Two reports from the same day paint a shared picture: the leaps in AI models are smaller where they have been most visible in recent years — in programming. At the same time, competition is shifting to another field: the cost per completed task.
Moonshot AI, provider of the Kimi models, explicitly cites the fact that little progress is still being made in coding as the reason for a strategic realignment. Alibaba, by contrast, presents a new flagship with Qwen4-Max, which surpasses its predecessor in all benchmarks but whose strongest argument is the price per task.
Moonshot shifts Kimi into the natural sciences
According to its own account, Moonshot AI is aligning its Kimi language models more strongly towards biology, chemistry and pharmacy. The reason: in software development, the gain from one model generation to the next has become small. In research, by contrast, the company still sees room for improvement.
One example is the integration into the research assistant BioMaster. There, Kimi K3.1 Pro analysed gene data from almost 3,700 cells in 90 instead of 240 minutes. That is a significant shortening of a concrete, clearly defined analysis — and that is exactly how the figure should be read.
Because the caveat comes with it: the models still do not achieve independent breakthroughs. AI accelerates work steps that experts plan, assess and take responsibility for anyway. It replaces neither the formulation of the question nor the interpretation of the results.
Qwen4-Max: more parameters, lower cost per task
Alibaba is taking a different route. Qwen4-Max works with 2.4 trillion parameters but activates only 71 billion of them per token. Behind this is an architectural principle in which only part of the model computes for each processing unit — each token, i.e. each text or code component. This keeps the effort per request low, even though the overall model is very large.
In addition, a new hybrid architecture lowers costs, especially for long contexts. Long context means: the model processes large amounts of accompanying information, such as entire codebases, documentation or contract documents, in a single request. This very case used to be expensive.
Alibaba puts the result at 62 percent lower costs per task. In all benchmarks, Qwen4-Max is above its predecessor. In the Intelligence Index from Artificial Analysis, the model lands just behind Kimi K3.1 Pro. Smaller models in the Qwen4 family are to follow as open source.
What the two reports show together
At first glance, the signals contradict each other: one company declares the coding field largely exhausted, while the other releases a model that improves in all benchmarks. But both can be thought of together.
The ranking at the top is close. Qwen4-Max is only just behind Kimi K3.1 Pro in the Intelligence Index — so two models from different providers are not worlds apart. When performance differences shrink, price becomes the decisive distinguishing feature. That is exactly what the 62 percent figure is aimed at.
Moonshot draws a different consequence from this: instead of continuing to run in the narrow race for coding benchmarks, the company is looking for application areas in which the benefit per model generation is still clearly visible. Both strategies confirm the same observation — a pure model upgrade carries less far in everyday development than it did some time ago.
Consequences for the use of AI assistants
For companies that use or want to introduce AI assistants in software development, the relevant question thus shifts. It is no longer which model stands at the top of a benchmark list, but how much a completed task costs and how reliably it succeeds in one's own context.
Three points are practically important here:
- Context beats model choice. An assistant only delivers usable results if it has access to clean codebases, documented conventions and clear task descriptions. If the context is disorganised, even the stronger model helps little.
- Long contexts become more affordable. If the costs for extensive inputs fall, scenarios that were previously too expensive become economically viable — for example the analysis of larger legacy systems or extensive documentation holdings.
- Open-source variants expand the options. Announced smaller models in the Qwen4 family can be interesting for companies that are considering their own operating models for data protection or cost reasons.
Equally important is the limit that both reports mark. The BioMaster example shows an acceleration, not a replacement of expertise. Transferred to software projects, this means: review, architecture decisions and quality assurance remain human tasks. Those who scale them along with everything else actually gain time through AI. Those who omit them merely shift effort to later.
Assessment
For web and software projects, the lever shifts from the next model update to processes, data quality and a sober cost calculation per task. Anyone using AI assistants should define metrics — throughput time, rework, cost per ticket — instead of reacting to benchmark placements.
The technical basis is becoming more interchangeable: what matters is an architecture that allows the model provider to be changed without rebuilding the entire workflow.