The problem with year-end AI roundups is that they mistake announcements for actual change. In any given year, announcements are loud, but most are forgotten by the next.
So, let’s establish some criteria first. In this post, “what actually changed” refers only to things that meet at least one of the following:
- Pricing changed — verifiable through public pricing sheets
- Specifications changed — measurable in numbers, like context length or rate limits
- Workflows changed — practitioners are working differently today than they did yesterday
Announcements, demo videos, and roadmaps do not count.
What Actually Changed
1. Context windows grew by orders of magnitude
Context windows for major models have scaled to the million-token mark. This isn’t just spec-sheet bragging; it is a fundamental shift in how we work.
Previously, handling long documents required building custom pipelines to chunk text, chain summaries, and retrieve relevant snippets via search. Now, in many cases, dumping the entire document and asking questions is both more accurate and faster. Whether to implement RAG has shifted from a default requirement to an optional choice.
2. Real-world costs plummeted
Even list prices show clear tiering. The cost per million tokens for flagship models versus small models differs by multiples. On top of this, we now have prompt caching (which drastically lowers input costs by reusing repeating prefixes) and batch processing discounts (offering 50% off for tasks that don’t require immediate responses).
As a result, simply applying three rules—”cheap models for easy tasks, caching for repetitive contexts, and batching for non-urgent jobs”—can slash your bill by an order of magnitude. This means model selection has become a core engineering discipline.
3. Coding tools became a daily reality
Moving far beyond simple autocomplete, tools that read entire repositories, modify multiple files, run tests, and self-correct have entered production environments. The signals are visible not just in the tools themselves, but in the surrounding ecosystem. Job postings now explicitly mention AI tool proficiency, teams are establishing new code review guidelines, and organizations are formalizing procedures for handling “AI-generated code.”
What Only Changed in Words (The Hype)
Fully Autonomous Agents — Demos claiming “just set a goal and it does the rest” surface every month. In practice, human-in-the-loop intervention is still mandatory. The reasons are technical: longer tasks accumulate errors in intermediate decision-making, and trust issues persist around irreversible actions like payments, deletions, or sending emails. Partial automation + human review remains the only realistic architecture.
“Prompting is dead” — While improved models have diminished the value of “hacks” and tricks, the need to clearly articulate requirements remains unchanged. What disappeared were the magic spells, not the ability to write clear specifications.
AI Search Replaced Traditional Search — While user numbers have certainly grown, it is not a replacement. Traditional search is still heavily used for queries requiring source verification or action-oriented tasks like transactions and bookings. However, the fragmentation of where information consumption begins is a highly observable shift.
AI Replaced Developers — While data confirms a decline in entry-level hiring, whether the cause is AI or macroeconomic factors remains inconclusive. This is the area where arguments conflating correlation with causation have circulated the most.
Where the Real Impact Will Hit Next Year
These are not mere predictions, but areas where signals are already visible.
An increasing number of sites see AI crawlers outnumbering search crawlers. Looking at raw web logs, it is now common for AI service crawlers to far outpace search engine crawlers. For content creators, the question is shifting from “How do we appear in search results?” to “How do we get cited in AI answers?”
Cost optimization will become a standard operational task. As integrating models has become easier, the resulting bills have caught the attention of finance departments. Model routing, caching, and batching are no longer optional optimizations but core design requirements.
As context windows grow, the challenge becomes “what to leave out.” Just because you can fit everything doesn’t mean you should; doing so hurts both cost and accuracy. Managing long context effectively is the next technical hurdle.
⚠️ The figures and specifications in this post are accurate as of August 2026, and this field moves rapidly. Please check the official pages of each provider for specific pricing and model specifications.
Leave a Reply