Aligning pricing with token usage is a standard competitive move in the market.
While pursuing a domestic listing and launching new models, MiniMax is intensifying its efforts in both technology and capital. On June 1st, MiniMax announced the launch of its new M3 model, emphasizing significant improvements in coding and agent capabilities. The company also initiated a limited-time 7-day, 50% discount promotion for the M3 API. MiniMax stated it would update the model's technical report and open-source the corresponding model weights within the next 10 days.
Industry reactions to the model's release have been mixed. Some users reported faster speeds and good contextual understanding, while many more noted a quicker consumption of tokens. Concurrently, the company revised its Token Plan subscription, effectively indicating a future price increase for the model. In early trading on June 1st, the share price of MiniMax (00100.HK) initially surged over 7% before reversing to a decline, closing down more than 15% at HKD 708 per share, resulting in a market capitalization of HKD 222.1 billion.
According to the official blog, M3 is a native multimodal model that supports image and video input and can operate a computer desktop. For instance, a user could instruct their phone to "open the local ERP client and batch-enter invoice information from this Excel file," and MiniMax Code would autonomously execute this cross-application, cross-file, cross-system task on the computer.
The AI coding product MiniMax Code, updated alongside M3, demonstrates strong autonomy in handling long and complex tasks. Its agent clusters can decompose a large task into multiple concurrent, dynamically adjustable stages, then collaboratively advance the process with capabilities for production, reflection, and correction. The company claims it can "run autonomously for days without human intervention."
MiniMax highlighted that M3 achieves cutting-edge performance in specialized tasks like programming and agent functionality. Official data shows that on the SWE-Bench Pro benchmark for evaluating coding ability, MiniMax M3 outperformed OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro, approaching the level of Claude's Opus 4.7. On the SVG-Bench benchmark, which assesses a model's ability to generate Scalable Vector Graphics, MiniMax M3 surpassed Opus 4.7.
However, numerous users domestically and internationally have expressed skepticism about benchmark scores, arguing that today's programming capabilities increasingly depend on training models with real-world user logic, making benchmarks an incomplete reflection of actual user experience. MiniMax also believes that the next generation of Agent Coding will compete not just on code generation, but more crucially on long-term collaboration, planning capabilities, and human-agent synergy.
To demonstrate this capability, MiniMax tasked M3 with independently replicating the award-winning paper "Learning Dynamics of LLM Finetuning," which studies the learning dynamics during large language model fine-tuning. M3 autonomously ran for nearly 12 hours, successfully conducted the core experiment, observed the "squeezing" effect discussed in DPO experiments, and validated the Extend mitigation method proposed in the original paper.
This replication process required multimodal abilities to interpret the paper's graphs, data, and formulas. Long context windows allowed the paper, code, and experiment logs to be processed in a single instance, while strong programming and agent capabilities enabled the long-threaded, even concurrent, completion of the task.
Enabling a model to handle such complex agent tasks necessitates advancements in context scaling. To address this, MiniMax employed a novel sparse attention architecture called MSA (MiniMax Sparse Attention). Based on this, the M3 model supports up to 1 million tokens of ultra-long context, with a computational cost per token at this length being only 1/20th that of the previous generation model.
Regarding pricing, the API call cost for MiniMax M3 is tiered based on context length. For contexts within 512k tokens, the standard rate is ¥4.2 per million input tokens and ¥16.8 per million output tokens, currently offered at a 50% discount for 7 days. For contexts between 512k and 1 million tokens, the price is ¥8.4 per million input tokens and ¥33.6 per million output tokens.
MiniMax also updated its Token Plan, recommending it for individual developers. According to the company's calculations, at maximum usage, the Token Plan can be over 10 times more cost-effective than pay-as-you-go pricing. At an equivalent price point, the usage allowance from MiniMax is approximately 15 times that of a subscription to the leading overseas model, Claude. The official positioning is to "offer the highest usage allowance among subscription products at the same price point."
However, many users have noted that this move by MiniMax effectively raises model prices, as the previous "high-volume, all-you-can-use" packages are no longer available. Some industry observers view the previous subscription models as largely subsidized. The rapid development of agents has accelerated token consumption, and model providers, unable to sustain the pressure, are forced to raise prices to mitigate losses. MiniMax's adjustment is thus seen as an expected step, with the shift towards token-based billing being a normal competitive adaptation.
MiniMax is currently operating at a loss. Financial results released in March showed that for 2025, MiniMax's revenue was approximately $79.04 million, a 159% year-over-year increase. However, its annual loss widened by 302% to $1.87 billion, with an adjusted net loss of $250 million.
Alongside its technological advancements, MiniMax's capital activities have been frequent. Just last week, the company submitted an IPO tutoring filing report to the Shanghai Securities Regulatory Bureau, initiating the process for a domestic listing. On May 31st, MiniMax released an insider information announcement stating it is continuously evaluating capital markets, including an assessment for listing on the Science and Technology Innovation Board (STAR Market). As of the announcement date, the company has engaged professional advisors to consult on meeting STAR Market listing requirements and has signed a tutoring agreement.
Since its debut on the Hong Kong Stock Exchange in January 2026, MiniMax's stock performance has been closely watched. Its IPO price was HKD 165, and it reached a high of HKD 1330 on March 18th. Following the June 1st early trading session decline of 15.7% to HKD 708 per share, the stock still maintains a gain of over three times its issue price.
Comments