Search "DeepSeek V4 Pro benchmarks" right now and you will get charts that disagree with each other. One post says it is the smartest model ever tested. Another says it is slow. A third quotes a coding score you cannot find on any official page. All three are "about" the same model — the difference is where each number came from, and whether the person quoting it even knows.
DeepSeek V4 Pro (version -0813, released August 13, 2026) is a genuinely important release: a 1.6-trillion-parameter Mixture-of-Experts flagship that ranks #2 of 104 on the independent Artificial Analysis Intelligence Index. But a #2 ranking means nothing if you cannot tell which numbers are independent measurements, which are vendor self-reports, and which are made up. This article decodes the 0813 numbers that actually matter, explains what each one measures, and gives you a rule of thumb for reading any model benchmark — so the next leaderboard you see does not mislead you.
What This Article Solves
Conflicting numbers: vendor self-reported scores versus independent third-party measurements, benchmark terminology you do not recognize (II, TTFT, verbosity), and no idea what a run score means for your actual work. This article uses one independent data source — Artificial Analysis, captured 2026-08-13 — and explains each number in plain terms instead of just quoting it.
Why This Data Set Is Trustworthy (and Most Others Aren't)
Almost every "benchmark" number circulating about DeepSeek V4 Pro fails the first test of credibility: who measured it, and how?
DeepSeek publishes specifications and pricing, not its own performance scores — and that is the correct behavior. The numbers in this article come from Artificial Analysis (AA), an independent benchmarking organization that runs models itself on a controlled, standardized harness rather than accepting vendor-provided results. That distinction matters: a score is only as trustworthy as the person who ran the test, and AA has no incentive to inflate any vendor's numbers.
The second question is which version of the index you are looking at. AA's Intelligence Index is currently v4.1.1, a composite of nine separate evaluations designed to cover different capability areas:
- GDPval-AA v2 — graduate-level, evidence-driven reasoning (AA's own version of GDPval)
- 𝜏³-Banking — agentic tool use and multi-step workflows on real banking tasks
- Terminal-Bench v2.1 — agentic coding in a shell: reading files, chaining commands, fixing errors
- SciCode — scientific coding: implementing research problems as runnable code
- Humanity's Last Exam (HLE) — frontier knowledge across dozens of domains
- GPQA Diamond — graduate-level science questions (physics, chemistry, biology)
- CritPt — critical-point reasoning on difficult, ambiguous problems
- AA-Omniscience — broad knowledge retrieval under time pressure
- AA-LCR — long-context recall, testing whether the model actually uses what is in a huge prompt
We are deliberately not quoting per-test scores: AA publishes an overall index plus per-test results on its live page, which change as runs accumulate. What is stable and dated is the composite — and the composite is the right thing to reason with.
Intelligence: 53 and the #2 of 104 Ranking
On the Artificial Analysis Intelligence Index (v4.1.1), DeepSeek V4 Pro 0813 scores 53, ranked #2 of 104 models tested. To calibrate: the median score across all 104 models is 27. In other words, V4 Pro sits at roughly double the median capability of everything AA has tested — a two-standard-deviation class jump, not a marginal improvement.
Reading this correctly:
- #2 of 104 is a statement about the composite index across all nine evaluations, not about any single task. It means V4 Pro is the second-strongest all-around model AA has measured.
- A composite ranking is the best "which model is generally smarter" signal available, and also the worst tool for choosing a model for one specific job — a model ranked #40 could beat #2 on your narrow task (more on this in the decision framework below).
- The score was captured with Reasoning Max Effort — AA runs V4 Pro in its thinking mode at maximum reasoning effort, which is the setting you should use for hard tasks anyway. DeepSeek's API supports switching between thinking and non-thinking modes; the benchmark reflects the thinking profile.
Capability Dimensions: What V4 Pro Is Actually Good At
The nine evaluations inside the index are grouped by capability. DeepSeek V4 Pro's known strengths and the honest caveats:
Coding and agentic work (Terminal-Bench v2.1, SciCode, 𝜏³-Banking). This is where the model is designed to win: agentic shell operations, scientific coding, and multi-step tool workflows with real transactions. Terminal-Bench-style tasks reward models that plan before acting and self-correct after errors — a natural fit for a 1.6T MoE with 49B active parameters running reasoning at max effort.
Graduate reasoning and knowledge (GPQA Diamond, Humanity's Last Exam, CritPt). GPQA covers graduate-level science; HLE is the hardest general knowledge test in circulation. A strong composite score here means V4 Pro can hold up under genuinely hard, ambiguous questions — the kind that break smaller models into confident wrong answers.
