AdvertiseFee Deal Desk

Macro

Alibaba's Qwen4-Max Trails Top Claude Model by 22 Points on Independent Index

Alibaba unveiled Qwen4-Max, its new flagship model, on November 16, and the company's own benchmarks placed it level with Anthropic 's Claude Opus 5.5.

Be a creator
November 18, 2026, 04:12 AM UTC4 min read
AI SummaryAI
  • Alibaba released flagship model Qwen4-Max on November 16, claiming parity with Claude Opus 5.5.
  • Qwen4-Max scored 43 on the Artificial Analysis Intelligence Index, 22 points behind Fable 5.2.
  • Anthropic says full-reasoning training reached 80.4% on SWE-bench Verified versus 71.8% with summaries.
  • Anthropic's October 12 defenses cut stealable reasoning traces by roughly 90% to 99%.

Qwen4-Max Scores 43 Against Claude's 65

Alibaba unveiled Qwen4-Max, its new flagship model, on November 16, and the company's own benchmarks placed it level with Anthropic's Claude Opus 5.5. Independent evaluation tells a different story. Artificial Analysis, a third-party benchmarking service, assigned the model 43 on its Intelligence Index, 16 points below Opus 5.5 at 59 and 22 points below the strongest Claude model, Fable 5.2, at 65.

The shortfall has revived a specific theory among AI observers: that withholding Claude's reasoning traces has cut off the distillation channel Chinese labs used to close capability gaps. Chubby, an AI commentary account followed by roughly 160,000 people, made that argument in a November 16 post, saying Qwen looked equal to Opus 5.5 on paper while the index showed a very different picture, and that the distillation cut-off illustrated the failure. The claim is not settled fact; it remains an interpretation of one benchmark reading.

The timeline behind the defenses is documented. Anthropic removed visible reasoning from the Claude Code interface after February, citing lower latency, and dropped thinking summaries from Fable 5's web and app versions in July without an explanation. Its October 12 distillation defense notice stated the purpose plainly: reasoning records, the step-by-step internal process a model runs before answering, are the most valuable material for distillation. Anthropic reported that a model trained on complete reasoning scored 80.4% on SWE-bench Verified against 71.8% for one trained only on summaries, and said the new protections cut stealable reasoning by roughly 90% to 99%. Fable 5.1 and Opus 5.5 now ship with preserved thinking, and since August 31 newly opened API accounts cannot rewrite Claude's prior context.

Alibaba is a named object of Anthropic's suspicion. A June 10 letter to US senators alleged the company used about 25,000 fake accounts to drive roughly 28.8 million conversations, and a September threat intelligence report put the count at about 151 million conversations between May and July. Alibaba has not responded publicly, and the defense notice itself did not name the company.

Amazon and Google to Hold 28.4%

Investor demand for Anthropic's initial public offering has kept building. A prospectus updated on November 17 shows Amazon and Google on track to hold 14.7% and 13.7% respectively once the listing and a companion private placement are complete, a combined 28.4% stake. The two companies are also Anthropic's core cloud suppliers: the filing commits Anthropic to spend at least $110 billion with Amazon and at least $121.3 billion with Google on compute over the contract periods, so each giant profits through equity appreciation and long-term cloud revenue at once. Amazon additionally holds convertible notes that would lift its theoretical stake to 23.8% if fully converted, although ownership caps prevent a full conversion at listing. Neither giant's shares carry ordinary voting rights; the founding team retains 50.1% of votes through a special share structure.

The subscription picture has reversed the valuation debate. In early November some investors argued the valuation might need to fall from $1.5 trillion to about $1 trillion. Orders now approach seven times the shares on offer, and reports indicate the company is weighing a price above $80 per share, which would value it at roughly $1.5 trillion to $1.6 trillion. The filing still lists the original $70 to $78 range, with final pricing due on November 18 and Nasdaq trading set for November 19.

Existing backers are adding rather than exiting. Sequoia Capital, Iconiq, Lightspeed Venture Partners, Accel and Altimeter Capital Management plan to subscribe for several billion dollars of stock in the offering, according to people familiar with the plans. Lightspeed's cumulative investment exceeds $3.5 billion and Iconiq's tops $6 billion, while Sequoia has committed about $3 billion for a stake near 1%. Accel's total outlay is undisclosed, but its Anthropic position stands as the largest single investment in the firm's history.

Read together, the two stories describe a company whose market position and market value are being tested in the same week. The 43 score against 65 for Fable 5.2 is a data point on whether withholding reasoning actually slows rivals, and observers themselves frame the Qwen4-Max shortfall as a possible effect of that policy rather than settled proof. The offering book suggests investors are not waiting for that question to close: near-seven-fold demand and a possible price above the $70 to $78 range put the valuation back at $1.5 trillion or more. Final pricing lands on November 18 and the Nasdaq debut follows on November 19, the next checkpoints for both debates.

Primary sources

COINOTAG's editorial and research desk.

AI-Assisted

AI-generated, AI-reviewed, under COINOTAG editorial oversight.

Price-Impacting News

Live
All breaking news
Breaking News