Qwen4-Max vs Qwen3.8-Max: Complete Comparison and Report on Pricing, Cost per Task, Benchmarks, and Reasoning Effort

Qwen4-Max is the flagship model Alibaba released on November 16, 2026, a sparse mixture-of-experts system with 2.4 trillion total parameters and about 71 billion active per token, sold through QwenCloud and Alibaba Cloud Model Studio with a one-million-token context window. Qwen3.8-Max is the model it replaces at the top of the Qwen line: released on August 3, 2026, with the same 2.4 trillion total parameters but 95 billion active, and refreshed on September 2 with the qwen3.8-max-0902 snapshot. Both remain on Alibaba's official price list, with the same model IDs callable today, so developers who built on the older model now face a real choice between staying put and migrating.
The verdict favors Qwen4-Max for nearly every API workload. On the International (Singapore) deployment it costs $1.50 per million input tokens and $4.50 per million output tokens, 25% less than Qwen3.8-Max at $2 and $6, and its cache read price of $0.075 is 70% below the older model's $0.25 implicit cache rate. Independent testing points the same way: Vals AI scores Qwen4-Max at 63.20% on the Vals Index against 59.05% for Qwen3.8-Max 0902, and Artificial Analysis gives it 43 points on its Intelligence Index against 34, at $3.27 per index task against $10.02. Qwen3.8-Max keeps three practical advantages: a priced batch mode at half the real-time rate, an open-weight sibling checkpoint on Hugging Face for teams that must self-host, and stronger Vals results on its legal and tax benchmarks.
For a team running a coding agent or a customer-support assistant on Alibaba's API, Qwen4-Max is the better default, and the savings grow with heavy prompt caching. For offline batch jobs that tolerate delay, or for organizations that require weights they can run on their own hardware, Qwen3.8-Max still has a case. All prices below were verified on Alibaba Cloud's official Model Studio pricing page and the QwenCloud model pages on November 17, 2026.
··········
QWEN4-MAX AND QWEN3.8-MAX RELEASE TIMELINE AND MODEL IDENTITY.
The two models share a total parameter count and a context length, which makes the comparison unusually direct. Alibaba describes Qwen4-Max as a new architecture generation, with the mixture-of-experts design upgraded across attention, residual connections and embeddings, and says it activates 25% fewer parameters per token than its predecessor at the same 2.4 trillion scale. The QwenCloud changelog lists Qwen4-Max on November 16, five days after a Qwen4-Max-Preview opened on November 11 for Token Plan subscribers only.
Qwen3.8-Max arrived on August 3 as what the changelog called "Qwen's most capable flagship model to date", with hybrid thinking enabled by default. The qwen3.8-max-0902 snapshot followed on September 2, described by Alibaba as an upgraded snapshot with better coding, multi-tool orchestration and vision. On October 8 the base model gained a US deployment scope in the Virginia region. Independent evaluators now benchmark the 0902 snapshot, and Artificial Analysis labels it deprecated in its own tracking, even though Alibaba's documentation carries no retirement notice for either Qwen3.8-Max ID.
........
Item | Qwen4-Max | Qwen3.8-Max |
API model ID | qwen4-max | qwen3.8-max, qwen3.8-max-0902 |
Release date | November 16, 2026 | August 3, 2026 (0902 snapshot: September 2, 2026) |
Architecture | Sparse MoE, Qwen4 design | MoE built on the Qwen 3.5 foundation |
Total parameters | 2.4 trillion | 2.4 trillion |
Active parameters | About 71 billion | 95 billion |
Input modalities | Text, image, video | Text, image, video |
Output | Text | Text |
Open weights | No (smaller Qwen4 models promised) | Sibling checkpoint Qwen3.8-2.4T-A95B on Hugging Face |
Retirement notice | None | None on Alibaba's pages |
........
The parameter figures come from Alibaba's November 16 press release for Qwen4-Max and from the August 3 announcement and Hugging Face model card for Qwen3.8-Max. The 71 billion versus 95 billion active parameters is the hardware-level reason Alibaba gives for the lower price: fewer active parameters per token means less compute per generated token at the same model size.
··········
QWEN4-MAX VS QWEN3.8-MAX API PRICING BY REGION.