Long context (AA-LCR). With a 1M-token context window, the relevant test is not "can it fit the prompt" but "can it recall facts buried in the middle of it." AA-LCR measures exactly that, and it is the dimension most relevant to anyone planning to use V4 Pro with huge codebases or long documents.
The honest caveat: DeepSeek V4 Pro is text-only. It does not accept image input — so it cannot reason about screenshots, diagrams, or UI mockups, and none of the nine index evaluations cover multimodal abilities. If your workflow needs vision, this model is not a complete answer regardless of its index score.
Speed and Latency: 83.2 tok/s and a 1.63s TTFT
Intelligence is only half the picture. The other half is whether the model is usable interactively.
- Output speed: 83.2 tokens per second (median across models: 66.2). Ranked #19 of 104 — comfortably above average. For a 1.6T-parameter flagship that is impressive: larger models usually sacrifice throughput, and V4 Pro is fast enough for real-time interactive coding.
- TTFT (time to first token): 1.63 seconds (median: 1.89s). This is the time between sending a request and the first token arriving. Under two seconds puts V4 Pro in the same latency class as much smaller models.
Why these two numbers matter: TTFT is what your perception of speed depends on, and tok/s is what your bill depends on. A model that is slow to start but fast to stream feels sluggish in chat but cheap on long generations — V4 Pro is decent on both, which is why it scores well on "speed of use" combined metrics. Treat these as dated measurements from AA's harness; absolute throughput varies by provider and serving configuration.
Cost Efficiency: What a #2 Ranking Actually Costs
A ranking without a price tag is a hobby. Here is the economics of running a #2-ranked model, from AA's measured cost data and DeepSeek's official pricing (per 1M tokens, 2026-08-13):
| Metric | DeepSeek V4 Pro | Context |
|---|---|---|
| Output price | $0.87 / 1M | Category median is $2.20 — under half |
| Input price (cache miss) | $0.435 / 1M | Slightly above median $0.33 |
| Cache hit input | $0.003625 / 1M (−99%) | Ranked #8 of 104 on cache price |
| Blended rate (7:2:1 cache hit/input/output) | ≈ $0.18 / 1M | Effective cost with caching |
| Total cost to run all II tasks | $135.03 | AA's measured spend for the full evaluation |
The two numbers that deserve attention:
- Blended ≈ $0.18/1M. Because DeepSeek's context caching discount is enormous (−99% on cached input), a workload with heavy repeated context — agents, coding sessions, multi-turn chat — lands far below the nominal $0.435 input price. Cache hits are what make this model cheap in production.
- $135.03 total for the entire II evaluation. AA published this as its actual spend running all nine suites at max reasoning effort. It is a sanity check: a genuine full benchmark run of a frontier model costs roughly a hundred dollars in API fees. Any vendor posting "independent benchmark results" without this kind of cost footprint is probably not running real evaluations.
One critical warning: DeepSeek has announced a significant price increase for V4 Pro in the near future. The figures above are the official 2026-08-13 prices; budget against the official pricing page, not this article.
The Verbose Flag: 130M Tokens and What It Means for You
Every benchmark has a footnote. For DeepSeek V4 Pro, it is verbosity: the II evaluation produced ~130M output tokens, versus a median of ~100M across tested models. Translation: at max reasoning effort, V4 Pro writes about 30% more text per task than the average model.
This is the benchmark detail most review posts skip, and it changes your cost math:
- More tokens per answer means longer answers — genuinely useful on hard reasoning tasks, where thinking out loud helps, but wasteful on routine work like summarization or extraction.
- Your API bill scales with output tokens. At $0.87/1M, a 30% verbosity premium on a heavy workload is real money — the cost table above is why DeepSeek's answer is "cache aggressively and switch to non-thinking mode for simple tasks."
- The practical lever: use thinking mode only when the task needs it. DeepSeek's API lets you switch between thinking and non-thinking per request. Cheap, short-answer work does not need max-effort reasoning, and non-thinking mode cuts both latency and token count.
How to Use Benchmarks to Pick a Model (Decision Framework)
Stop treating a composite index like a high-score competition and use it as a filter, then a test:
- Check the source. Is the number from an independent runner (Artificial Analysis) or an unlabeled chart? No primary source, no trust. This alone eliminates most social-media "benchmarks."
- Match the task, not the total. Coding agent? Look at Terminal-Bench/SciCode and ignore knowledge tests. Long documents? Look at AA-LCR. A single #2 composite tells you the model is strong everywhere — but the deciding number is the sub-test for your job.
