibaba publishes. Licence terms, exact release date and hardware requirements were never specified in the launch announcement and remain unpublished as far as we can verify.
Pricing verified . That is the date the plans were last re-checked against the vendor's own pages, not today's date. Confirm on the official site before you pay.
Is Qwen 3.8 Max worth it?
Worth it for Engineering teams running high-volume coding agents who are already paying frontier per-token rates and want to cut inference cost, assuming China-hosted inference clears their compliance review.
You can test that on the free tier before paying anything. The recorded trade-offs are listed below, and any one of them can settle the question on its own.
The 8.4/10 AI Score is an editorial read of published capability, price and shipping pace. Nobody here has hands-on hours with Qwen 3.8 Max. How we verify.
Worth it if
The strengths recorded against this entry.
- $2/$6 per million tokens undercuts every other frontier-tier model in this directory on list price (GPT-5.6 Sol $5/$30, Kimi K3 $3/$15)
- The output-token gap compounds hardest in agentic loops, which is exactly the workload Alibaba tuned it for
- API was live on announcement day rather than gated behind a waitlist
- Native multimodal input in the same checkpoint, with no separate vision model to route to
- Explicitly trained for terminal and repository work rather than retrofitted for it after the fact
Not worth it if
Any one of these blocks your use case.
- TerminalBench 86.6 and the rest of the benchmark table remain vendor-reported, with no independent replication a week after launch
- The open-weights release was dated to the week after launch and we cannot confirm it has happened; on that axis Kimi K3 already ships weights today
- 2.4T total parameters puts practical self-hosting out of reach for individuals and small teams even if weights do land
- China-hosted inference raises data-residency questions that regulated buyers must resolve before adoption
What sets Qwen 3.8 Max apart
- $2/$6 per million tokens is the lowest list price among the frontier-tier models tracked here, against GPT-5.6 Sol at $5/$30 and Kimi K3 at $3/$15
- 2.4T total parameters with ~95B active: the sparsity ratio, not a smaller model, is what carries the price down
- The launch led with TerminalBench rather than a general-knowledge benchmark, positioning it as an agent backend rather than a chat assistant
- Open weights are pledged but unconfirmed, so on openness it does not currently differentiate from Kimi K3, which ships downloadable weights at comparable scale
Key features
Sparse MoE architecture
2.4 trillion total parameters with about 95 billion active per token. The sparsity ratio is what makes the low per-token price plausible, since you pay for a fraction of the network on each forward pass rather than a smaller model overall.
Agentic coding focus
Alibaba led the announcement with TerminalBench (86.6, vendor-reported) rather than a general-knowledge benchmark, signalling the model is tuned for multi-step terminal and repository work. No third-party replication of that figure has appeared in the week since launch.
Long-horizon tool use
Built to hold state across extended tool-calling chains, the failure mode that separates usable coding agents from demos. Independent evidence on how far it holds up is still not available, so benchmark it on your own task before committing a fleet to it.
Open weights, pledged not shipped
Alibaba announced a weights release for the week following the 3 August 2026 API launch. That window has now arrived and no confirmed download, licence text or hardware guidance has surfaced in the sources we track. Judge this feature when the artefact exists.
How it compares
| Tool | Best for | Pricing | Score |
|---|---|---|---|
| Qwen 3.8 Max | Engineering teams running high-volume coding agents who are already paying frontier per-token rates and want to cut inference cost, assuming China-hosted inference clears their compliance review. | $2 / $6 per 1M tokens (input/output); open-weights release announced, not yet confirmed | 8.4/10 |
| Claude vs Claude → | Developers and technical teams who want a coding agent running in their own repo, terminal and IDE rather than a hosted editor, plus a chat assistant strong enough for long-form writing and 1M-token codebase analysis. | Free + Pro $17-20/mo + Max from $100/mo + Team $20-25/seat | 9.4/10 |
| GPT-5.6 Sol vs GPT-5.6 Sol → | Engineering teams building long-horizon coding or tool-use agents who need a frontier reasoning tier for the hard steps and can route easier calls to Terra or Luna on the same API key. | API usage-based: Sol $5/$30 per 1M tokens; cheaper Terra and Luna tiers cut July 30 | 9.3/10 |
| Gemini 3.7 Flash | — | Intro $0.75 input / $3.75 output per 1M tokens through Dec 31, 2026; then $1.50 / $7.50. Half the original 3.6 Flash token price during the intro window. | 8.9/10 |
Compare head-to-head
Related reading
Qwen3.8-Max: $2/$6 API, Open Weights Next Week
Alibaba priced Qwen3.8-Max at $2/$6 per million tokens on August 3 and dated the open-weight release to next week. What that sequence means for builders.
Qwen3.8-Max Pricing Lands, the Benchmarks Don't
Alibaba priced Qwen3.8-Max at $2/$6 per million tokens on August 3, two weeks after selling a preview on a ranking it never showed.
Claude Riemann Hypothesis Result: 41.6% to 67.2%
Anthropic says an unreleased Claude raised the proven fraction of Riemann zeta zeros on the critical line from 41.6% to 67.2%. Here is what that means.
Ready to try Qwen 3.8 Max?
Head to the official site to start with Qwen 3.8 Max — pricing and plans are listed above.
Visit Qwen 3.8 Max