DeepSeek V4 and Qwen 3.8: How China's AI Pricing War Reshaped the Global Frontier in 2026

Pricing (per million tokens)

DeepSeek also ran a 75% discount promotion on V4-Pro through early May 2026 . For comparison, GPT-5.5 output pricing is ~$30/MTok, making V4-Flash roughly 107× cheaper on output .

Performance and Competitive Positioning

DeepSeek V4's benchmark results tell a nuanced story:

• Codeforces: V4-Pro-Max scores 3206 — the highest rating of any tested model, ahead of GPT-5.4 (3168) .
• SWE-bench Verified: 80.6%; SWE-bench Pro: 55.4% (vs. GPT-5.5's ~57.7% on SWE-bench Pro) .
• GPQA Diamond: V4-Pro-Max scores 90.1% — competitive with GPT-5.4 xHigh (93.0%) and Opus-4.6 Max (91.3%) .
• Terminal-Bench 2.0: V4 at 67.9% vs. GPT-5.5 at 82.7%, a notable gap in agentic coding .

Important caveat from NIST CAISI evaluation (May 2026): DeepSeek's self-reported benchmarks claim parity with GPT-5.4 and Opus 4.6, but CAISI's independent evaluation found V4 "substantially less capable" on non-public benchmarks, suggesting some self-reported scores may be inflated . Artificial Analysis's Intelligence Index rates GPT-5.5 (high) at 53 vs. DeepSeek V4 Pro (Reasoning, High Effort) at 41* — GPT-5.5 is more intelligent and faster (69 vs. 56 tok/s), but DeepSeek is dramatically cheaper ($0.18 vs. $4.35/MTok under reasoning mode) .

Alibaba Qwen 3.8: A Preview with Bold Claims and No Verifiable Data

Alibaba's Qwen team previewed Qwen 3.8 Max at the World AI Conference in Shanghai on July 19, 2026 . It is the team's first flagship model above 1 trillion parameters.

Known Specs

Pricing

No pay-as-you-go pricing has been published. Access is currently via subscription through Qwen Chat and select API partners. BenchLM lists Qwen 3.8 Max Preview pricing as "Not listed" . Some third-party trackers show auto-pricing at $1.50/$5.00 per million input/output tokens, but this is not confirmed by Alibaba .

Performance: A Marketing Claim with Zero Verifiable Scores

Alibaba states Qwen 3.8 is "second only to Fable 5" (Anthropic's top Claude model), which would place it ahead of GPT-5.5 and DeepSeek V4 in Alibaba's own ranking . Gains are claimed in coding, professional productivity, full-stack development, data analysis, and office workflows over Qwen 3.7 Max .

However, no independent benchmarks have been published. BenchLM, a benchmark tracking site, tracks 321 benchmarks for Qwen 3.8 Max Preview with 0 verified results as of July 20, 2026 . As one analysis put it, "Qwen 3.8 is worth testing, but it is too early to treat it as a stable production replacement for Kimi K3, GPT-5.6 Sol, or Claude Fable 5" .

For context, Qwen's predecessors set a strong foundation: Qwen 3.6-Max-Preview (April 20, 2026) was a 35B total / 3B active parameter text-only model with 262K context that claimed #1 on six coding and agent benchmarks . Qwen 3.7-Max (May 2026) expanded to a 1M token context window with reasoning agent capabilities . But neither was generally considered frontier-level outside of agentic coding.

Competitive Landscape: Who Leads and Who Sets the Price

DeepSeek V4 vs. GPT-5.5

• Pricing advantage: DeepSeek V4 is 97–107× cheaper than GPT-5.5 on output pricing .
• Coding strength: V4 leads on competitive programming (Codeforces, LiveCodeBench), while GPT-5.5 leads on applied agentic coding (Terminal-Bench 82.7% vs. 67.9%, SWE-bench Pro 58.6% vs. 55.4%) .
• General intelligence: Third-party evaluations (Artificial Analysis, NIST CAISI) consistently place GPT-5.5 ahead .
• Context window parity: Both offer 1M-token context, but GPT-5.5 is the first OpenAI model where the full 1M window is genuinely usable without degradation .

Qwen 3.8 vs. Claude Fable 5 and GPT-5.5

Alibaba explicitly positions Qwen 3.8 as "second only to Fable 5" . If this claim holds up under independent evaluation, it would place Qwen 3.8 ahead of GPT-5.5 and well ahead of DeepSeek V4. But with zero verified benchmarks, this remains a marketing claim, not an established fact. Anthropic's Fable 5 appears to be the current recognized frontier leader per Alibaba's own framing .

The Pricing War Has Changed Everything

DeepSeek's aggressive pricing has driven a race to the bottom on API costs. V4-Flash at $0.14/M input is radically cheaper than any Western frontier model. Alibaba has not yet set Qwen 3.8 pricing, but Qwen 3.6 was already priced aggressively (~$1.25/$7.50 input/output) .

The open-weight MIT license for DeepSeek V4 means self-hosting is also possible, further compressing margins for closed-source providers . As one analysis put it, the practical effect is that "developers can now access near-frontier capability (on coding tasks) at a fraction of the cost of GPT-5.5 or Claude."

Key Takeaways

• DeepSeek V4 is a shipped, open-weight, 1.6T MoE model that sits approximately 10–20% behind GPT-5.5 on general intelligence but leads on competitive coding and is orders of magnitude cheaper. Independent NIST CAISI evaluation tempers its self-reported benchmark scores.
• Qwen 3.8 is a preview-phase 2.4T MoE model that Alibaba claims is "second only to Fable 5" — but with zero verified benchmarks and no published pricing, this is currently aspirational, not established.
• GPT-5.5 remains the overall intelligence leader (Artificial Analysis Index 53 vs. 41 for DeepSeek) and the leader in agentic coding (Terminal-Bench 82.7%), but costs 100× more.
• The Chinese AI pricing war has permanently lowered inference costs globally, making frontier-competitive capabilities accessible to far more developers and forcing Western providers to compete on both performance and price.