- Price the workload, not the token. Compute blended cost with your cache-hit ratio (DeepSeek's −99% cache discount changes everything for repeated context), then multiply by your real monthly volume. Verbose models cost more per task even at the same per-token price.
- Check the boundaries the score doesn't show. For V4 Pro: text-only (no images), 500 concurrent requests (vs. 2,500 for V4 Flash), thinking/non-thinking modes, and an announced price increase. Benchmarks never include these, and they decide real usability.
Rule of Thumb: a benchmark score tells you what a model could do; the task relevance of the sub-test tells you whether it will do it for you. Total scores rank models for bragging rights — sub-test scores and pricing buy you production systems.
If you want a flagship without the boundaries above, GLM 5.2 — Zhipu AI's flagship, ~750B MoE with ~40B active — is MIT-licensed with a 1M context, and you can try it free with no API key on glm5.app. Different design, same league; run your own task on both before you budget. (New to the family? Our DeepSeek V4 Flash benchmarks piece explains the efficiency tier, and V4 Flash vs. V4 Pro covers the family split.)
Frequently Asked Questions
Is DeepSeek V4 Pro the strongest model right now? Per Artificial Analysis (v4.1.1, captured 2026-08-13), it scores 53 — #2 of 104, double the 27 median. "Strongest" depends on your task: the composite says elite all-around; sub-tests decide specific jobs; and it is text-only, so multimodal work is out of scope regardless.
Is DeepSeek V4 Pro fast? Yes, for its size: 83.2 tok/s output (median 66.2) and 1.63s TTFT (median 1.89s), measured by AA. Speeds vary by provider and serving configuration — treat these as AA's dated measurements.
Why is DeepSeek V4 Pro so verbose? At Reasoning Max Effort it produces ~130M tokens on the II suite vs. a ~100M median — about 30% more text per task. It is a side effect of thinking-mode reasoning. Switch to non-thinking mode for routine work to cut tokens and cost.
Can I trust the DeepSeek V4 Pro benchmark numbers I see? Only if they cite an independent source. AA's figures are reproducible measurements with published methodology; unlabeled charts and vendor self-reports are not. Check the sub-test matches your task, and note DeepSeek's announced price increase before budgeting.
How much does a #2-ranked model cost to run? Official prices (2026-08-13): $0.87/1M output, $0.435/1M cache-miss input, $0.003625/1M cache-hit input (−99%, ranked #8/104). Blended with caching ≈ $0.18/1M; AA's full nine-suite run cost $135.03 total.
Bottom Line
The 0813 story on DeepSeek V4 Pro, from independent measurements: Intelligence Index 53 (#2/104, median 27), 83.2 tok/s, 1.63s TTFT, ~$0.18/1M blended with caching, and a 30% verbosity premium at max reasoning effort. It is a genuinely elite, cheap-to-serve flagship — with three honest boundaries: text-only, cache-dependent economics, and an announced price increase.
If you use its reasoning mode for hard tasks and its non-thinking mode for routine work, the economics work out. If you want a flagship to try right now without an API key, test GLM 5.2 free on glm5.app — same league, different boundaries, and your own prompt is the only benchmark that ships.
By the GLM 5 Team. Benchmark figures are Artificial Analysis measurements captured 2026-08-13 (Intelligence Index v4.1.1); prices are official DeepSeek pricing as of 2026-08-13. Rankings, versions, and prices change frequently — DeepSeek has announced an upcoming price increase, so verify the linked sources before quoting any figure.
Sources
- Artificial Analysis — DeepSeek V4 Pro — Independent measurements of the Intelligence Index (53, #2/104), 83.2 tok/s output speed, 1.63s TTFT, verbosity (~130M tokens), and evaluation cost ($135.03). Primary source for all performance figures.
- Artificial Analysis — Methodology — How the Intelligence Index v4.1.1 and its nine evaluations (GDPval-AA v2, 𝜏³-Banking, Terminal-Bench v2.1, SciCode, HLE, GPQA Diamond, CritPt, AA-Omniscience, AA-LCR) are measured.
- DeepSeek Models & Pricing — Official pricing for
deepseek-v4-pro(cache hit/miss input, output), context limits, and concurrency. Note: a significant price increase has been announced. - DeepSeek-V4-Pro on Hugging Face — Official model card: 1.6T-parameter MoE with 49B active per token, 1M context, MIT license.
- DeepSeek Thinking Mode — Official documentation for switching between thinking and non-thinking modes, the setting AA used for its measurements.