DeepSeek V4 release strengthens the competitiveness of China’s frontier large models, but API price hikes raise near-term usage costs

Overview

This report focuses on the release of DeepSeek V4 within China’s AI pathway. It positions DeepSeek V4 as an open-weight MoE large language model, highlighting its breakthroughs in agent capabilities, world knowledge, and reasoning efficiency, and compares it with frontier models such as Opus 4.7, Qwen 3.6, Kimi 2.6, M2.7, and GLM-5.1.

Core views

The core view is that DeepSeek V4 improves the competitive positioning of China’s frontier large models, especially as V4-pro delivers significant upgrades in parameter scale, context window, and capability benchmarks; however, the upgrade comes with sharply higher API pricing, which may raise near-term costs for developers and enterprises. If Ascend 950 Super Node is widely deployed in 2H26, V4-pro pricing could have substantial room to decline.

Analysis framework

The report uses a cross-model comparison framework for frontier large models, comparing release timing, total parameters, active parameters, context window, input pricing, and output pricing, while incorporating DeepSeek’s disclosed pricing outlook to assess the trade-off between model capability improvements and commercialization costs.

Methodology notes

• Cross-model comparisonFrontier LLM capability and cost comparison

  Compare different frontier large models through parameter scale, active parameters, context window, and API pricing.

  The report’s table lists models such as DS-V3.2, DS-V4-pro, DS-V4-flash, Opus 4.7, Qwen 3.6, Kimi 2.6, M2.7, and GLM-5.1 to observe capability upgrades and changes in usage costs.

• Industry viewMorgan Stanley Industry View

  In-Line means the analyst expects the industry coverage universe to perform broadly in line with the relevant broad market benchmark over the next 12-18 months.

  The report states that the Greater China IT Services & Software industry view is In-Line and does not directly translate the DeepSeek V4 release into an overweight conclusion for the industry.

Asset mapping & comparison

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

• DeepSeek V4-pro

  The core version behind the upgrade of China’s frontier large models

  Strengths

  1.6T total parameters, 49B active parameters, a 1000K context window, and it has reached or exceeded Opus-4.6 on coding and agent benchmarks.

  Weaknesses

  API pricing is meaningfully higher than V3.2, with input and output prices rising 5x and 7x, respectively.

  Comparison

  Versus DS-V3.2, model scale and context window have expanded significantly; versus Opus 4.7, the report emphasizes its competitiveness on certain coding and agent benchmarks.

  Risks

  High usage costs may limit developer adoption speed, while price improvement depends on progress in compute supply such as Ascend 950 Super Node.

• DS-V4-flash

  The low-cost version in the DeepSeek V4 series

  Strengths

  Input at Rmb0.6/million tokens and output at Rmb2.0/million tokens, with costs significantly below V4-pro while still offering a 1000K context window.

  Weaknesses

  Its parameter scale is 284B total parameters and 13B active parameters, so its capability positioning may be below V4-pro.

  Comparison

  More focused on cost efficiency versus V4-pro, while offering a longer context window versus DS-V3.2.

  Risks

  If performance cannot meet the needs of high-complexity agent or reasoning scenarios, enterprises may still need to use the more expensive V4-pro.

• Ascend 950 Super Node

  AI compute infrastructure that could potentially affect V4-pro’s cost curve

  Strengths

  If widely available in 2H26, it could drive a meaningful decline in V4-pro pricing.

  Weaknesses

  The report only provides a forward-looking availability expectation and has not yet offered validation of actual deployment scale or the cost curve.

  Comparison

  Its significance lies in potentially easing inference costs for high-end models, rather than directly comparing with large-model capability parameters.

  Risks

  Supply timing, deployment scale, software ecosystem adaptation, and actual inference efficiency could all affect whether price cuts are realized.

Key data

• DS-V4-pro parameter scale1.6T total parameters, 49B active parametersA clear increase versus DS-V3.2’s 685B total parameters and 37B active parameters.
• DS-V4-flash parameter scale284B total parameters, 13B active parametersAlso released in April 2026, with a 1000K context window.
• Context windowBoth DS-V4-pro and DS-V4-flash are at 1000KAbove DS-V3.2’s 128K.
• V4-pro input pricingRmb6.5/million tokensOn a 1:1 blended basis of cached and non-cached usage, up 5x versus V3.2.
• V4-pro output pricingRmb24/million tokensUp 7x versus V3.2.
• DS-V4-flash pricingInput Rmb0.6/million tokens, output Rmb2.0/million tokensBelow V4-pro, reflecting its positioning as a cost-friendly version.
• Industry viewIn-LineApplies to Greater China IT Services & Software coverage.
• Potential price-cut timing2H26DeepSeek says V4-pro pricing could decline meaningfully once Ascend 950 Super Node becomes widely available.

Impact & implications

The release of DeepSeek V4 could enhance the visibility of China’s large-model ecosystem in frontier capabilities and continue to drive the evolution of AI applications, agents, and the developer ecosystem. From an investment perspective, stronger capabilities support the AI software and applications narrative, but higher API pricing means near-term commercialization cost pressure still needs evaluation; compute supply and the availability of domestic AI infrastructure will be key variables for future price declines and higher penetration.

Risks

• Sharp API price increases may suppress near-term usage volume and developer adoption speed.
• V4-pro price cuts depend on the broad availability of Ascend 950 Super Node in 2H26, creating execution uncertainty.
• This is an industry update and deep-dive report and does not provide a target price or clear buy/sell recommendation for any single company.
• Morgan Stanley discloses that it has or may have investment banking, market-making, or other service relationships with multiple covered companies, and investors should consider potential conflicts of interest.
• Large-model benchmark performance may differ from real commercial use-case performance.

What to watch

• Adoption of DeepSeek V4-pro and V4-flash in real enterprise applications, agent use cases, and coding scenarios.
• Whether V4-pro API pricing declines meaningfully as Ascend 950 Super Node adoption expands.
• Progress in China’s AI compute supply, inference costs, and domestic infrastructure ecosystem.
• How Greater China IT Services & Software companies convert stronger model capabilities into revenue, margins, and product competitiveness.
• Whether Morgan Stanley subsequently adjusts ratings, prices, or industry views for relevant covered companies.