Alibaba prices both models in a single band from zero to one million input tokens, with no surcharge for long prompts and the same rate for thinking and non-thinking output. The price gap depends on where the model is deployed. On the International scope served from Singapore, the cut is a flat 25% on both input and output. On the Global scope, which Alibaba lists for Frankfurt, Virginia, Tokyo and Hong Kong, the reduction is about 14%, from $1.65 to $1.413 on input and from $4.951 to $4.24 on output.
........
Deployment scope (region) | Qwen4-Max input | Qwen4-Max output | Qwen3.8-Max input | Qwen3.8-Max output | Change |
International (Singapore) | $1.50 | $4.50 | $2.00 | $6.00 | -25% |
Global (Frankfurt, Virginia, Tokyo, Hong Kong) | $1.413 | $4.24 | $1.65 | $4.951 | -14.4% |
US scope (Virginia) | Not listed | Not listed | $2.00 | $6.00 | No Qwen4-Max US-scope row |
Chinese mainland (Beijing), USD list | $1.413 | $4.24 | $1.65 | $4.951 | -14.4% |
........
Prices are USD per million tokens as listed on the Model Studio pricing page, last updated November 16, 2026, and verified on November 17. The qwen3.8-max-0902 snapshot carries the same rates as the base Qwen3.8-Max ID in every region. One gap is visible in the table: Qwen3.8-Max has a dedicated US deployment scope in Virginia, which keeps processing inside the United States, while the Qwen4-Max documentation lists Virginia only under the Global scope. Companies with a US data-residency requirement should confirm the scope before migrating, because the Global scope is the one Alibaba uses across several countries.
··········
CACHE, BATCH AND FREE-QUOTA TERMS FOR THE TWO QWEN MAX MODELS.
Caching is where the price gap widens most. Qwen4-Max charges $0.075 per million tokens for both implicit and explicit cache reads on the International scope, which Artificial Analysis describes as a 95% discount on the input price. Qwen3.8-Max charges $0.25 for implicit cache hits and $0.17 for explicit cache reads in Singapore, so a long agent loop that re-reads the same context on every turn pays between 56% and 70% less for those reads on the newer model.
Batch processing reverses the picture. Model Studio's model page for Qwen4-Max lists batch inference as unsupported, and the pricing page shows no batch row for it. Qwen3.8-Max has batch pricing: QwenCloud lists batch file input at $1 and output at $3 per million tokens, and Model Studio shows a 50% batch discount in the Beijing region, at $0.825 and $2.475. The QwenCloud page for Qwen4-Max includes "batch processing at reduced cost" in its general feature list but publishes no batch price, so the Model Studio documentation is the safer reference until Alibaba prices it.
........
Term (International scope unless noted) | Qwen4-Max | Qwen3.8-Max |
Implicit cache hit | $0.075 | $0.25 |
Explicit cache read | $0.075 | $0.17 |
Explicit cache write, 5 minutes | $1.875 | $2.50 |
Explicit cache write, 1 hour | $3.00 | Not listed |
Batch input / output | Not supported on Model Studio | $1.00 / $3.00 (QwenCloud); $0.825 / $2.475 (Beijing) |
Free quota | 1 million tokens, 90 days | 1 million tokens (0902 snapshot: none) |
........
The one-hour cache option is new with Qwen4-Max. A five-minute cache suits chat sessions with frequent turns, while the one-hour version, at $3 per million tokens written, fits agents that pause for tool calls, human review or slow external systems and would otherwise rebuild their context after the shorter cache expires.
··········
CONTEXT WINDOW, OUTPUT LIMITS AND REASONING EFFORT CONTROLS.
Both models accept up to one million tokens of context, with a maximum input of 991,808 tokens without thinking and 983,616 with thinking. The difference is on the output side. Qwen4-Max doubles the maximum output in thinking mode to 262,144 tokens, where Qwen3.8-Max stays at 131,072 in both modes. For agents that write long files, full reports or large code diffs in a single response, that is the most concrete capability change between the two.
