Qwen4-Max vs DeepSeek V4 Pro: Complete Comparison and Report on Pricing, Benchmarks, Cost per Task, and Open-Weight Access

Qwen4-Max is Alibaba's new flagship model, released on November 16, 2026 as a proprietary sparse Mixture-of-Experts system with 2.4 trillion total parameters, 71 billion active per token, native image and video input and a one-million-token context window. DeepSeek V4 Pro is DeepSeek's flagship, whose generally available version (DeepSeek-V4-Pro-0813) rolled out to the DeepSeek app, web chat and API on August 13, 2026, with open weights under the MIT license and the same one-million-token context. The two models sit in a similar list-price band and are the most direct flagship-to-flagship pairing among the large Chinese labs this month.
The independent results point clearly in one direction. On the Artificial Analysis Intelligence Index v5.0.1, Qwen4-Max scores 43 against 23 for DeepSeek V4 Pro at maximum effort, and on the Vals Index Qwen4-Max scores 63.20% against 52.97%, with the widest single gap on Terminal-Bench 4.0 (48.48% against 14.14%). DeepSeek V4 Pro wins on price and openness: its peak-hour API rates of $1.32 input and $3.96 output per million tokens are already below Qwen4-Max's $1.50 and $4.50 in Singapore, and its off-peak rates are exactly half the peak rates, so outside a seven-hour weekday window Qwen4-Max costs about 2.3 times as much per token. Measured per completed task, Qwen4-Max still costs more, at $3.27 per Artificial Analysis index task against $1.77 for DeepSeek V4 Pro.
The practical split follows from those numbers. Qwen4-Max is the stronger choice for terminal agents, finance and tax workflows, image or video input, and teams that need inference pinned to the US, Germany or Japan. DeepSeek V4 Pro is the better buy for high-volume text workloads that can be scheduled off-peak, for anyone who needs downloadable weights to self-host, and for teams already invested in its 384K-token output ceiling. All prices in this report were verified on the official Alibaba Cloud and DeepSeek pricing pages on November 16, 2026.
··········
QWEN4-MAX AND DEEPSEEK V4 PRO: RELEASE TIMELINE AND MODEL IDENTITY.
Alibaba presented Qwen4-Max in a press release dated November 16, 2026 in Hangzhou, describing it as the most capable model in the Qwen family to date. A preview build, qwen4-max-preview, had opened on November 11 through Alibaba's Token Plan subscriptions only; Alibaba's documentation now marks the preview as unsuitable for production and states that it will be retired after 24:00 on November 25, 2026 (UTC+8). The production model, called with the ID qwen4-max, replaces Qwen3.8-Max as the top of the Qwen line, and Alibaba says it activates 25% fewer parameters per token than Qwen3.8-Max at the same 2.4-trillion-parameter scale. The weights are not released, although the press release says smaller open-weight Qwen4 models will follow.
DeepSeek's change log records the general-availability update of DeepSeek V4 Pro on August 13, 2026, rolled out across the app, the web chat and the API under the unchanged model name deepseek-v4-pro. The same update added native Responses API support, three thinking-effort levels and, from 16:00 UTC on August 16, the peak and off-peak price structure that still applies. The Hugging Face model card for DeepSeek-V4-Pro-0813 lists 1.7 trillion total parameters, an MIT license for both code and weights, and a speculative-decoding module attached to the earlier V4 Pro structure; it does not state the active parameter count, which Artificial Analysis lists as 49 billion (alongside a 1.6 trillion total figure that differs from the card).
........
Item | Qwen4-Max | DeepSeek V4 Pro |
Developer | Alibaba (Qwen team) | DeepSeek |
Production release | November 16, 2026 | August 13, 2026 (GA update) |
API model ID | qwen4-max | deepseek-v4-pro (version DeepSeek-V4-Pro-0813) |
Architecture | Sparse MoE, 2.4T total, 71B active | MoE, 1.7T total per model card |
Weights | Proprietary, not downloadable | Open weights, MIT license |
Input modalities | Text, image, video | Text only (vision is listed as not supported) |
Predecessor | Qwen3.8-Max (August 3, 2026) | DeepSeek V4 Pro preview (April 2026) |
........
The modality row is a real separator. Alibaba's model page lists text, image and video as inputs for Qwen4-Max, while DeepSeek's pricing page marks vision as unavailable on deepseek-v4-pro and available only on the cheaper deepseek-flash model. Any workflow that reads screenshots, scanned pages, charts or video frames therefore has only one candidate in this pair.
