The Model Gap

Alibaba

Qwen3.8-Max

Eight of the ten benchmark scores tracked here, including the core coding results (SWE-bench Verified, LiveCodeBench, DeepSWE, Terminal-Bench 2.1), are independently confirmed — only the Toolathlon-Verified number behind Alibaba's agentic claim, and HLE with tools, are still self-reported. The hosted endpoint has served the Qwen3.8-Max-0902 snapshot since 2026-09-05; rows observed before that were earned by the 0803 build.

Qwen3.8-Max benchmarks and pricing, every number sourced: 10 tracked Qwen3.8-Max benchmark scores (8 independently run, 2 still resting on a vendor’s own claim), priced at $2.00 per million input tokens and $6.00 per million output.

Qwen3.8-Max’s 10 benchmark scores on this page were each verified against their sources between 2026-08-13 and 2026-09-29.

Released
2026-08-03
License
proprietary
Context window
1M tokens
Knowledge cutoff
Not disclosed
Parameters
Not disclosed
Architecture
Not disclosed

Qwen3.8-Max’s verified record

Qwen3.8-Max’s most-compared rival is Kimi K3: 2 trails, 2 ties, and 6 not callable across their 10 shared comparisons. Qwen3.8-Max is priced at $2.00/$6.00 per 1M tokens (in/out) vs Kimi K3’s $3.00/$15.00.

Against the 219 head-to-head comparisons Qwen3.8-Max shares with other tracked models: 55 real gaps, 42 inside the noise band, and 122 we will not call.

A gap counts for Qwen3.8-Max only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — Qwen3.8-Max trails on 39 of them.

HLE · no tools Reasoning

±2 is noise
Behind
Step 5 Preview −3.4 · Kimi K3 −3.8 · Gemini 3.1 Pro Preview −3.9 · Gemini 3.8 Flash −4.7 · GPT-6 Sol −4.8 · Muse Spark 1.3 −5.6 · MiMo-V2.6-Pro −6.3 · GPT-5.6 Sol −6.4 · GPT-6.1 Sol −9.8 · GPT-6 Astra −11.6 · Claude Opus 5 −11.8 · Claude Sonnet 5.5 −11.9 · Claude Fable 5.1 −16.0 · Claude Opus 5.5 −18.3 · 4 superseded: Muse Spark 1.2 −2.4 · Gemini 3.7 Flash −4.8 · Claude Opus 4.8 −5.6 · Claude Fable 5 −12.4
Ahead
DeepSeek V4 Pro (0813) +2.1 · GLM-5.3-Flash +3.2 · GPT-5.6 Luna +3.6 · GPT-6 Luna +4.6 · Qwen3.8-Flash-Next +5.1 · 2 superseded: GLM-5.2 +2.0 · DeepSeek V4 Flash (0731) +4.5
Tie
4 models within ±2
Unverified
2 models — vendor-reported on one side

LiveBench Composite score across 7 domains

±2.7 is noise
Behind
Muse Spark 1.3 −3.1 · GPT-6.1 Sol −3.1 · Claude Opus 5.5 −4.7 · Claude Fable 5.1 −4.9 · 1 superseded: Claude Fable 5 −4.5
Ahead
GPT-5.6 Luna +4.9 · GPT-6 Luna +6.5 · GLM-5.3-Flash +6.9 · MiniMax M3 +11.2 · 2 superseded: DeepSeek V4 Flash (0731) +4.3 · GLM-5.2 +5.3
Tie
15 models within ±2.7

DeepSWE Long-horizon coding

±9.5 is noise
Behind
Grok 4.6 −10.0 · GPT-5.6 Luna −10.0 · Kimi K3 −12.0 · GLM-5.3 −12.0 · GPT-5.6 Sol −16.0 · Claude Opus 5 −17.0 · Gemini 3.8 Flash −17.0 · GPT-6 Astra −17.0 · 1 superseded: Claude Fable 5 −13.0
Ahead
1 superseded: GLM-5.2 +13.0
Tie
6 models within ±9.5
Unverified
10 models — vendor-reported on one side
Setup-dependent
2 models — scored on a different harness

Agents' Last Exam Professional work

±3.2 is noise
Behind
GPT-5.6 Luna −3.3 · GPT-5.6 Sol −3.6 · Claude Opus 5 −5.2 · GPT-6 Sol −5.2 · Muse Spark 1.3 −5.2 · GPT-6 Astra −7.2 · Claude Opus 5.5 −11.2
Ahead
Gemini 3.1 Pro Preview +10.6 · 1 superseded: GLM-5.2 +6.6
Tie
4 models within ±3.2
Unverified
6 models — vendor-reported on one side

No verdict for Qwen3.8-Max anywhere on Terminal-Bench 2.1 (every independently confirmed comparison inside the noise band); GPQA Diamond, LiveCodeBench, SWE-bench Verified (saturated); HLE · with tools, Toolathlon-Verified (nothing independently confirmed on both sides).

