PraxonCoverage of this industry leans on four terms constantly: reasoning, agents, multimodal, inference cost. Most articles use them without defining them, on the assumption that the reader already tracks the field closely. Here's what each one actually refers to technically, and what changed under the hood to produce it.
Chain-of-thought reasoning at inference time
Early chat models generated an answer in a single forward pass: read the prompt, emit tokens, done. Reasoning models insert an intermediate step, an extended chain-of-thought sequence generated before the final answer, where the model works through sub-steps and can revise its own intermediate conclusions. That's the mechanism behind the "thinking" trace some interfaces now show you before the answer lands.
The measurable effect shows up on multi-step tasks: math proofs, debugging, anything requiring several dependent steps in sequence. It's additional inference-time compute spent checking work, not a new kind of intelligence. It reduces a specific, well-documented failure mode without eliminating error generally.
Tool use and function calling
For most of this field's history, a model's output was text and nothing else. The current shift is function calling: a model can now emit a structured call to an external tool, an API, a code interpreter, a browser, receive the result, and incorporate it into the next step. That's the technical substance behind the word "agent." Instead of describing the code you'd need, it invokes an interpreter, runs the code, reads the stack trace if it fails, and revises.
That's also where the risk profile changes. A model confined to text can only be wrong in words. A model with function-calling access can execute the wrong action against a live system. Most of the serious engineering work right now sits in permissioning and sandboxing what a model is allowed to call, not in making the underlying model smarter.
Multimodal transformers
"Multimodal" means the same transformer architecture that tokenizes text now tokenizes other modalities too: image patches, audio spectrograms, video frames, all projected into the same embedding space the model reasons over. Practically, that means you can hand a model a screenshot of a stack trace or a photo of a whiteboard and get a text response grounded in it. The architectural line between "a language model" and "a model that perceives multiple input types" has effectively collapsed into a single unified network.
Falling inference cost
This gets the least press and matters the most. Techniques like distillation, quantization, and mixture-of-experts routing have pushed the cost per token down by orders of magnitude since the first GPT-3-class models shipped. Capability that once required a data center now runs on consumer hardware. That's the actual driver behind the recent flood of AI-powered products: falling marginal cost, not a single breakthrough release.
What the coverage leaves out
All four of these are real, benchmarked improvements. None of them make the underlying model infallible. A reasoning model can chain its way to a confidently wrong conclusion just as easily as it can catch an error. An agent with tool access can execute the wrong action with the same fluency a chatbot uses to state the wrong fact, except now it might actually send the email instead of drafting one you'd catch first. Lower cost and higher capability don't reduce the error rate. They just mean more systems shipping with that error rate built in, at greater scale.
The takeaway isn't "AI got smarter, so trust it more." It's that the judgment calls around where to apply it now matter more than they did two years ago, not less.