··········
QWEN4-MAX VS DEEPSEEK V4 PRO API PRICING BY REGION AND HOUR.
Alibaba prices Qwen4-Max by deployment region, with a single tier covering inputs up to one million tokens. In the Singapore (International) region the list price is $1.50 per million input tokens and $4.50 per million output tokens, with cache reads at $0.075, five-minute explicit cache creation at $1.875 and one-hour cache creation at $3. The Global deployment scope (served from Hong Kong, Frankfurt, US Virginia and Tokyo) and the China (Beijing) region charge $1.413 input and $4.24 output, with cache reads at $0.071. A free quota of one million tokens applies in Singapore only. Batch inference for Qwen4-Max is supported only in Beijing, and the pricing page gives no batch rate for the model.
DeepSeek prices deepseek-v4-pro by time of day instead of by region. Off-peak, a cache-miss input token costs $0.66 per million, a cache hit $0.022 and an output token $1.98; during peak hours each figure doubles to $1.32, $0.044 and $3.96. DeepSeek publishes no batch API price for the model.
........
Price per 1M tokens (USD) | Qwen4-Max, Singapore | Qwen4-Max, Global and Beijing | DeepSeek V4 Pro, off-peak | DeepSeek V4 Pro, peak |
Input (cache miss) | $1.50 | $1.413 | $0.66 | $1.32 |
Input (cache hit) | $0.075 | $0.071 | $0.022 | $0.044 |
Output | $4.50 | $4.24 | $1.98 | $3.96 |
Explicit cache write (5 min / 1 h) | $1.875 / $3.00 | $1.767 / $2.827 | Automatic caching, no write fee listed | Automatic caching, no write fee listed |
Batch | Beijing only, no price listed | Beijing only, no price listed | Not offered | Not offered |
Free quota | 1M tokens | None listed | None listed | None listed |
........
For comparison inside each family, Alibaba's page still lists Qwen3.8-Max at $2 input and $6 output in Singapore, so Qwen4-Max is 25% cheaper than its predecessor on both input and output there. DeepSeek's cheaper sibling, deepseek-flash (DeepSeek-V4.1-Flash-1017), costs $0.15 input and $0.60 output off-peak, and Artificial Analysis currently scores it above V4 Pro, a point covered in the benchmark sections below.
··········
DEEPSEEK V4 PRO PEAK-HOUR WINDOW IN US AND EUROPEAN TIME.
DeepSeek defines peak hours as 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays; every other hour, including weekends, is billed at the off-peak rate. That makes seven peak hours on each weekday, or 35 of the week's 168 hours (20.8%). Because the window is defined in UTC, it falls in the evening and night for American users, and the Monday window starts on Sunday evening in US time zones. In US Eastern time (EST) the two peak blocks run from 8:00 to 11:00 p.m. on the previous evening and from 1:00 to 5:00 a.m.; in US Pacific time (PST) they run from 5:00 to 8:00 p.m. on the previous evening and from 10:00 p.m. to 2:00 a.m.; in Central European Time they run from 2:00 to 5:00 a.m. and from 7:00 to 11:00 a.m.
For a US business running daytime traffic, almost all requests land off-peak, which is where DeepSeek V4 Pro's price advantage is largest. A European team working mornings pays the peak rate from 07:00 to 11:00 CET. Workloads that run evenly around the clock pay a weighted average of the two rates; with uniform traffic, about 79% of tokens are billed off-peak.
··········
CONTEXT WINDOW, OUTPUT LIMITS AND EFFORT CONTROLS IN QWEN4-MAX AND DEEPSEEK V4 PRO.
Both models advertise a one-million-token context, but the usable limits differ in the details. Alibaba's Model Studio page for qwen4-max gives a maximum input of 991,808 tokens without thinking and 983,616 tokens with thinking, a maximum output of 131,072 tokens in non-thinking mode, and a combined ceiling of 262,144 tokens for reasoning plus final answer, with a default thinking budget of 131,072. The QwenCloud model page for the same model lists a 393K maximum output and a 262K maximum reasoning length, so the two official pages do not agree on the output ceiling; the Model Studio figures are the more detailed of the two. DeepSeek lists a one-million-token context and a 384K-token maximum output for deepseek-v4-pro, and its model card recommends that output ceiling for high and max effort.
