JPMorgan: Zhipu's Price Hike Succeeds While MiniMax's Price Cut Fails — Pricing Power Determines Valuation Divergence
N.R. Finch
J.P. Morgan on June 12 used pricing power as the deciding variable, maintaining Zhipu at Overweight while downgrading MiniMax to Neutral — the two companies ran the same experiment, raising prices after a new model launch, and got opposite results, splitting their valuations.
Same experiment — why did the results diverge?
Zhipu doubled its API prices year-to-date. Customers stayed, and usage kept growing. This means → the market accepted the premium. J.P. Morgan calls this "price-setter behavior."
MiniMax launched its flagship M3 at roughly twice the price of its predecessor M2.7, then permanently cut that price by 50% about a week later, falling back near M2.7 levels.
In plain terms = Zhipu raised prices and customers kept buying — the product is worth it. MiniMax raised prices and had to retreat — the market rejected the premium. Same pricing experiment, opposite verdicts.
Where does the capability gap lie?
Zhipu has kept refreshing China's state-of-the-art (SOTA) — the best-performing frontier model — through GLM-5 and 5.1. Even after Kimi K2.6 and DeepSeek V4 launched, GLM-5.1 still ranks among the top Chinese models on Code Arena and WebDev Arena.
Since launching M2, MiniMax has not reclaimed SOTA status in China in subsequent iterations. This means → its model is no longer the best in class, so a price hike loses its hardest support.
This reflects a pattern: while AI demand outstrips inference supply, only the model that consistently leads on capability earns the right to charge a premium — ranking and pricing power are directly linked.
How low has DeepSeek pushed the "good enough" price floor?
Blended token prices among China's major LLM providers (assuming 80% cache-hit rate, 10:1 input-to-output ratio) show a clear tiering: Qwen3.7-Max at roughly ¥7.2 per million tokens, GLM-5.1 at ¥5.45, MiniMax M3 post-cut at roughly ¥1.45, DeepSeek V4 Pro at roughly ¥1.11, and V4 Flash at just ¥0.38.
After the cut, MiniMax sits in the same price band as DeepSeek. In plain terms = it chose to compete in the cheapest tier of the market.
J.P. Morgan argues this is most damaging for models positioned on value-for-money — DeepSeek squeezes from the price end, SOTA models squeeze from the high-value-task end. Pressure from both sides at once.
How wide is the valuation gap?
J.P. Morgan assigns Zhipu a premium valuation at 57× estimated 2027 price-to-sales, raising the target price from HK$950 to HK$1,400.
MiniMax's target price drops sharply from HK$1,100 to HK$400, at 29× P/S — grouped with "anchor-priced providers."
The permanent 50% price cut on M3 directly widens loss forecasts: adjusted net losses for 2026–2028 rise from US$309 million, US$596 million, and US$512 million to US$432 million, US$940 million, and US$972 million respectively.
What three tests must a company pass to earn a premium valuation?
J.P. Morgan's framework: premium valuation requires passing three tests simultaneously — repeated SOTA delivery, validated pricing power, and sustainable workflow adoption.
The report stresses that list prices are continuous, public, and set by the party that best understands its own demand curve — making them a harder-to-fake market signal than benchmarks.
In plain terms = benchmarks can be gamed, but pricing cannot — daring to charge a premium and having customers keep buying is the hardest evidence there is.
What would flip either rating?
MiniMax's path back to Overweight: launch a next-generation flagship at a premium price and hold that price for a full quarter, or achieve a China-frontier capability reset validated by both third-party benchmarks and user data.
Zhipu's downside trigger: a DeepSeek release that forces GLM's premium tier into a sharp price cut, or usage attrition that exposes demand elasticity at the doubled price point.
Near-term catalysts center on GLM and M-series launch windows plus monthly third-party leaderboard updates. This means → whether the valuation split holds will be tested in the next model-release cycle.
Content is for reference only, not financial advice.