Some 2025 takeaways in LLMs: reasoning as a signature feature, coding agents were useful, subscriptions hit $200/month, and Chinese open-weight models impressed
This is the third in my annual series reviewing everything that happened in the LLM space over the past 12 months.
Context & Ripple Effects
The prior year’s review described multimodal vision becoming routine while model prices fell; this year’s account shifts the emphasis from broad capability access to differentiated reasoning, agent usefulness and premium service tiers. A separate year-end review likewise identified Claude Code as an early convincing agent example, reinforcing that coding had become a practical test case rather than merely a demo category.
Earlier coverage noted that LLM-assisted coding could be difficult and unintuitive, even as it exposed new ways to explore models’ capabilities. The reported usefulness of coding agents suggests that those hands-on coding experiments matured into a more consequential product benchmark, while impressive Chinese open-weight models broadened the set of credible model sources.
First-order effects
- Developers and teams evaluating LLM tools now have stronger evidence to assess reasoning and coding-agent performance as distinct product capabilities, rather than treating chat quality as the sole criterion.
- Providers can support higher-priced subscription tiers where users perceive agentic coding and reasoning workflows as materially useful; Chinese open-weight model makers gain visibility as credible alternatives.
Second-order effects
- Closed-model vendors face pressure to justify premium plans through measurable workflow value, while open-weight models give developers more options for balancing capability, control and operating cost.
- Coding-tool vendors and enterprise buyers will increasingly compare products by the cost and reliability of completed tasks, not just model access—an instance of lessons from real-world LLM application building.
Third-order effects
- If useful agents continue to concentrate value in repeatable work, LLM competition may shift from generic model subscriptions toward integrated workflow products and the infrastructure that runs them.
- The combination of premium hosted tiers and capable open-weight alternatives could make inference economics and distribution—not raw model availability alone—more decisive sources of market power.
The trend: LLMs are moving from broadly accessible conversational models toward a segmented market where reasoning, agents, deployment choice and cost per useful task determine adoption.