Effort control works similarly in both. Qwen4-Max accepts low, high and xhigh, with high as the default, and maps the aliases minimal, medium and max onto those three; the setting is passed as reasoning_effort in Chat Completions, reasoning.effort in the Responses API and output_config.effort in the Anthropic Messages format. DeepSeek V4 Pro offers low, high and max, with thinking on by default and a non-thinking mode available.
........
Limit or control | Qwen4-Max | DeepSeek V4 Pro |
Context window | 1,000,000 tokens | 1,000,000 tokens |
Max input | 991,808 (no thinking) / 983,616 (thinking) | Not separately listed |
Max output | 131,072 non-thinking; 262,144 reasoning + answer (Model Studio) | 384K tokens |
Effort levels | low, high (default), xhigh | low, high, max |
Non-thinking mode | Yes (thinking on by default) | Yes (thinking on by default) |
Throughput limit | 5,000,000 tokens per minute (QwenCloud page) | 500 concurrent requests |
Fine-tuning via API | Not supported | Not offered via API; weights can be fine-tuned locally |
........
The output ceiling is the one place where DeepSeek V4 Pro has a documented edge on limits. For long generated artifacts such as full code files, translations of long documents or multi-part reports produced in one call, its 384K-token output allowance is larger than the 262,144-token combined reasoning-plus-answer budget that Model Studio documents for Qwen4-Max.
··········
ACCESS BY APP, API AND REGION FOR QWEN4-MAX AND DEEPSEEK V4 PRO.
Consumers can reach both models without writing code. Alibaba says Qwen4-Max is available in QwenStudio at chat.qwen.ai and in QwenWork, and through Token Plan subscriptions on QwenCloud. DeepSeek's August change log states that the V4 Pro update went live in the DeepSeek app and web chat as well as the API; our guide to what is free and what is paid on DeepSeek covers the app side in detail.
For developers, the API surfaces overlap more than they differ. Both accept OpenAI-style Chat Completions, the OpenAI Responses API and the Anthropic Messages format, so either can sit behind tooling written for those interfaces. Qwen4-Max adds the DashScope protocol and built-in server-side tools: web search, a web extractor, a code interpreter and image search through the Responses API, and web search and web fetch through the Anthropic format. DeepSeek V4 Pro supports JSON output, tool calls, chat prefix completion and fill-in-the-middle completion in non-thinking mode, and DeepSeek sells separate search, fetch, OCR and speech APIs.
........
Access route | Qwen4-Max | DeepSeek V4 Pro |
Consumer chat | QwenStudio (chat.qwen.ai), QwenWork | DeepSeek app and web chat |
Subscription route | Token Plan on QwenCloud | None (API is pay-as-you-go) |
API formats | OpenAI Chat, OpenAI Responses, Anthropic Messages, DashScope | OpenAI Chat, OpenAI Responses, Anthropic format |
Built-in tools | Web search, web extractor, code interpreter, image search, web fetch | Separate paid search and fetch APIs ($0.75 per 1,000 calls) |
Inference regions | Beijing, Singapore, Frankfurt, Virginia, Tokyo, Hong Kong | DeepSeek's own service (no region choice listed) |
Self-hosting | Not possible | Possible with the MIT-licensed weights |
........
Region choice is the clearest enterprise difference. Alibaba's regions page says static data stays in the region a customer selects and that the deployment scope decides where inference runs, with US-only and EU-only scopes in Virginia and Frankfurt. DeepSeek's privacy policy states that it collects, processes and stores personal data in the People's Republic of China, and the only way to keep DeepSeek V4 Pro traffic out of that pipeline is to run the open weights on infrastructure the customer controls.
··········
HOW THE QWEN4-MAX AND DEEPSEEK V4 PRO BENCHMARKS WERE PRODUCED.
The two vendors published almost no overlapping figures, which makes the independent evaluators the only source for direct comparison. Alibaba's launch materials give no benchmark numbers against named competitors. The press release says Qwen4-Max ranks sixth on the Code Arena coding leaderboard and scores strongly on Agents' Last Exam, the Vals Index, JobBench and Chartography without stating the scores, and it reports internal Work360 results (a proprietary set covering 721 specialized fields) in which Qwen4-Max beats Qwen3.8-Max while using 35% fewer tokens at high effort and cutting the average cost per task by about 60%.
DeepSeek's model card, by contrast, publishes a full table for V4-Pro-0813, run at max effort with temperature 1.0 and top_p 0.95, with code-agent tasks executed in the minimal mode of DeepSeek's own agent scaffold. Two of its rows, DSBench-FullStack and DSBench-Hard, are internal DeepSeek test sets. None of DeepSeek's competitors in that table is a Qwen model.
