MiniMax Faces Triple Threat: Pricing Backlash, Benchmark Doubts, and a July Unlock
Once the brighter half of Hong Kong's AI twin listing, MiniMax has shed roughly 64% from its March peak, as a self-inflicted pricing controversy, contested model benchmarks, and an imminent flood of unlockable shares converge to stress-test a valuation that was always built on scarcity rather than earnings.
The stock closed at HK$451.8 on June 10, 2026, down from an intraday high of HK$1,238 reached on March 18 — erasing more than HK$2,300 per share in market capitalization in under three months. Its former co-listing peer on the Hong Kong main board, Zhipu AI, has held up comparatively better, closing at HK$1,048 on the same day after touching a record HK$1,993 on May 29, though it too has begun retreating. The divergence between the two stocks, once near-identical in price trajectory at the time of their January 2026 debuts, is now the defining story of China's AI capital markets.
Botched Repricing Ignites Developer Revolt
The immediate catalyst for MiniMax's latest leg down was a pricing overhaul that coincided with the June 1 launch of its new foundation model, MiniMax M3. The company simultaneously announced a structural shift in billing — abandoning per-use or time-period subscriptions in favor of token-based pricing — while quietly canceling the RMB 29/month (approximately US$4.03) Starter plan for existing subscribers without prior notice. The new floor subscription stands at RMB 49/month (US$6.81).
Under the revised API schedule, input tokens are priced at RMB 4.2 per million tokens for contexts up to 512k, rising to RMB 8.4 per million for the 512k–1M range; output tokens are priced at RMB 16.8 and RMB 33.6 respectively. Users on social media reported that token consumption for equivalent tasks had increased materially, accelerating credit burn beyond expectations.
MiniMax subsequently issued a public apology, acknowledging it had failed to communicate the changes in advance and had handled legacy user quotas poorly. To contain the damage, the company introduced a permanent 50% promotional discount on M3 API pricing — bringing input token costs to US$0.30 per million tokens and output token costs to US$1.20 per million tokens. At those levels, M3 undercuts Claude Sonnet 4.5 (US$3.00 input / US$15.00 output) by a factor of 12.5x on output, and GPT-5.2 (US$1.75 input / US$14.00 output) by 11.7x. However, it remains more expensive than domestic rivals: DeepSeek-V4-Flash charges just US$0.14 per million input tokens and US$0.28 per million output tokens.
The pricing episode is symptomatic of a deeper structural tension. With GPU supply unable to keep pace with surging token call volumes, repricing serves a dual function: rationing compute capacity and buying margin breathing room for a company that reported an adjusted net loss of approximately US$251 million (RMB 1.73 billion at reference rate) in fiscal year 2025. The promotional discount, meanwhile, is a short-term developer retention tool — one that defers, rather than resolves, the underlying unit economics problem.
M3 Benchmark Claims Draw Independent Scrutiny
MiniMax's assertion that M3 achieved a 59.0% score on SWE-Bench Pro — surpassing GPT-5.5 (58.6%) and Gemini 3.1 Pro (54.2%), and approaching Claude Opus 4.7 — was met with immediate skepticism from the global technology press. TechTimes, Startup Fortune, and DataNorth each flagged within hours of the announcement that the results were self-reported and that portions of the evaluation used external agent scaffolding tools, including Claude Code and Mini-SWE-Agent. Independent third-party verification remains pending.
This is not MiniMax's first credibility challenge. In February 2026, Anthropic publicly accused MiniMax, along with DeepSeek and Moonshot AI, of conducting what it described as an "industrial-scale distillation attack" on its Claude models — alleging that MiniMax alone was responsible for more than 13 million interactions across approximately 24,000 synthetic accounts. MiniMax did not publicly respond. The reputational overhang has been noted by investors: Wang Jie, a backer of both Moonshot AI and Moore Threads, told Caixin that "market confidence in MiniMax pulled back after the Anthropic distillation allegation."
The combined effect — unverified benchmark claims layered on top of an unresolved distillation controversy — creates a credibility discount that promotional pricing alone cannot offset.