Reasoning effort is configured differently. Qwen4-Max exposes a reasoning_effort parameter with three values, low, high and xhigh, and defaults to high, with a default thinking budget of 131,072 tokens. Alibaba's Qwen3.8-Max announcement lists xhigh as the default, with medium and low as the alternatives. A team that migrates without setting the parameter therefore moves from the older model's highest effort level to the newer model's middle one, which lowers token use by default and can change answer quality on hard tasks.
........
Limit or control | Qwen4-Max | Qwen3.8-Max |
Context window | 1,000,000 tokens | 1,000,000 tokens |
Max input (non-thinking / thinking) | 991,808 / 983,616 | 991,808 / 983,616 |
Max output, non-thinking | 131,072 | 131,072 |
Max output, thinking | 262,144 | 131,072 |
Max chain of thought | 262,144 | 262,144 |
Effort levels | low, high, xhigh | low, medium, xhigh |
Default effort | high | xhigh |
Thinking default | On | On |
Rate limit, Global scope | 30,000 RPM, 5,000,000 TPM | 30,000 RPM, 5,000,000 TPM (0902: 150,000 TPM) |
........
The rate limits come from the Model Studio model pages. Qwen4-Max lists monthly-spend tiers that raise the tokens-per-minute ceiling to 10 million above $8,000 of monthly consumption and to 20 million above $140,000. The 0902 snapshot of Qwen3.8-Max is capped at 150,000 tokens per minute outside Beijing and Singapore, which is low for production agents, so teams pinned to that snapshot gain headroom simply by switching.
··········
ACCESS: QWENCLOUD, MODEL STUDIO REGIONS, TOKEN PLAN AND OPEN WEIGHTS.
Both models are served from the same six Model Studio regions: Beijing, Singapore, Frankfurt, Virginia, Tokyo and Hong Kong. Both accept requests through OpenAI-compatible Chat Completions and Responses endpoints and through Anthropic's Messages format, which lets developers plug them into Claude Code, Codex and similar coding tools by changing a base URL. Alibaba also offers Qwen4-Max inside its Token Plan subscription and in its QwenWork and QwenStudio applications.
........
Access route | Qwen4-Max | Qwen3.8-Max |
QwenCloud API | Yes | Yes |
Model Studio regions | Beijing, Singapore, Frankfurt, Virginia, Tokyo, Hong Kong | Same six regions |
US deployment scope | Not listed | Yes, Virginia (since October 8) |
OpenAI Chat Completions and Responses | Yes | Yes |
Anthropic Messages format | Yes | Yes |
Token Plan subscription | Yes | Not stated for this model |
Web search tool | Yes | Base ID: Beijing and Singapore; 0902: all regions |
Fine-tuning on Model Studio | Not supported | Not supported |
Downloadable weights | None | Qwen3.8-2.4T-A95B (different checkpoint) |
........
The open-weight row needs a precise reading. The Hugging Face card for Qwen3.8-2.4T-A95B states that Qwen3.8-Max is "the official version based on Qwen3.8-2.4T-A95B with more features", adding vision input, non-thinking support, a one-million-token default context and built-in tools. The downloadable checkpoint has a native context of 262,144 tokens, extensible to 1,010,000, and requires thinking mode. It is a related checkpoint, and Alibaba does not present it as identical to the API version. Qwen4-Max has no downloadable weights at all; Alibaba says only that smaller open-weight Qwen4 models will follow, without a date.
The QwenCloud page for Qwen4-Max lists fine-tuning among its features while the Model Studio model page marks fine-tuning and model deployment as unsupported. The same kind of conflict appears for Qwen3.8-Max. Until Alibaba aligns the two pages, the Model Studio documentation, which gives per-region detail, is the more conservative source for planning.
··········
HOW THE QWEN4-MAX AND QWEN3.8-MAX BENCHMARK FIGURES WERE PRODUCED.
Three kinds of evidence exist for this pair, and they carry different weight. The first is Alibaba's own material. The Qwen4-Max announcement gives no numeric score on any public benchmark in its text: it names Agents' Last Exam, Vals Index, JobBench and Chartography as areas of strong results and reports a sixth place on the Code Arena leaderboard without a score. Its only direct comparison with Qwen3.8-Max uses Work360, an internal benchmark covering 721 specialized fields, reported in relative terms. The Qwen3.8-Max model card on Hugging Face, by contrast, publishes a full table of vendor scores, which has no Qwen4-Max counterpart.
