AussieBytes Blog ◖ᵔᴥᵔ◗

The Gap Is Widening: 2025 AI Recap and 10 Predictions for What's Next (Part 1)

2025 was an incredible year of AI progress. I wanted to take the time to map out the changes that mattered this year, summarise the key trends, and take a punt on where things might land for 2026.

I've also asked my vibe-coded LLM council to provide its predictions for 2026, which I have included, which will be a bit of fun to track as we go into next year.

There is a bit to cover off, so we are going to do this in two parts - a 2025 recap in part 1 and 2026 predictions in part 2 here.

TL;DR

2025 in summary:

Part 1: The 2025 Recap

A colorful timeline outlines key events and release dates for various technology companies from Q1 to Q4 of 2025. Events include announcements, releases, and summaries with company names like Google, Meta, and others highlighted in different colors across the timeline. AI 2025 Timeline

Intelligence Progress: A Year of Saturation

At the start of 2025, GPT-4-class models (notably GPT-4o) were predominant. By year-end, we've seen an explosion of capability:

Models at the start of the year could handle many simple creative and document-writing tasks, but had high hallucination rates and an inability to say "I don't know." As we close out 2025, almost all key intelligence benchmarks across coding, general knowledge, and graduate-level science, maths, and physics are saturating - and fast, with hallucination rates significantly lower and certain models starting to perform well on "I don't know responses" and safety protocols.

A line graph illustrating the progression of various artificial intelligence models over time, tracking their intelligence index scores against release dates. Different colored lines represent multiple organizations, showing trends and comparisons in model performance from 2015 to 2023. Artificial Analysis - Intelligence Index, Dec 2025

The real story was behind the benchmarks - models became dramatically more dependable when wrapped in proper harnesses: tools, eval loops, critique, tests, and supervision. This has enabled software engineers to shift from writing code to supervising agents that write it for them – focusing on outcomes over keystrokes, especially for commodity work across the most popular languages, systems and patterns.

The Competition Heats Up

OpenAI has led most of the year, but Anthropic and Google have made significant ground, if not surpassed OpenAI in some areas. Open-source options have proliferated and provide a compelling choice and price differential for those comfortable with these models' provenance.

The DeepSeek Moment (January 2025):

The year's first major shock came when Chinese startup DeepSeek released DeepSeek-R1 – an open model that DeepSeek claimed matched Western closed models at a fraction of the training cost (though analysts debated what that figure captured). It quickly ranked among the top models on major public leaderboards and proved the U.S. was "not as far ahead in AI" as assumed - a genuine "Sputnik moment" for the industry. DeepSeek has followed up with some compelling new versions and innovation on mathematics.

On the open-source side, Chinese and Asian model labs are now dominating:

A bar chart comparing the Artificial Intelligence Index by open weights and proprietary weights for various companies. The black bars represent proprietary weights, while the blue bars illustrate open weights, showing varying values across the listed companies. Artificial Analysis - Intelligence Index by Open Weights vs Proprietary, Dec '25

These models offer excellent general and coding capabilities and are rapidly advancing, putting competitive pressure on frontier labs to not just stay ahead, but to ensure their cost-to-value ratio makes sense when developers and enterprises have choice.

OpenAI Goes Open:

In a major pivot in August and going back to its roots, OpenAI released its own open-weight models (gpt-oss-120b / gpt-oss-20b) under Apache 2.0. This landed amid intensifying competition from open-weight models, including DeepSeek.

The Compute Crunch

Unbeknownst to most consumers, there are huge constraints on the ability of labs to access the latest chips and run cost-effective inference at scale. This is compounded by an insatiable demand for tokens by customers, which appears to be ever growing as new capabilities emerge.

A bar graph illustrating estimated global data center capacity demand in gigawatts under a 'continued momentum' scenario, projecting growth from 2022 to 2030. The graph highlights different bars for non-AI and AI workloads, showing significant increases in both categories, especially for AI workload by 2030. Mckinsey, The cost of compute: A $7 trillion race to scale data centers, April 2025

This compute crunch puts pressure on:

  1. Labs' ability to launch new products which require even more inference
  2. Balancing trade-offs across R&D and commercial needs in a highly competitive market, where R&D is crucial to long-term success

We've seen a dramatic increase in CapEx allotments to data centres in 2025 - total AI data centre investment commitments approaching $1 trillion over the next five years:

Power availability is also not evenly distributed – China's grid has significant spare capacity and can connect new demand faster than most Western markets. Memory challenges (SRAM and DRAM supply chain constraints) are also becoming a bottleneck on data centre growth.

