#AI horizons 25-08 – OpenAI busy August

Open-Weight Models, Market Power, and Real-Time Voice AI

Executive Summary

August 2025 was a milestone month for OpenAI, with the headline event being the release of GPT-5—a development significant enough to merit its own dedicated article, which I will cover separately. Alongside that landmark, OpenAI delivered three other major announcements: its first open-weight models since GPT-2, a strategic partnership with the U.S. government to roll out ChatGPT Enterprise at symbolic pricing, and the launch of a real-time voice API capable of humanlike interaction. These moves strengthen OpenAI’s dominance, challenge China’s growing influence in open models, and set the stage for voice-driven enterprise applications that could redefine how businesses and governments interact with AI

Key Points

• OpenAI launched gpt-oss-120B and gpt-oss-20B, its first open-weight models since GPT-2 (2019).
• Both are Apache 2.0 licensed, downloadable on Hugging Face, but training data remains closed.
• The 120B model fits on a single Nvidia H100 GPU, making it efficient relative to Chinese and Meta open-weight competitors.
• Pricing is aggressive: $0.15/M input and $0.69/M output tokens, 10x cheaper than OpenAI’s o4-mini.
• OpenAI reports 700 million weekly ChatGPT users and $1 billion monthly revenue, with a projected $500B valuation.
• Competition increasingly mirrors a duopoly: OpenAI (like Apple’s iPhone) vs. Google Gemini (like Android).
• New GPT Real-time API enables humanlike voice applications, phone call handling, and multimodal conversations.
• The Realtime model costs ~$0.75/hour, potentially disruptive for call centers and enterprise workflows.

In-Depth Analysis

Open-Weight Models: A Strategic Play

OpenAI’s release of gpt-oss-120B and gpt-oss-20B marks its first return to open weights since GPT-2. The move is less about transparency and more about strategic positioning. Alibaba’s Qwen3-235B currently tops global benchmarks, but OpenAI’s smaller, more efficient models are optimized for Western hardware and APIs. Apache 2.0 licensing makes them legally accessible, but withholding training data keeps OpenAI’s moat intact.

By offering lower inference costs and hardware-efficient scaling, OpenAI addresses one of the biggest concerns for European and U.S. enterprises: affordability of AI adoption at scale. This creates a counterweight to Chinese open-weight dominance, which poses both competitive and national security risks.

X: Artificial Analysis.

Monopoly Dynamics and Market Positioning

With 700 million weekly users and $1B in monthly revenue, OpenAI is cementing itself as the consumer AI default. Venture voices like Keith Rabois already call it a “monopoly business.” Comparisons to the smartphone market are apt: ChatGPT as the iPhone, Gemini as Android, and others (xAI, Meta) relegated to niche roles. For Europe, this concentration raises dependency concerns, especially as most EU-native models lag behind in both scale and usability.

The U.S. federal partnership, priced symbolically at $1 per year, is less about revenue and more about ecosystem lock-in. Once agencies standardize on ChatGPT Enterprise, alternatives face immense adoption barriers.

GPT Real-Time: Humanlike Voice as the Next Interface

OpenAI’s GPT Real-time API could redefine digital interaction. Unlike today’s robotic assistants, the model can laugh, sigh, code-switch languages mid-sentence, and sustain context-rich dialogue. With SIP support, businesses can route real phone calls through it, integrating CRM and ERP systems in ways previously impossible without custom engineering.

For enterprises, the economics are striking: ~$0.75 per hour of conversation. That undercuts traditional call centers while adding 24/7 multilingual coverage. For Europe’s service-heavy economies, the labor implications are profound. Voice-first AI is no longer a gimmick but a practical replacement for a segment of human labor.

Business Implications

The combination of open weights, monopoly-scale adoption, and real-time AI creates both opportunity and risk. Enterprises gain cheaper, more flexible options for deploying AI at scale, particularly in customer interaction and automation. However, dependency on U.S. providers deepens, while Chinese competitors like Alibaba continue to push larger open-weight alternatives that may appeal to price-sensitive or geopolitically aligned markets.

For European policymakers, the developments test the AI Act’s relevance: are open weights with withheld training data compliant? Are subsidies for domestic players enough when OpenAI offers 10x cheaper API costs? The risk of strategic lock-in to U.S. infrastructure is clear, even if it brings short-term gains in productivity.

Why It Matters

OpenAI’s August announcements underscore a new phase in the AI race: platform consolidation. By blending affordability (open weights), accessibility (federal partnerships), and usability (real-time voice AI), OpenAI is building structural advantages that few can counter. For businesses, this means faster adoption cycles and new possibilities in customer engagement, but also fewer alternatives in the long run. The EU must accelerate investments in sovereign AI infrastructure if it wants to avoid repeating its dependency on U.S. cloud platforms. For executives, the message is clear: experiment aggressively with real-time AI, capture early productivity gains, but prepare contingency strategies against platform monopolization. The voice-first revolution is not theoretical—it is now priced and packaged for mass adoption.