The second kind is independent evaluation on identical test suites. Vals AI ran both Qwen4-Max and the Qwen3.8-Max 0902 snapshot through the same versions of its benchmarks, which makes those rows directly comparable. The settings differ by model: Vals ran Qwen4-Max at temperature 1, xhigh reasoning effort and a 262,144-token output cap, and ran the 0902 snapshot at temperature 0.7 with a 128,000-token cap and no stated effort level. Artificial Analysis ran both through version 5.0.1 of its Intelligence Index, a composite of nine evaluations, and reports token use, cost and speed from the same runs.
The third kind is the absence of evidence. No independent organization has yet published a head-to-head on long-context retrieval, vision or safety for these two models, and Alibaba has not released a system card for Qwen4-Max.
··········
[[ADPLUS_300x250_1]]
VALS AI SCORES FOR QWEN4-MAX AND QWEN3.8-MAX 0902.
On the Vals Index, Vals AI's composite of professional tasks, Qwen4-Max scores 63.20% and ranks 11th of 38 models, against 59.05% and 18th place for the Qwen3.8-Max 0902 snapshot. The gains are concentrated in agentic and coding work: Terminal-Bench 4.0 rises from 38.89% to 48.48%, Finance Agent v2.1 from 58.83% to 66.68%, and Vibe Code Bench v1.1 from 77.14% to 86.66%.
........
Vals AI benchmark | Qwen4-Max | Qwen3.8-Max 0902 | Difference (points) |
Vals Index | 63.20% (11/38) | 59.05% (18/38) | +4.15 |
Vals Finance Index | 61.99% | 61.66% | +0.33 |
Vals Legal Index | 70.71% | 73.90% | -3.19 |
Terminal-Bench 4.0 | 48.48% | 38.89% | +9.59 |
Finance Agent v2.1 | 66.68% | 58.83% | +7.85 |
Vibe Code Bench v1.1 | 86.66% | 77.14% | +9.52 |
HLAB-Vals-1.1 | 21.25% | 15.42% | +5.83 |
EMB v1.1 | 70.83% | 71.98% | -1.15 |
Legal Research Bench | 41.35% | 46.15% | -4.80 |
Tax Agent Bench | 40.35% | 49.66% | -9.31 |
Cost per Vals Index test | $9.33 | $8.73 | +6.9% |
Average latency | 64 min 39 s | 105 min 46 s | 39% faster |
........
The table also shows where the older model holds up. Qwen3.8-Max 0902 leads by 9.31 points on Tax Agent Bench, by 4.80 points on Legal Research Bench and by 3.19 points on the Vals Legal Index. On Tax Agent Bench the gap exceeds the two published confidence intervals combined (±3.60 and ±3.54 points), while the legal gaps fall within the combined intervals and should be read as tendencies. Vals' cost per test is slightly higher for Qwen4-Max, at $9.33 against $8.73, a figure produced at the xhigh effort setting used in its runs, and its average latency is 39% lower.
··········
ARTIFICIAL ANALYSIS INDEX, TOKEN USE AND COST PER TASK.
Artificial Analysis places Qwen4-Max at 43 on its Intelligence Index, 38th of 81 models and above the comparable-model median of 41. The Qwen3.8-Max 0902 snapshot scores 34, ranked 51st. The cost data from the same runs are more striking than the score gap: running the index cost $3.27 per task on Qwen4-Max against $10.02 on Qwen3.8-Max, a 67% reduction.
........
Artificial Analysis metric (Index v5.0.1) | Qwen4-Max | Qwen3.8-Max 0902 |
Intelligence Index | 43 (#38 of 81) | 34 (#51 of 81) |
Cost per Intelligence Index task | $3.27 | $10.02 |
Output tokens to run the index | 127 million | 150 million |
Input tokens to run the index | 9.9 billion | 14.5 billion |
Blended price (7:2:1 cache hit, input, output) | $0.80 | $1.18 |
Output speed | 54 tokens/s | 61 tokens/s |
Time to first answer token | 159.36 s | 102.66 s |
Time per index task | 32.1 min | 47.0 min |
........