Consumer and Enterprise Adoption

ChatGPT has reached ~800 million weekly active users – up from zero just three years ago, making it one of the top 10 most visited websites globally. Both OpenAI and Anthropic have made major gains in the enterprise space - their enterprise app offerings and API spaces have been major drivers of growth.

A bar graph showing the traffic share of various generative AI platforms over the past year, with time intervals indicated as "12mo ago," "6mo ago," "3mo ago," "1mo ago," and "Today." Each bar is color-coded to represent different platforms, including OpenAI, Google Gemini, and others, with the majority of the traffic share assigned to OpenAI.

SimilarWeb, Dec '25

The Data Wall

One big trend is the data constraint powering model development. By some estimates, there is enough data to support two major training runs over the next 18–24 months across the major model labs. Without major breakthroughs in the interim, we won't know how model progress will continue.

A line graph showing projections for the stock of public text and data usage from 2020 to 2028. It includes multiple lines representing different models (like GPT-3 and PaLM) and their respective estimated amounts of human-generated text, along with confidence intervals. Several shaded areas indicate key data points for various years and trends in data growth. Epoch AI - Projections of the stock of public text and data usage, 2024

This still represents a major order-of-magnitude improvement on core model capabilities today. Take coding – at the start of the year, models could barely code. By year-end 2025, models like Claude Opus 4.5 and OpenAI's GPT-5.1-Codex-Max - when properly harnessed with tools, evals, and supervision - can now sustain autonomous coding runs of 24 hours or more, shipping production-grade code with minimal human intervention. A further order-of-magnitude improvement from today is very hard to imagine.

Multimodal Progress

We've seen continued advancement in multimodal models – video, speech-to-text, text-to-speech, music / sound and image generation. We are still very early on the development of these capabilities. Video and early research in gaming and world model simulations show huge promise as we go into 2026.

On the R&D front, we have never had so many research and engineering professionals all focused on AI at the same time. NeurIPS 2025 received a record ~21,500 paper submissions. Promising research vectors are emerging across memory, recursive learning, and new model architectures.

AI Achieves Scientific Milestones

In July, AI achieved gold-medal standard on IMO problems (in evaluation) – marking the first time AI approached the level of the world's top teenage mathematicians. The model labs have now turned their attention to the sciences with a number of developments:

What Models Can Do Now

Models today can execute a wide range of tasks they couldn't at the start of 2025:

Even if no new models emerged in 2026, what exists today—properly harnessed with the right products and workflows—can already do extraordinary things. There are months, if not years, of untapped potential in the current technology alone. Contact centres, digital channels, marketing, back-office processes, software development—the applications already are vast.

Enterprise Adoption - Surging But Still Early

In 2024, 78% of organisations reported using AI in some form (up from 55% in 2023), according to the Stanford AI Index 2025. Microsoft's CEO revealed that 20-30% of all code at Microsoft is now written by AI assistants. We're also starting to see credible pilots of end-to-end automation – AI systems completing tasks autonomously rather than merely suggesting or editing human work – though augmentation remains the dominant pattern today. Enterprises are still at the very start of their adoption curve.

Regulation: Moving from Principles to Implementation

Governments have shifted from a light touch to active implementation:

The full risks and benefits are not yet clear to governments and the average consumer, with much of the regulatory response preparatory in nature.

Closing Thoughts

2025 marked the moment AI shifted from impressive demos to a durable, general-purpose capability layer. Intelligence gains saturated the benchmarks. Compute and data constraints emerged as the new bottlenecks. And enterprise adoption—while surging—remains at the very start of its curve.

The gap is widening. Not between those who use AI and those who don't, but between those who redesign how work gets done and those still experimenting at the surface.

We don't need to wait for new models. Years of productivity, creativity, and competitive advantage are already locked up in what exists today. The question heading into 2026 isn't whether AI will matter—it's whether you're positioned to compound alongside it.

In Part 2, I'll share my 10 predictions for where the biggest shifts will land in 2026 —from agentic workflows and job displacement to the IPOs that will define the next chapter.

Sample Sources & References

2025 Developments

Frontier Models & Labs:

Open Source & China:

Compute & Infrastructure:

Data Constraints:

Scientific Breakthroughs:

Robotics & Multimodal:

Adoption & Enterprise:

Regulation & Policy: