Reuters DeepSeek-V4-Pro-0813: Official Pricing vs. Independent Index

One-sentence takeaway

Reuters cites Artificial Analysis's independent pricing and index data: V4-Pro-0813 scores 53 on the reasoning Intelligence Index, versus 40 for V4 Flash, but Pro's input and output prices are roughly 9 and 14 times those of Flash, respectively. Model selection must account for both quality and cost.

Use cases

• Suitable tasks: API vendor selection, Agent cost budgeting, and price/capability tier comparisons between V4-Pro and V4-Flash.

• Unsuitable tasks: Treating the Intelligence Index as the success rate for a specific codebase, or deciding production routing based solely on media reports.

• Applicable model versions: DeepSeek-V4-Pro-0813 and V4 Flash as reported by Reuters.

• Applicable clients, Agents, or APIs: DeepSeek API, app, and web; the report does not provide a calling harness.

• Recommended reasoning levels and parameters: Not disclosed in the report; follow the official low/high/max rules and record actual token costs on the target task.

Test environment

• Source data: Reuters's report on DeepSeek's GA release; pricing and the Intelligence Index are from Artificial Analysis.

• Metric: The Artificial Analysis Intelligence Index combines nine capability categories, including agentic work, tool use, coding, scientific reasoning, and long-context tasks.

• Reproduction status: Reuters did not rerun the tasks and did not disclose Artificial Analysis's complete sample set or weighting.

Input/configuration

Artificial Analysis's per-task inputs, model parameters, tool configuration, and number of repetitions were not disclosed. Reuters reports prices of $1.32 per million input tokens and $3.96 per million output tokens, compared with $0.14/$0.28 for V4 Flash; the pricing page and peak/off-peak rates may change over time.

Results data

Conclusion

Pro's independent composite index is higher than Flash's, but the price gap is also substantial. In practice, routing can use Flash first for simple, high-concurrency execution, then send complex planning, long-horizon Agent, or high-cost-of-failure tasks to Pro, provided that the same business eval confirms the switch is worthwhile.

Limitations and reproduction steps

• Limitation: This is a second-hand report, and the metric comes from Artificial Analysis; there are no per-task data or independent run traces, so it should not be treated as a reproducible benchmark.

• Reproduction steps: Lock the date on the current API pricing page; define simple/complex task buckets; record Pro/Flash success rates, tokens, latency, tool failures, and human remediation; calculate the routing threshold using business-weighted actual costs.

• Pricing boundary: DeepSeek separately announced peak/off-peak rates effective 2026-08-16; Reuters's static unit prices should not replace current billing calculations.

Original evidence and data

The Reuters report explicitly gives $1.32/$3.96, Flash $0.14/$0.28, and reasoning Intelligence Index scores of 53/40, and says that the metric covers nine capability categories. This article preserves the report's source and Artificial Analysis's attribution rather than splitting the index into fabricated subscores.

Applicability boundaries

• The index is not a substitute for any single measure such as code completion rate, tool success rate, or long-context accuracy.

• Prices are affected by peak/off-peak periods, caching, the supplier, and time; production budgets must check the current official pricing.

• Routing strategies must retain failure fallback and human review; do not automatically move all tasks to Flash because of its lower price.

Source excerpt or observation (compliance short quote only)

Reuters's core fact is that Pro is “pricing it several times higher than its V4 Flash model,” while also reporting a higher reasoning index than Flash. Together, these establish a quality-cost trade-off rather than a one-dimensional ranking.