AI in 2025: The New Competitive Stage

SUAD SEFERI

Artificial Intelligence in 2025 has accelerated from a promising tool into a critical societal infrastructure, shaping economics, governance, research, and culture. The last two months, in particular, have seen developments that redefine the capabilities of large language models (LLMs), multimodal AI systems, and autonomous agents. These are no longer isolated breakthroughs — they signal a structural shift in how AI is designed, deployed, and integrated into everyday life. The question for leaders and policymakers is no longer whether to engage with AI, but how quickly and responsibly they can adapt.

A Surge in Capability

The headline event is undoubtedly OpenAI’s release of GPT‑5 in August 2025 — widely regarded as the most capable AI model made available to the public so far. With over 1.8 trillion parameters and a one‑million‑token context window, GPT‑5 merges unprecedented scale with drastically reduced “hallucinations” (false or fabricated outputs). For industry professionals, this is transformative:

Lawyers can analyse and cross‑reference thousands of pages of legal text in real time.
Scientists can run multi‑variable literature reviews in hours, not weeks.
Creative teams can generate scripts, visuals, and supporting research in a seamless workflow.

Crucially, GPT‑5’s alignment training has been redesigned with advanced safety layers that filter bias, improve truthfulness, and ensure closer adherence to user instructions.

In parallel, Anthropic’s Claude 4 Opus has consolidated its role in the “trust‑critical” AI market. Built on the Constitutional AI framework — embedding ethical principles deep in the training — Claude’s 200,000‑token memory enables sustained, complex conversations and document review without losing context. This makes it indispensable for regulated sectors such as healthcare, compliance, and public administration. The latest update, Claude Opus 4.1, adds advanced autonomous reasoning capabilities for policy simulation and complex coding.

Meanwhile, Google DeepMind’s Gemini 2.5 Pro excels through real‑time multimodal intelligence — processing text, images, audio, and video simultaneously. With one million tokens of memory and deep integration into Google’s ecosystem, Gemini powers industrial simulations, large‑scale media creation, and live data‑driven decision support. In recent weeks, DeepMind unveiled natural voice interaction, making machine conversation sound and feel closer to human dialogue.

Finally, the open‑source frontier took a bold leap. Meta’s Llama 4 Scout shattered records with a 10‑million‑token context window, enabling single‑session analysis of entire legal archives or scientific repositories. Alongside, models like DeepCogito v2 have delivered strong reasoning and code‑generation capabilities at a fraction of the cost of closed platforms — signalling that top‑tier AI power is no longer reserved for the world’s largest tech companies.

Three Defining Shifts (Aug – Oct 2025)

The Leap to Million‑Token Multimodality
LLMs capable of integrating vast textual data with images, video, and audio are now context‑aware reasoning engines, not just text generators.
The Rise of Ultra‑Long‑Memory Models
From Claude’s 200K tokens to Llama’s 10M, retaining relevant context over hours or days transforms legal research, academic analysis, and policy planning.
The Quiet Domination of AI Agents
Autonomous AI agents — able to plan and execute multi‑step tasks without micromanagement — are now in ~70% of Fortune 500 companies, driving efficiency in research, customer service, and operations.

By the Numbers – AI 2025 Snapshot

Global AI Market Growth (2025): +37% YoY (Gartner)
Autonomous Agent Adoption (Large Enterprises): ~70%
Reduction in GPT‑5’s Hallucination Rate vs GPT‑4.5: ~40%
Cost Drop in Mid‑Tier LLM API Calls (Since Jan 2025): ~30%
Energy Efficiency Gains in New AI Chips: +22% (FLOPS/Watt)

Regional Relevance: The Western Balkans in the AI Era

For North Macedonia, Albania, Serbia, Montenegro, Bosnia and Herzegovina, Kosovo, and the wider Balkan region, global AI progress represents a rare opportunity to leapfrog into the next stage of the digital economy.

Across the region, governments are moving toward formal national AI strategies:

North Macedonia is embedding AI literacy in schools, incentivising AI startups through grants and tax relief, and aligning governance with the EU AI Act.
Albania has launched a roadmap to link AI adoption with economic diversification.
Serbia invests heavily in AI hubs in Belgrade and Novi Sad to attract top talent.
Montenegro pilots AI tools in e‑governance and tourism analytics.
Bosnia and Herzegovina tests AI in agriculture and energy efficiency.
Kosovo reforms education policy to prioritise digital skills and integrate into European tech value chains.

Local industry is equally poised for transformation. Agritech firms use AI to optimise yields; fintech startups deploy AI for fraud detection; manufacturing adapts AI for predictive maintenance. Crucially, affordable open‑source models like Llama 4 Scout and DeepCogito v2 allow SMEs to compete internationally without prohibitive infrastructure costs.

Yet there’s a shared challenge: talent. The region needs engineers, data scientists, AI ethicists, and policy specialists capable of guiding responsible adoption. Without them, Western Balkan nations risk importing AI solutions without shaping them — and losing strategic advantage.

A coordinated cross‑border AI ecosystem — through research networks, training programmes, and shared standards — would position the Western Balkans as an emerging AI hub in Europe, not just a consumer of global technology.

Leading Large Language Models – Mid‑2025

Model	Context	Multimodal	Key Strengths
GPT‑5	1M tokens	Yes	Superior reasoning, low error rate
Claude 4 Opus	200K	Limited	Ethical AI, long-context memory
Gemini 2.5	1M	Full	Real-time multimodal processing
Llama 4 Scout	10M	Yes	Ultra-long document analysis
DeepCogito v2	250K	No	Logic & code generation
GPT‑4.5	128K	Basic	Creative tasks, education, content
Grok 3	1M	Partial	Web-enhanced live reasoning

Looking Ahead

The evolution of AI in 2025 proves that sheer scale is no longer enough. Leadership in this field now depends on:

Integrating multiple modalities
Maintaining deep contextual memory
Building trustworthy reasoning frameworks

For AI NOW members, policymakers, educators, and leaders in our region, the path forward is clear:

Adopt tools that genuinely improve output and decision-making,
Regulate with precision to safeguard rights,
Skill‑build across every sector,
Collaborate internationally to remain competitive.

This decade will define whether AI becomes a platform for equitable progress — or a new axis of inequality.

Follow us