Some LLM takeaways for 2025: reasoning as a signature feature, coding agents were useful, subscriptions hit $200/month, and Chinese open-weight models impressed
It's that time. It's been a hell of a year. — At the start we barely had reasoning models. X: Simon Willison / @simonw : Here's my enormous round-up of everything we learned about LLMs in 2025 - the third in my annual series of reviews of the past twelve months https://simonwillison.net/... This year it's divided into 26 sections! This is the table of contents: [image] LinkedIn: William Robertson : I keep up with developments in the GenAI/LLM world exclusively through Simon Willison's weblog. His end-of-year recap is a must read to catch up on what's happened this year: … Arvind Balasundaram : “If 2023 and 2024 were defined by AI prophecy—that is, by sweeping claims about imminent superintelligence and civilizational rupture … Simon Willison : I published my third annual roundup of the last twelve months in LLMs. This one has 26 sections, starting with reasoning models and coding agents … Bluesky: @daniloc.xyz : Claude Code alone has a billion dollars in run rate, I had no idea — Imagine telling anyone in 2020 that you could build a unicorn scale product on the command line — simonwillison.net/2025/Dec/31/ ... Dare Obasanjo / @carnage4life : Simon Willison has a great summary of LLM progress in 2025: — • Reasoning models have made LLMs useful for web search. — • AI agents now work well for Deep Research and Coding. — • AI image generation went mainstream. — • Vibe coding became a big business. … Mike Hadlow / @mikehadlow.com : Enjoyed @simonwillison.net 's wrap up of the year in AI. His is probably my favourite developer-focussed AI blog. Definitely worth a regular read: — simonwillison.net/2025/Dec/31/ ... Jeremy Morrell / @jeremymorrell.dev : I'm glad Simon puts these together. It's honestly hard to believe how much things have changed in just a year [embedded post] Mastodon: @Khrys@mamot.fr : From prophet to product: How AI came back down to earth in 2025 — https://arstechnica.com/... In a year where lofty promises collided with inconvenient research, would-be oracles became software tools. Forums: Hacker News : 2025: The Year in LLMs Lobsters : 2025: The year in LLMs
Context & Ripple Effects
This roundup extends an earlier 2025 assessment in which multimodal vision became routine and LLM prices fell sharply. It marks a changed competitive baseline: capability gains are now being judged alongside the cost and packaging of access.
The practical emphasis also builds on lessons from teams deploying LLM applications, where workflow fit and prompting constraints mattered as much as model novelty. The reported usefulness of coding and research agents gives that implementation lens more weight.
First-order effects
- Developers and knowledge workers gain more credible support for coding and deep-research workflows as reasoning-oriented models and agents become practically useful.
- Premium LLM providers can charge substantially more for high-end access, while Chinese open-weight models give users and builders additional model options.
Second-order effects
- Model vendors face pressure to demonstrate agent reliability on concrete tasks, not just broad benchmark capability; coding tools become a more important distribution channel for models.
- Higher subscription tiers sharpen the trade-off between hosted premium services and lower-cost or self-managed open-weight alternatives, linking product choice to the prior decline in model pricing.
Third-order effects
- If agent usefulness continues to improve, competition may shift from standalone chat interfaces toward integrated systems that combine reasoning models, tools and workflow context.
- The coexistence of expensive subscriptions and strong open-weight models points to a more segmented LLM market, where inference economics and deployment control matter alongside raw model quality.
The trend: LLMs are moving from broadly capable assistants toward priced, workflow-specific agent products, while open-weight competition widens the supply of capable models.