DeepSeek V4 released: significant capability improvement, but API prices raised sharply

Overview

This report comments on the release of DeepSeek V4 and its implications for China's AI path. The core message is that DeepSeek V4 has been significantly upgraded in model capabilities, parameter scale, and context length, and continues to compete in the frontier large-model race via the open-weight MoE route; however, V4-pro API pricing has risen materially versus V3.2, making the pace of commercialization adoption and subsequent compute cost reduction key variables.

Core views

The report's main views include: first, DeepSeek V4 stands out in agent capabilities, world knowledge, and reasoning efficiency, with coding and agent benchmarks reaching or exceeding Opus-4.6; second, V4-pro's parameter scale has expanded significantly to 1.6T total parameters and 49B active parameters, indicating that improved model capability comes with higher compute requirements; third, API price increases make V4-pro's short-term cost clearly higher than V3.2, though it remains below some leading overseas frontier models; fourth, if Ascend 950 Super Node is widely deployed in 2H26, there is meaningful downside potential for V4-pro pricing.

Analysis framework

The report uses an event-commentary and cross-model comparison approach, comparing DS-V4-pro, DS-V4-flash, DS-V3.2, and models such as Opus, Qwen, Kimi, and GLM across release date, total parameters, active parameters, context window, input price, and output price, while also incorporating DeepSeek's comments on improved compute supply in 2H26 to assess the future price path.

Methodology notes

• Cross-model comparison of large modelsComparison of frontier LLM capabilities and pricing

  Parameter scale, active parameters, context window, and API pricing jointly measure model capability and usage cost.

  By comparing total parameters, active parameters, context window, input pricing, and output pricing across different models, the report assesses changes in DeepSeek V4's performance, cost, and competitive positioning.

• Event-driven analysisImpact assessment of a new product release

  A new product release changes both the model performance curve and the commercialization cost curve.

  Following the release of DeepSeek V4, investors need to assess technical leadership, API pricing changes, potential conditions for price cuts, and the impact on China's AI ecosystem simultaneously.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• DeepSeek V4-pro

  Core release target

  Strengths

  Large parameter scale, long context window, and strong agent and coding capabilities; the report says it has matched or exceeded relevant Opus-4.6 benchmarks.

  Weaknesses

  API pricing has risen sharply versus V3.2, creating clear short-term usage cost pressure.

  Comparison

  Compared with DS-V3.2, model scale and context window have improved significantly; compared with some overseas frontier models, pricing still retains a certain relative advantage.

  Risks

  If price declines fall short of expectations or performance advantages are matched by competing models, commercialization appeal may weaken.

• DeepSeek V4-flash

  Lower-cost version

  Strengths

  Input and output pricing are significantly lower than V4-pro, potentially making it more suitable for cost-sensitive scenarios.

  Weaknesses

  Parameter scale and active parameters are lower than V4-pro.

  Comparison

  Compared with V4-pro, cost is lower but capability positioning may be more lightweight.

  Risks

  If performance cannot meet high-complexity tasks, the scope of application may be limited.

• Ascend 950 Super Node

  Potential cost-reduction catalyst

  Strengths

  If widely available in 2H26, it may support a significant decline in V4-pro pricing.

  Weaknesses

  The timing of availability, supply scale, and actual magnitude of cost reduction remain uncertain.

  Comparison

  Its role is more reflected in inference cost and domestic compute supply, rather than direct comparison of model capability.

  Risks

  Delays in supply progress or insufficient scale would weaken expectations for subsequent price cuts.

• Greater China IT Services and Software sector

  Covered sector affected by changes in AI model capability and cost

  Strengths

  Stronger models and longer context may drive upgrades in software and AI applications.

  Weaknesses

  Higher API pricing may compress the economics of some application scenarios.

  Comparison

  The sector view is In-Line and was not upgraded to a more positive stance because of this event.

  Risks

  Intensifying competition, slower-than-expected cost declines, regulatory disclosures, and customer adoption pace may all affect investment judgment.

Key data

• DS-V4-pro parameter scale1.6T total parameters, 49B active parametersA significant increase versus DS-V3.2's 685B total parameters and 37B active parameters.
• DS-V4-pro context window1000KHigher than DS-V3.2's 128K and on the same order of magnitude as Opus 4.7's 1000K.
• DS-V4-pro API pricingInput Rmb6.5 per million tokens, output Rmb24 per million tokensVersus V3.2, blended input pricing is up about 5x and output pricing about 7x.
• DS-V4-flash parameters and pricing284B total parameters, 13B active parameters; input Rmb0.6 per million tokens, output Rmb2.0 per million tokensPositioned as a lower-cost version.
• Potential price-cut timing2H26DeepSeek said V4-pro pricing may decline significantly once Ascend 950 Super Node is widely available.
• Sector viewIn-LineCoverage scope is Greater China IT Services and Software.

Impact & implications

The release of DeepSeek V4 strengthens the competitive position of Chinese AI models in frontier capabilities, especially in open weights, agent capabilities, and long context. For investment, the short-term impact is more reflected in AI application development costs, model-calling choices, and expectations for China's AI industry chain; the medium-term key is whether domestic compute supply can bring down API prices, thereby expanding use cases and improving application-side economics.

Risks

• V4-pro API pricing is sharply higher than V3.2, which may suppress short-term demand for usage.
• There is uncertainty around the broad availability of Ascend 950 Super Node in 2H26 and the actual extent of cost reduction.
• Competition among frontier large models is intense, and models such as Opus, Qwen, Kimi, and GLM may continue to iterate rapidly.
• The report is an event commentary and does not constitute personalized investment advice; Morgan Stanley discloses business or potential conflicts of interest with multiple covered companies.

What to watch

• Actual developer adoption and usage volume changes for DeepSeek V4-pro and V4-flash.
• Supply progress of Ascend 950 Super Node in 2H26 and its actual impact on API pricing.
• Subsequent benchmark performance of V4 in coding, agents, reasoning efficiency, and long-context tasks.
• Whether Chinese AI application companies can translate stronger model capabilities into improved product revenue and margins.
• Ratings, pricing, and subsequent research updates for companies covered in Greater China IT Services and Software.