The token counts explain most of the cost gap. Qwen4-Max used 15% fewer output tokens and about 32% fewer input tokens to complete the same index, and each token cost less. Artificial Analysis still rates both models as very verbose, with output token use well above the 78.6 million median of comparable models. Alibaba's own claim, a roughly 60% lower average cost per task on Work360, is consistent in direction with this independent figure, which is higher at 67% on a different test set.
··········
ALIBABA'S OWN CLAIMS: WORK360, CODE ARENA AND THE QWEN3.8-MAX MODEL CARD.
Alibaba states that Qwen4-Max outperforms Qwen3.8-Max on Work360 while using 35% fewer tokens in high-effort reasoning mode. The company publishes no absolute Work360 scores, no per-field breakdown and no description of how tasks are graded, so the claim cannot be checked against the model's results elsewhere. The Code Arena sixth place is reported without a score or a date.
The Qwen3.8-Max card on Hugging Face gives a fuller vendor picture for the older model, which helps readers understand what Qwen4-Max is being measured against. The figures below are Alibaba's, with the settings it states.
........
Qwen3.8-Max vendor benchmark | Score | Setting stated by Alibaba |
SWE-bench Pro | 67.7 | Claude Code, temperature 1.0, 256K context |
Terminal Bench 2.1 | 86.6 | Claude Code, avg@10, 5-hour timeout |
GPQA Diamond | 92.6 | Not stated |
HLE (without tools / with tools) | 43.6 / 56.2 | Not stated |
SkillsBench v1.1 | 70.2 | OpenCode, average of 3 runs |
Toolathlon Verified (Pass@1) | 72.5 | Not stated |
MRCR v2 256K (8-needle) | 92.9 | Not stated |
Agents' Last Exam (pass / score) | 27.0 / 52.4 | Not stated |
........
Alibaba has not published Qwen4-Max numbers for any of these rows in text form, so no vendor-to-vendor comparison is possible on SWE-bench Pro, GPQA Diamond or HLE. Readers comparing the two models on those benchmarks will find only the Qwen3.8-Max side.
··········
WHAT THREE TYPICAL WORKLOADS COST ON QWEN4-MAX AND QWEN3.8-MAX.
To turn the price lists into budgets, the calculation below uses International-scope list prices and three monthly workloads. A customer-support assistant handles 10,000 requests a day for 30 days, each with 3,000 input tokens, of which 60% hit the implicit cache, and 400 output tokens. A long-document pipeline processes 200 documents a day of 400,000 tokens each with 3,000-token summaries and no caching. A coding agent runs 1,000 sessions a month of 40 turns, each turn reading a 120,000-token context with 90% cached and writing 1,500 tokens. The first run assumes both models produce the same number of tokens; a second Qwen4-Max column applies Alibaba's claim of 35% fewer output tokens.
........
Monthly workload | Qwen3.8-Max | Qwen4-Max, same tokens | Saving | Qwen4-Max, 35% fewer output tokens | Saving |
Support assistant (900M input, 120M output) | $1,575.00 | $1,120.50 | 28.9% | $931.50 | 40.9% |
Long documents (2.4B input, 18M output) | $4,908.00 | $3,681.00 | 25.0% | $3,652.65 | 25.6% |
Coding agent (4.8B input, 60M output) | $2,400.00 | $1,314.00 | 45.2% | $1,219.50 | 49.2% |
Long documents on Qwen3.8-Max batch ($1 / $3) | $2,454.00 | Batch not offered | Qwen3.8-Max 33% cheaper | Batch not offered | Qwen3.8-Max 33% cheaper |
........
The coding agent shows the largest saving because cached reads dominate its bill and the cache price fell by 70%. The long-document pipeline gains the least, since it pays almost entirely for uncached input, where the cut is the flat 25%. The last row is the exception to the general verdict: when the same document job can run asynchronously, Qwen3.8-Max at QwenCloud's batch rates costs $2,454 a month against $3,681 for real-time Qwen4-Max. The 35% token reduction is Alibaba's figure for one internal benchmark at one effort level; the independent Artificial Analysis runs measured a 15% reduction in output tokens, so the first Qwen4-Max column is the more conservative planning number.
