AA's measurement of DeepSeek V4 Pro 0813 (Reasoning, Max Effort): Intelligence Index 53, ranked 3rd in its pool of 107 models; speed 80.3 tok/s; output $3.96/1M (peak); 1M context; 1600B total parameters with 49B active — top-tier intelligence, but on the expensive and verbose side (a single index evaluation generated 130M tokens).
Suitable tasks: Comparing V4-Pro with other frontier models on the same index scale (intelligence/speed/cost/verbosity); providing third-party evidence for “what Pro is good at” (high intelligence + long context + text-only).
Unsuitable tasks: Treating the Intelligence Index as the success rate for a specific codebase; mixing data from the 2026-04-24 older version with the new 0813 version; using peak prices to replace all cost calculations (off-peak and cache discounts differ substantially).
Applicable model versions: DeepSeek V4 Pro 0813 (the page measures it under a max effort reasoning configuration; the official non-reasoning variant also exists but is not tested on this page).
Applicable clients, Agents, or APIs: AA's unified API Provider Benchmark harness; the model outputs text, and the official API supports the Responses/Anthropic formats.
Recommended reasoning levels and parameters: This page is the max effort tier; AA notes that lower tiers (e.g., high/low) produce different intelligence, token, and cost figures, so each tier must be queried separately.
Metrics: Artificial Analysis Intelligence Index (composite intelligence), cost per task, output speed, and Verbosity (total tokens generated during the index evaluation).
Measurement results (page text): Intelligence Index 53, ranked 3/107 (peer median 27); Speed 80.3 tok/s, ranked 22/107 (median 67); Cost $1.32/1M input, $3.96/1M output (peak pricing, 97% cache discount), $0.25 per Intelligence Index task, ranked 24/107 (input median $0.30, output median $1.20); Verbosity 130M tokens (median 100M, on the verbose side); total cost of the full index evaluation $604.51.
Technical specs: 1M context (roughly 1,500 A4 pages); 1600B total parameters, 49B active; text input and output; MIT license; weights on Hugging Face.
Comparison baseline: A pool of 107 peer models (open weights, Large >150B, or reasoning class); the “Highlights” comparison chart lists Claude Opus 5 (max), Claude Fable 5, GPT-5.6 Sol (max), Grok 4.6 (high), Kimi K3 (max), GLM-5.3 (max), Muse Spark 1.2 (xhigh), and others.
The page's measurements were run through AA's unified API Provider Benchmark harness with a fixed max effort configuration. What is disclosed: the Intelligence Index methodology, per-task cost, output speed, and verbosity figures. What is not disclosed: the full sample sets, weighting, and harness details, which are controlled by AA and not fully public (the Reuters report makes the same point).
| Metric | Value |
|---|---|
| Intelligence Index | 53, ranked 3/107 (peer median 27) |
| Speed | 80.3 tok/s, ranked 22/107 (median 67) |
| Cost | $1.32/1M in / $3.96/1M out (peak pricing, 97% cache discount); $0.25 per Intelligence Index task; ranked 24/107 (input median $0.30, output median $1.20) |
| Verbosity | 130M tokens per index evaluation (median 100M) |
| Total evaluation cost | $604.51 |
| Context | 1M (roughly 1,500 A4 pages) |
| Parameters | 1600B total / 49B active |
| Modality | Text in / text out |
| License | MIT (weights on Hugging Face) |
Under AA's measurement at max effort, DeepSeek V4 Pro 0813 delivers top-tier intelligence (Index 53, 3rd of 107) but is expensive and verbose: peak prices of $1.32/$3.96 per 1M tokens, $0.25 per task, and 130M tokens generated per evaluation are all above the peer medians. The index score of 53 is consistent with the Reuters 2026-08-13 report, so the AA data cross-checks. Note the version distinction: the 0813 snapshot on this page is not the same as the older 2026-04-24 preview snapshot (ChatBench AA Index 43.75), so the two must not be mixed when comparing.
Limitation: AA is a third-party index; the samples, weights, and harness are controlled by AA and not fully disclosed (the Reuters report makes the same point). The index is a composite score, not a task success rate.
Limitation: “max effort” is the page's fixed configuration; the max tier's token consumption and cost are significantly higher than the official default high tier. Production model selection should re-check data at the actual tier in use.
Limitation: Verbosity of 130M tokens shows that max-tier output is wordy; long conversations and batch tasks must factor token costs into the budget, using low/high tier comparisons when needed.
Limitation: Page prices are point-in-time data as collected; DeepSeek moved to peak/off-peak billing effective 2026-08-16, so later price changes require re-verification.
Reproduction steps: Re-check the live Artificial Analysis page for DeepSeek V4 Pro per effort tier (max effort as shown here, plus high/low for comparison), and cross-check current prices against the official pricing page for the effective peak/off-peak rates.
Page positioning statement (verbatim): “DeepSeek V4 Pro 0813 (Reasoning, Max Effort) is amongst the leading models in intelligence, but somewhat expensive when comparing to other open weight models of similar size. It's also faster than average, however somewhat verbose.”
The same index score of 53 matches the Reuters 2026-08-13 report (see review 03), showing the AA data can be cross-checked; the ChatBench 2026-08-12 snapshot shows the older version (Released 2026-04-24, AA Index 43.75), which differs from this page's 0813 — mind the version distinction.
Cost basis: page prices are the peak tier ($1.32/$3.96), consistent with the peak prices on the official pricing page effective 2026-08-16; off-peak is half ($0.66/$1.98), and cache-hit input is $0.022/$0.044.
AA is a third-party index: the samples, weights, and harness are controlled by AA and not fully disclosed (the Reuters report makes the same point); the index is a composite score, not a task success rate.
“max effort” is the page's fixed configuration: the max tier's token consumption and cost are significantly higher than the official default high tier; production model selection should re-check data at the actual tier in use.
Verbosity of 130M tokens shows that max-tier output is wordy: long conversations and batch tasks must factor token costs into the budget, using low/high tier comparisons when needed.
Page prices are point-in-time data as collected: DeepSeek implemented peak/off-peak billing effective 2026-08-16, and later price adjustments require re-verification.
Page text: “In total, it cost $604.51 to evaluate DeepSeek V4 Pro 0813 (Reasoning, Max Effort) on the Intelligence Index.”
DeepSeek V4 Pro