Two independent organizations have now tested both models on identical suites. Artificial Analysis runs its Intelligence Index v5.0.1 across nine evaluations (AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2 Hard Subset and Terminal-Bench-Science 0.1). Vals AI ran Qwen4-Max at xhigh effort with a 262,144-token output ceiling and DeepSeek V4 Pro at max effort with a 384,000-token ceiling, so each model was tested at its highest setting. Neither Artificial Analysis nor Vals AI states a test date on the model pages; Vals AI lists Qwen4-Max's entry with a November 15, 2026 date and DeepSeek V4 Pro's with August 13, 2026.
··········
ARTIFICIAL ANALYSIS AND VALS AI RESULTS FOR QWEN4-MAX AND DEEPSEEK V4 PRO.
........
Independent measure | Qwen4-Max | DeepSeek V4 Pro (max effort) | Source |
Intelligence Index v5.0.1 | 43 | 23 (high: 21; non-reasoning: 13) | Artificial Analysis, Nov 16, 2026 |
Output tokens to run the index | 127M | 119M | Artificial Analysis |
Cost per index task | $3.27 | $1.77 | Artificial Analysis |
Vals Index | 63.20% (rank 11 of 38) | 52.97% (rank 29 of 38) | Vals AI |
Cost per Vals Index test | $9.33 | $2.50 | Vals AI |
Vals Index latency | 64 min 39 s | 58 min 3 s | Vals AI |
........
A 20-point gap on the Artificial Analysis index is large at this level of the scale, and the Vals Index gap of just over 10 percentage points points the same way. Token consumption is close: Qwen4-Max used 127 million output tokens to complete the Artificial Analysis suite against 119 million for DeepSeek V4 Pro at max effort, so the cost difference per task comes mostly from Qwen4-Max's higher per-token price. Artificial Analysis prices DeepSeek V4 Pro at its peak rates of $1.32 and $3.96, and Vals AI lists the same peak figures, which means both cost-per-task numbers for DeepSeek would roughly halve for work run off-peak.
The Artificial Analysis page for DeepSeek V4 Pro also shows that the newer and cheaper DeepSeek V4.1 Flash 1017 scores 33 on the same index, ten points above V4 Pro at max effort. Buyers comparing Qwen4-Max against DeepSeek's lineup should therefore treat V4 Pro as DeepSeek's largest current model, which on Artificial Analysis is no longer its highest-scoring one.
··········
[[ADPLUS_300x250_1]]
VALS AI BENCHMARK-BY-BENCHMARK SCORES FOR QWEN4-MAX AND DEEPSEEK V4 PRO.
Vals AI publishes per-benchmark accuracy for both models on the same task sets, which gives the most granular like-for-like view available.
........
Vals AI benchmark | Qwen4-Max | DeepSeek V4 Pro | Gap (points) |
Terminal-Bench 4.0 | 48.48% | 14.14% | +34.34 Qwen |
Tax Agent Bench | 40.35% | 26.76% | +13.59 Qwen |
HLAB-Vals-1.1 | 21.25% | 10.83% | +10.42 Qwen |
Finance Agent v2.1 | 66.68% | 57.12% | +9.56 Qwen |
Vals Finance Index | 61.99% | 52.94% | +9.05 Qwen |
EMB v1.1 | 70.83% | 65.13% | +5.70 Qwen |
Vibe Code Bench v1.1 | 86.66% | 82.30% | +4.36 Qwen |
Vals Legal Index | 70.71% | 69.01% | +1.70 Qwen |
Legal Research Bench | 41.35% | 40.87% | +0.48 Qwen |
........
Qwen4-Max leads on every benchmark where both scores exist, but the size of the lead varies sharply. The terminal-agent result is the decisive one for coding agents: on Terminal-Bench 4.0, Qwen4-Max completes more than three times as many tasks as DeepSeek V4 Pro under Vals AI's setup. Finance and tax agent work shows a lead of nine to fourteen points. On legal tasks the two models are close: the 0.48-point difference on Legal Research Bench and the 1.70-point difference on the Vals Legal Index fall within or near the published error margins of roughly 1.8 to 3.4 points.
Vals AI also lists results for DeepSeek V4 Pro that have no Qwen4-Max counterpart yet, including 96.40% on SWE-bench (rank 2 of 88), 92.42% on GPQA Diamond and 87.53% on LiveCodeBench. Those scores show DeepSeek V4 Pro remains strong on established single-turn coding and reasoning sets, but they cannot be set against Qwen4-Max until Vals AI publishes the same rows for it.