··········
SPEED AND LATENCY MEASUREMENTS FOR THE TWO QWEN MAX MODELS.
The two independent sources disagree on speed, and the disagreement comes from what they measure. Artificial Analysis records raw generation on Alibaba's API: Qwen4-Max produces 54 tokens per second against 61 for Qwen3.8-Max 0902, and takes 159.36 seconds to the first answer token against 102.66 seconds, largely because it reasons before answering. Both figures sit far above the 12-second median for comparable models.
Task-level timing favors the newer model. Artificial Analysis measured 32.1 minutes per index task for Qwen4-Max against 47.0 minutes for Qwen3.8-Max, and Vals AI measured an average latency of 64 minutes 39 seconds against 105 minutes 46 seconds on its benchmark suite. Fewer tokens per completed task outweigh the slower per-token rate. Interactive applications that stream a short answer will feel the longer wait for the first token, so lowering reasoning_effort to low is the available control for chat interfaces where responsiveness counts more than depth.
··········
GAPS IN THE PUBLISHED QWEN4-MAX DOCUMENTATION.
Several facts a buyer would expect are missing. Alibaba has published no system card or safety evaluation for Qwen4-Max, no absolute Work360 scores, and no Qwen4-Max scores in text form on the public benchmarks it reported for Qwen3.8-Max. It has not priced batch inference for Qwen4-Max, and its two official pages disagree on whether batch and fine-tuning are available. The Qwen4-Max model page lists no US deployment scope, while Qwen3.8-Max has one.
On the Qwen3.8-Max side, Alibaba has issued no deprecation date. The model remains on the price list, and its US-scope and batch options are still documented. Independent trackers have started to treat it as superseded: Artificial Analysis marks the 0902 snapshot deprecated and now refreshes its measurements every 18 hours instead of every six. Teams that stay on the older model should expect less frequent independent data over time.
The independent results also cover a narrow slice of use. Neither Vals AI nor Artificial Analysis has published a head-to-head of these two models on long-context retrieval at one million tokens, on image and video understanding, or on multilingual tasks, which are areas Alibaba promotes for both.
··········
CHOOSING BETWEEN QWEN4-MAX AND QWEN3.8-MAX BY WORKLOAD.
The evidence supports moving most real-time API traffic to Qwen4-Max. It is cheaper per token in every region where both are priced, much cheaper on cached reads, faster per completed task in both independent test suites, and stronger on the agentic and coding benchmarks that Vals AI and Artificial Analysis ran under identical conditions. Qwen3.8-Max remains the better fit for a short list of cases defined by features the new model lacks.
........
Workload or requirement | Recommended model | Reason |
Coding agents and long tool-calling loops | Qwen4-Max | Terminal-Bench 4.0 +9.59 points, cache reads 70% cheaper |
Customer-support and chat assistants | Qwen4-Max | 25% lower list price, about 29% lower monthly cost in the example |
Finance agent workflows | Qwen4-Max | Finance Agent v2.1 66.68% vs 58.83% |
Outputs longer than 131,072 tokens | Qwen4-Max | 262,144-token output cap in thinking mode |
Overnight batch processing | Qwen3.8-Max | Batch at $1 / $3; no Qwen4-Max batch price |
Tax and legal research agents | Qwen3.8-Max, after testing | Leads on Tax Agent Bench and Legal Research Bench |
US-only data processing scope | Qwen3.8-Max | Dedicated US scope in Virginia |
Self-hosting on own hardware | Qwen3.8-2.4T-A95B | Only downloadable weights in either family's top tier |
........
For the full limits of each Qwen model, our guide to Qwen context windows and token limits lists every current model. The launch details and regional prices for the new model are in our report on Qwen4-Max pricing at $1.50 input and $4.50 output, and the comparison with a competing Chinese flagship is in Qwen4-Max vs DeepSeek V4 Pro.
A team migrating should set reasoning_effort explicitly, because the default drops from xhigh on Qwen3.8-Max to high on Qwen4-Max, and should move long shared prompts into the explicit cache, where the one-hour option is new. Keep Qwen3.8-Max only for batch pipelines, US-scope processing, or tax and legal workloads where your own tests confirm the older model's lead.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org