Qwen3.8-Max API pricing

$2.00 in / $6.00 out per 1M tokens — official pricing

What Qwen3.8-Max costs per job

Qwen3.8-Max cost for three reference workloads, computed from its list rates
WorkloadTokens in / outCost
One long chat turn100K / 10K$0.260
A codebase review1,000K / 100K$2.60
A day of agent work10,000K / 1,000K$26.00

Computed from Qwen3.8-Max’s list rates above — cache discounts and batch tiers are not applied.

Qwen3.8-Max is one of 2 Alibaba models tracked on this site, at these official list prices.

Alibaba model pricing, official list rates
ModelIn / 1MOut / 1M
Qwen3.8-Flash-Next——
Qwen3.8-Max$2.00$6.00

Qwen3.8-Max benchmark scores

Qwen3.8-Max benchmark scores, provenance, and source links
BenchmarkScore
HLE(no tools)[1]
Reasoning · ±2 is noise
43.1artificialanalysis.ai · 2026-09-29
HLE(with tools)
Reasoning · ±2 is noise
56.2huggingface.co · 2026-08-17
GPQA Diamondsaturated[2]
Expert science Q&A — not ranked at any gap size
93.7vals.ai · 2026-08-17
Terminal-Bench 2.1[3]
Terminal ops · ±10.6 is noise
88.8artificialanalysis.ai · 2026-09-29
DeepSWE[4]
Long-horizon coding · ±9.5 is noise
57.0deepswe.datacurve.ai · 2026-08-13
Toolathlon-Verified[5]
Multi-tool chores · ±9.7 is noise
72.5huggingface.co · 2026-08-17
Agents' Last Exam[6]
Professional work · ±3.2 is noise
27.0snorkel.ai · 2026-08-17
SWE-bench Verifiedsaturated[7]
Bug fixing — not ranked at any gap size
85.6vals.ai · 2026-08-17
LiveCodeBenchsaturated[8]
Contest coding — not ranked at any gap size
87.8vals.ai · 2026-08-17
LiveBench[9]
Composite score across 7 domains · ±2.7 is noise
78.5livebench.ai · 2026-08-24

Who ran these numbers: 8 of 10 independent — vals.ai (3), artificialanalysis.ai (2), deepswe.datacurve.ai (1), snorkel.ai (1), livebench.ai (1); vendor self-reported (2).

  1. HLE: AA's own run of 'Qwen3.8 Max (0902)' (text-only, no tools; underlying 0.4310), the snapshot the hosted endpoint has served since 2026-09-05. AA's separate entry for the 0803 build (slug qwen3-8-max-0803) read 43.0, and Alibaba's self-reported launch figure for that build was 43.6 — all three agree closely.
  2. GPQA Diamond: vals.ai run, rank 5/133 (updated 2026-08-15). Replaces Alibaba's self-reported 92.6.
  3. Terminal-Bench 2.1: AA's own run of 'Qwen3.8 Max (0902)' at max effort (underlying 0.8876; Terminus 2 harness). Replaces Alibaba's self-reported 86.6 for the 0803 build (Claude Code harness, avg@10); AA's entry for the 0803 build reads 81.3. Not on tbench.ai's official board as of 2026-09-29. Cross-check (2026-10-01): vals.ai's archived Terminus 2 table lists 'Qwen 3.8 Max' at 67.42 without naming a build.
  4. DeepSWE: Independent score (57% ± 3%, $3.73/task) — debuted as the highest-scoring new entry on this leaderboard at the time.
  5. Toolathlon-Verified: Vendor-reported. Does not appear on the official Toolathlon-Verified leaderboard as of Aug 2026.
  6. Agents' Last Exam: Overall pass rate 27.0 (Claude Code, XHigh; score 52.5; $486). Corrects an earlier version of this page, which showed 52.5 — that was the partial-credit score, not the pass rate.
  7. SWE-bench Verified: vals.ai run, rank 13/83, bash-only harness; slowest of the top group at 41m42s/test (updated 2026-08-14).
  8. LiveCodeBench: vals.ai run, rank 8/138 (updated 2026-08-15).
  9. LiveBench: Board row "Qwen 3.8 Max" on the 2026-06-25 LiveBench release.

Notes on the record

The $2/$6 per-million-token price shown on this page is Alibaba Cloud's Singapore/international endpoint rate. China (Beijing) region pricing is lower ($1.65/$4.951 per 1M tokens) — and per Alibaba Cloud's pricing page (checked 2026-08-20), so is pricing on the Frankfurt (Germany), Virginia (US), Tokyo (Japan), and Hong Kong endpoints; only the Singapore endpoint bills at the higher $2/$6 rate, so "international" pricing is really a Singapore-specific surcharge rather than a blanket non-China rate.