July Lock-Up Expiry Threatens to Structurally Reprice the Float
Beyond the immediate operational noise, the more structurally significant risk materializes in July 2026. According to analysis by China International Capital Corporation (CICC), approximately 63% of MiniMax's Hong Kong-listed share capital becomes eligible for sale on July 9 — of which financial investors account for more than one-third. Zhipu AI faces a smaller but still meaningful unlock on July 8, with approximately 11.6% of its share base freed, predominantly held by state-backed cornerstone investors.
UBS Securities China internet analyst Xiong Wei has noted that both companies have traded at elevated multiples partly because listed AI model pure-plays remain globally scarce, and partly because low free-float artificially suppressed liquidity-adjusted valuations. That scarcity premium is eroding rapidly. OpenAI and Anthropic have both reportedly filed confidential IPO applications. Among Chinese peers, StepFun is expected to file a Hong Kong prospectus imminently, while Moonshot AI — which had previously ruled out fundraising and listing — has reopened a new financing round at a pre-money valuation of US$30 billion and is reported to be dismantling its VIE and red-chip structure, widely interpreted as preparation for a Hong Kong listing.
As the investable universe of AI model companies expands, the scarcity premium that inflated both MiniMax and Zhipu AI at IPO will compress. The July unlock compounds this by introducing a large, motivated seller base at a time when the stock has already lost significant ground.
Revenue Mix Shift Defines the Long-Term Thesis
MiniMax reported total revenue of US$79.04 million in fiscal year 2025, up 158.9% year-on-year. Revenue from its B2B open platform business reached US$25.96 million, representing 32.8% of total revenue and growing 197.8% year-on-year. More than 70% of total revenue was derived from international markets across more than 200 countries, a geographic profile that differentiates MiniMax from most Chinese AI peers.
At listing, this consumer-led, globally diversified revenue base was viewed as a relative strength — more tangible than the enterprise pipeline narratives offered by competitors. But as the dominant commercial narrative in AI has pivoted toward Agentic applications, Vibe Coding, and enterprise software deployment, the market has begun to re-rate companies with heavier B2B DNA more favorably. The result is not that MiniMax's business has deteriorated, but that its existing revenue mix is being discounted against a new valuation framework.
MiniMax co-founder and COO Yuan Yeyi disclosed in late May that the company's user base had exceeded 300 million, that enterprise and developer clients had surpassed one million — a fivefold increase in six months — and that annualized recurring revenue had doubled over the preceding two months. Critically, Yuan indicated that enterprise revenue had reached parity with consumer revenue, suggesting the B2B ramp is accelerating faster than the headline mix implies.
A-Share IPO Push Signals Funding Urgency
On May 31, MiniMax announced via a Hong Kong Stock Exchange filing that its board had resolved to explore a preliminary proposal to issue RMB-denominated shares, and that it had engaged advisors and signed a sponsorship agreement in connection with a potential STAR Market listing on the Shanghai Stock Exchange.
Zhipu AI filed a parallel announcement the following day, disclosing plans to raise RMB 15 billion (US$2.08 billion) on the STAR Market — allocating RMB 12 billion (US$1.67 billion) to general-purpose AI foundation model development, RMB 2 billion (US$278 million) to its MaaS platform, and RMB 1 billion (US$139 million) to working capital.
The dual-listing push reflects a straightforward capital imperative. With both companies still burning cash — Zhipu AI posted an adjusted net loss of RMB 3.182 billion (US$442 million) in 2025 alongside MiniMax's RMB 1.73 billion loss — and with the AI industry's infrastructure spending cycle showing no signs of abating, broadening access to domestic A-share capital markets is less a growth strategy than a financial necessity.
The deeper question that neither the A-share filing nor the M3 launch has answered is the one investors are increasingly pressing: at what point does the high-investment, low-output model of frontier AI development produce a self-sustaining business? For MiniMax, the path forward requires not just closing the gap between consumer scale and enterprise monetization, but doing so before the July unlock, the scarcity premium compression, and the ongoing credibility questions compound into a structurally lower valuation floor.
Related Coverage:
MiniMax M3 Debuts With 9.4X CUDA Acceleration and Autonomous Model Training