··········
BENCHMARKS ONLY DEEPSEEK OR ALIBABA PUBLISHED.
DeepSeek's own table for V4-Pro-0813 covers agentic coding, tool use and long-horizon tasks. Alibaba has not published numeric scores on any of these sets for Qwen4-Max, so the figures below describe DeepSeek V4 Pro only and come from the vendor.
........
DeepSeek-published benchmark (max effort) | DeepSeek V4 Pro | Qwen4-Max |
HLE without / with tools | 42.7 / 60.0 | Not published |
Terminal Bench 2.1 | 87.9 | Not published |
NL2Repo | 61.5 | Not published |
Cybergym | 83.3 | Not published |
DeepSWE | 62.7 | Not published |
Toolathlon-Verified | 74.1 | Not published |
Agents' Last Exam | 25.7 | Named by Alibaba, no score given |
AutomationBench (Public) | 31.8 | Not published |
DSBench-FullStack / DSBench-Hard (internal) | 71.1 / 67.2 | Not published |
........
The Terminal-Bench figures need care. DeepSeek's 87.9 is on Terminal Bench 2.1 with DeepSeek's own scaffold, while the 14.14% that Vals AI measured is on Terminal-Bench 4.0 with Vals AI's setup; these are different benchmark versions with different task sets, and the two numbers cannot be read as a contradiction. Alibaba's only specific coding claim, a sixth-place position on Code Arena, could not be confirmed in this turn because the public Code Arena leaderboard page we fetched does not list Qwen4-Max by name.
··········
COST PER COMPLETED TASK AND THREE WORKLOAD COST EXAMPLES.
List prices alone overstate how much a team saves with DeepSeek V4 Pro when it runs during peak hours, and understate the saving off-peak. The examples below use the official rates verified on November 16, 2026 and were computed in Python. Assumptions: a coding-agent session that sends 3 million input tokens (90% served from cache) and produces 200,000 output tokens; a support chatbot handling 10,000 requests a day for 30 days at 4,000 input tokens (60% cached) and 600 output tokens each; and a contract review of 200 documents of 300,000 tokens each with no cache reuse and 6,000 output tokens per document. Cache hits for Qwen4-Max are priced at the implicit-cache read rate, so no cache-write fees are included.
........
Workload | Qwen4-Max, Singapore | Qwen4-Max, Global | DeepSeek V4 Pro, off-peak | DeepSeek V4 Pro, peak |
Coding agent, one session | $1.55 | $1.46 | $0.65 | $1.31 |
Support chatbot, 300,000 requests a month | $1,584.00 | $1,492.56 | $689.04 | $1,378.08 |
Contract review, 200 documents | $95.40 | $89.87 | $41.98 | $83.95 |
........
Off-peak, DeepSeek V4 Pro costs about 44% of Qwen4-Max's Singapore price for the same token volume. During peak hours the difference shrinks to about 12%. With traffic spread evenly across the week, the weighted DeepSeek cost for the chatbot comes to $832.59 a month, and Qwen4-Max in Singapore costs 1.9 times that.
Those figures assume both models need the same number of tokens and the same number of attempts to finish a job, which the benchmarks above do not support. On the Artificial Analysis suite, where the measured token use was similar, Qwen4-Max cost 1.85 times as much per task ($3.27 against $1.77); on the Vals Index suite it cost 3.7 times as much per test ($9.33 against $2.50) while scoring about ten points higher. For tasks where DeepSeek V4 Pro fails more often, as on Terminal-Bench 4.0, a retry or a human fix can erase the token saving; for simpler, high-volume text work where both models succeed, the off-peak discount is a direct saving of more than half.
··········
SPEED AND LATENCY MEASUREMENTS FOR QWEN4-MAX AND DEEPSEEK V4 PRO.
Artificial Analysis measures DeepSeek V4 Pro on DeepSeek's API at 91 output tokens per second at max effort, 78 at high effort and 90 without reasoning, with a time to the first answer token of 51.48 seconds at max effort and 2.45 seconds in non-reasoning mode. Its average time per Intelligence Index task is 18.1 minutes at max effort. Artificial Analysis had not yet published speed or time-to-first-token figures for Qwen4-Max on the day of launch.