Knowledge cutoff not officially disclosed by Alibaba. The proprietary hosted API tracked on this page is a separate artifact from Alibaba's open-weight release of the same underlying mixture-of-experts architecture (2.4 trillion parameters total, 95 billion active per token), published on Hugging Face as Qwen/Qwen3.8-2.4T-A95B under a custom "qwen3.8-max" license rather than Apache 2.0; that open release is text-only (no vision/video input), requires thinking mode for all interactions, and has a smaller native context window (262,144 tokens, extensible to ~1,010,000) than the hosted API's 1,000,000-token default (source: huggingface.co/Qwen/Qwen3.8-2.4T-A95B).

Snapshot (checked 2026-09-29): Alibaba released Qwen3.8-Max-0902 on 2026-09-02 — a post-training upgrade with the architecture, context window and price unchanged — and Model Studio's update notice moved the qwen3.8-max endpoint to that snapshot on 2026-09-05 10:00 UTC+8 with billing unchanged. This page tracks the endpoint, so it now describes the 0902 build. Artificial Analysis scores the two builds separately: HLE 43.0 → 43.1, GPQA Diamond 92.7 → 92.8, Terminal-Bench 2.1 81.3 → 88.8 (max effort). The HLE and Terminal-Bench 2.1 rows below are AA's 0902 runs; the vals.ai, DeepSWE, Agents' Last Exam and LiveBench rows were observed before the switch and belong to the 0803 build unless the source names the snapshot.

Compare with

Qwen3.8-Flash-NextGPT-6.1 SolClaude Sonnet 5.5GPT-6 Lunaany model…

FAQ

Has Qwen3.8-Max been independently benchmarked?

Of the ten scores tracked for Qwen3.8-Max on this site, eight come from independent evaluators — Artificial Analysis (HLE no tools and Terminal-Bench 2.1, both on the 0902 snapshot), Vals.ai (GPQA Diamond, SWE-bench Verified, LiveCodeBench), the DeepSWE leaderboard, Snorkel AI's Agents' Last Exam, and LiveBench's own board — while two are self-reported by Alibaba, sourced to the Qwen3.8-2.4T-A95B model card on Hugging Face: HLE with tools and Toolathlon-Verified. The core software-engineering benchmarks (SWE-bench Verified, LiveCodeBench, DeepSWE) all fall on the independent side, and Terminal-Bench 2.1 joined them when Artificial Analysis ran the 0902 build; Toolathlon-Verified is the one agentic number still vendor-reported.

Is Qwen3.8-Max open source or a proprietary model?

The Qwen3.8-Max tracked on this page is the pay-per-token model on Alibaba Cloud Model Studio — proprietary, not open weight. Alibaba separately published the underlying architecture on Hugging Face as Qwen/Qwen3.8-2.4T-A95B, but under a custom "qwen3.8-max" license rather than Apache 2.0 (unlike the smaller Qwen3.8-27B, which is Apache 2.0 — source: huggingface.co/Qwen/Qwen3.8-27B), and that release drops vision/video input, forces thinking mode on, and caps native context at 262,144 tokens versus the hosted API's 1,000,000-token default. Source: huggingface.co/Qwen/Qwen3.8-2.4T-A95B.

Does Qwen3.8-Max cost the same price in every region?

No. The $2 in / $6 out per-million-token price on this page is Alibaba Cloud's Singapore endpoint rate. Its Beijing endpoint bills $1.65 in / $4.951 out, and — per the same Alibaba Cloud pricing page, checked 2026-08-20 — the Frankfurt, Virginia (US), Tokyo, and Hong Kong endpoints bill at that same lower rate. Singapore is the outlier, not the norm: of Qwen3.8-Max's six live regions, it's the only one billing at the higher rate.

What is Qwen3.8-Max's knowledge cutoff date?

Alibaba has not published one. Neither the Model Studio documentation for the hosted Qwen3.8-Max API nor the Hugging Face card for the related open-weight Qwen3.8-2.4T-A95B checkpoint states a training-data cutoff, even though both list parameter counts, context limits, and benchmark tables in detail.

Why is Qwen3.8-Max also listed as Qwen3.8-2.4T-A95B on Hugging Face?

"2.4T-A95B" is Alibaba's shorthand for the mixture-of-experts design behind Qwen3.8-Max: 2.4 trillion parameters total, with 95 billion active per token. It's the repository name for the open-weight release of that architecture; the hosted, billed API tracked on this page is versioned separately as "Qwen3.8-Max." Source: huggingface.co/Qwen/Qwen3.8-2.4T-A95B.

Further reading

Get the next verdict by email

One email per verdict — which launch-chart claims held up. No spam, unsubscribe anytime.