AI Predictions for 2025: A probabilistic look

Predicting AI developments in 2025 is a challenge, as the most important thing to know is that we really don’t know what’s going to happen. I have therefore decided - as the old data scientist that I am - to reflect my predictions in raw probabilities!

As because of this uncertainity, few will exceed 50% likelihood. These forecasts represent the most probable scenarios in a rapidly evolving landscape. And sometimes they showcase the opposite, a myth that I think is worth calling out how unlikely that is.

Here’s how the predictions are structured: I provide a headline summarising each prediction, a probability, and a brief explanation of the reasoning. For numeric predictions, I use ranges (e.g., [20-30%]), reflecting 90% confidence intervals.

1. A major new frontier model will be released

Prediction: A major new model will be released, similar in purpose to GPT-4, but universally better in most aspects. We will surely (99%) see "better" models, but we're touching 30% probability that all of 2025 will be spend largely moving efforts around. One model might be better at reasoning and score higher on benchmarks but be so slow and expensive it isn't worth it for everyday tasks. Again: Progress is almost certain (99%!) but near-universal progress is not.

Probability: 70%

Reason: Maybe we're hitting a wall, maybe we don't need progress according to the labs. Maybe it's worth doing progress different for a year. Who knows. But the industry should brace for the very real posibility that we won't see a major universal breakthrough in 2025.

2. Massive per-token price drops for equivalent intelligence

Prediction: Token prices for the same level of intelligence will fall dramatically.

Probability: YoY price for the same token will drop by 80-97%

Reason: This is basically continuing history (97%) or possibly with some drop-off (80%). A pretty safe prediction, if I am being honest.

3. Frontier model costs will increase

Prediction: New frontier models will cost a lot more than today’s most advanced models.

Probability: The most expensive tokens will be 10x current top prices.

Reason: I wanted to say 1-10x, but let's be bold and go for 10x. o1 is currently 15/60 USD per million input/output tokens, i.e. I'm predicting a model priced at 150/600 USD.

General note: The three first prediction is what I'd largely consider semi-permanent industry-trends. This is basically continuing the world we've seen in 2023 and 2024, and we should expect to see in 2026 and 2027 too. Fun fact, it logically means in around 2027, we might see models costing as much as 60.000 USD per million output tokens, or. 60 USD for one page worth of content. This is likely an extreme but no unthinkable scenario, if intelligence progress continues.

Also note, that it means current prices of something like GPT-4o will drop to 0,1 USD per million output tokens in 2027. I.e. current intelligence will be virtually free.

4. The AI service boom

Prediction: There will be a 30x increase in companies pitching AI products.

Probability: 80%

Reason: 2025 will be the year of the AI product, with a massive wave of B2B-focused services emerging. 2024 was a lot of obvious stuff and basic internal solutions. The startups are going to hit seriously in 2025 and many large enterprises too. (Bonus prediction: This will continue into 2026 too. The noise of people trying to sell you AI stuff is going to increase massively)

5. AI startups will dominate global VC funding

Prediction: 50-80% of globally funded startups in 2025 will be AI-focused.

Reason: AI will dominate venture capital markets. However, Europe will lag behind, with AI startups accounting for roughly half the global proportion. This is another almost free lunch when looking at Y Combinator batches, but it's worth stating here, because the impact on business life will be pretty profound.

6. No game-changing breakthroughs

Prediction: There will be no sudden revolutionary leaps in AI capabilities.

Probability: 85%

Reason: Progress will remain steady, but no major breakthroughs in planning, reasoning, or long-term task management will emerge. While niche modalities may develop, we won’t see game-changing capabilities like robotics models or transformative advancements in AI’s ability to handle complex tasks. That said, I'm 85% sure... So 15% left for a game-changer is still something I think a lot about.

7. Saturation of known benchmarks

Prediction: All current benchmarks, such as MMLU, will be saturated in 2025.

Probability: 75%

Reason: AI capabilities will continue to grow, effectively solving current benchmarks. While there may still be lingering questions, these benchmarks will no longer challenge or meaningfully differentiate models.

8. Saturation of benchmarks will stunt AI development

Prediction: Saturation of benchmarks will hinder AI development, and some model makers will publicly call for better benchmarks.

Probability: 50%

Reason: Without clear evaluation criteria, the AI landscape will become increasingly confusing for customers, with competing claims about model superiority. This lack of clarity will slow model progress and prompt public calls for improved benchmarking standards. We might even see industry darling Hugging Face get some backlash for their excitement about open source models doing well on common benchmarks then falling massively short in real world use.

9. Private benchmarks will begin to emerge

Prediction: Some businesses will develop private benchmarks to evaluate models for proprietary tasks.

Probability: 20%

Reason: With benchmark saturation, businesses in certain industries will need private metrics to assess models’ suitability for their specific needs. This trend will spark debate on whether to keep benchmarks private (protecting incumbents) or make them public to incentivize innovation.

Note: I have a whole write-up on this coming up, so stay tuned! I believe benchmarking to be a pivotal part of the near-term future of AI, and we'll be talking a lot more about them in 2025.

10. The emergence of second-generation applications

Prediction: Second-generation AI use cases will define much of 2025.

Probability: 75%

Reason: While vague, second-generation applications will likely center on purpose-built systems, including broad categories of "AI agents" that automate workflows, interact with systems, or perform as basic employees. This is going to be hell to evaluate in 2025, but let's see how it looks. If it's still all just RAGs and ChatGPT clones, I'll rest my case.

11. Creative AI goes mainstream

Prediction: Generative AI will become mainstream in social media and content creation.

Probability: 80%

Reason: Social media platforms will increasingly feature AI-generated content, which will largely be accepted—even celebrated—by users.

Sub-prediction: A fully AI-generated song will reach the Spotify Top 10.

Probability: 10%

Reason: While unlikely, this reflects the growing capabilities of generative AI in music and entertainment.

Final Thoughts

2025 would be considered insane progress if it wasn't for the hype and progress of 2024 and 2023. intelligence will increase and prices will fall at rates that could independently drive innovation in our field for years. But it will feel stagnent, because we've gotten used to acceleration.

The main change will be the massive onslaught of AI especially B2B but also B2C services. We've had about 2 years worth of time with this technology. If history is any indication, the first movers should exit development just about now. Get ready for a million sales pitches and get your sourcing and selection team in order.

The good news: While it's going to be terribly stressful for all the CIOs out there, a lot of these products are honestly going to be pretty awesome! There is A LOT of value to be generated in 2025.