The only comparable timing figure comes from Vals AI, which reports a total latency of 64 minutes 39 seconds for Qwen4-Max to run the Vals Index suite, against 58 minutes 3 seconds for DeepSeek V4 Pro. That puts Qwen4-Max about 11% slower on that suite, at a higher effort setting and with a smaller output ceiling. For interactive chat, the long wait before the first answer token at DeepSeek V4 Pro's max effort is a reason to drop to high effort or non-thinking mode, where its first token arrives in seconds.
··········
LICENSING, DATA HANDLING AND SAFETY POLICIES OF QWEN4-MAX AND DEEPSEEK V4 PRO.
........
Policy area | Qwen4-Max | DeepSeek V4 Pro |
License | Proprietary API model | MIT license for code and weights |
Training on API data | Alibaba Cloud states it will never use customer data for model training | Privacy policy allows training on personal data, with opt-out by email to privacy@deepseek.com |
Data location | Static data stays in the selected region; US-only and EU-only inference scopes available | Personal data collected, processed and stored in the People's Republic of China |
Published safety measures | Defensive cyber data kept and offensive content filtered; dedicated CBRN safety data and post-training | Model safety parameter documented for deepseek-flash only; no separate safety card for V4 Pro located |
........
The open-weight license gives DeepSeek V4 Pro an advantage that no price comparison captures: a company can download the MIT-licensed weights, run them on its own hardware or a cloud of its choice, fine-tune them and keep every prompt inside its own perimeter. The cost is infrastructure, since a 1.7-trillion-parameter model needs a multi-GPU deployment. Qwen4-Max cannot be self-hosted at all, so its data controls depend on Alibaba Cloud's regional deployment and its statement that customer data is not used for training.
··········
WHAT NEITHER VENDOR HAS PUBLISHED FOR THIS COMPARISON.
Several facts a buyer would want are missing from the official record. Alibaba has not released a model card or technical report for Qwen4-Max with numeric benchmark results, has not stated the score behind its Code Arena claim, and publishes conflicting maximum-output figures on two of its own pages (131,072 non-thinking tokens on Model Studio, 393K on QwenCloud). It lists no batch price for Qwen4-Max and limits batch inference to the Beijing region.
DeepSeek has not published an updated benchmark table for V4 Pro since August, even though its October change log reports newer results for V4.1 Flash on Terminal-Bench 4.0 (42.3) and Agents' Last Exam (35.4) that V4 Pro has no matching figures for. DeepSeek's own model card and Artificial Analysis disagree on V4 Pro's total parameter count (1.7 trillion against 1.6 trillion), and the card does not state the active count. Neither vendor has published a speed figure, and no third-party measurement of Qwen4-Max throughput existed on launch day.
··········
QWEN4-MAX OR DEEPSEEK V4 PRO: CHOICE BY WORKLOAD.
........
Workload or requirement | Recommended model | Reason from the verified data |
Terminal and coding agents | Qwen4-Max | 48.48% vs 14.14% on Vals AI Terminal-Bench 4.0 |
Finance, tax and accounting agents | Qwen4-Max | Leads by 9 to 14 points on Finance Agent, Finance Index and Tax Agent Bench |
Image, chart or video input | Qwen4-Max | DeepSeek V4 Pro does not accept vision input |
US-only or EU-only inference | Qwen4-Max | Virginia and Frankfurt scopes; DeepSeek stores data in China |
High-volume text, off-peak schedule | DeepSeek V4 Pro | $0.66 / $1.98 off-peak vs $1.50 / $4.50 |
Self-hosting or fine-tuning | DeepSeek V4 Pro | MIT-licensed open weights |
Very long single outputs | DeepSeek V4 Pro | 384K-token output ceiling |
Legal research and drafting | Either | Within about 2 points on Vals AI legal benchmarks |
........
For most teams buying an API model on capability, Qwen4-Max is the stronger flagship in this pair: it leads both independent indexes, wins every shared Vals AI benchmark and is 25% cheaper than the Qwen3.8-Max it replaces, as covered in our Qwen4-Max launch and pricing report. Its context and token limits are set out in our Qwen context window guide. DeepSeek V4 Pro keeps a clear role where the bill or the license decides: text pipelines that can run outside the seven weekday peak hours pay less than half of Qwen4-Max's rates, and only DeepSeek offers weights a company can host itself. Teams already weighing DeepSeek against other open-weight flagships can compare it with Mistral's model in our Mistral Large 4 vs DeepSeek V4 Pro report. For agentic coding and finance work, the 34-point Terminal-Bench 4.0 gap and the nine-to-fourteen-point finance leads make Qwen4-Max the model to choose even at roughly twice the off-peak price.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org




