News at a Glance
OpenAI released GPT-5 on September 24, 2025, cutting inference costs to one-tenth of GPT-4 while setting new records on math, coding, and scientific reasoning benchmarks. The model uses a mixture-of-experts architecture with adaptive reasoning duration, marking a shift in the large model competition from parameter scal
What Happened
OpenAI unveiled GPT-5 at its DevDay event in San Francisco on September 24, 2025. The model scores 92.3% on MMLU-Pro, up from GPT-4's 84.1%, and 88.7% on GPQA scientific reasoning, up from 78.5%. On SWE-bench Verified, GPT-5 solves 82.4% of software engineering tasks versus GPT-4's 67.0%. GPT-5 employs a mixture-of-experts architecture with 3.5 trillion total parameters but activates only about 35 billion per inference. Its core innovation is an adaptive reasoning duration mechanism that dynamically adjusts compute based on task complexity, responding in 0.3 seconds for simple queries and extending to 5 minutes for complex mathematical proofs. API pricing is set at $1.25
Background
GPT-5's release comes as the large model industry faces dual pressures of compute costs and inference efficiency. In the first half of 2025, OpenAI's annualized revenue surpassed $12 billion, but inference costs continued to rise as a share of operating expenses. Google released Gemini 2.5 Pro in June 2025, approaching GPT-4's performance on multiple benchmarks, while Anthropic's Claude 4 Opus excelled in coding tasks. Industry competition has shifted from model capability to intelligence output per unit of compute. OpenAI reorganized its research teams
Deep Dive
Liu Gong believes GPT-5's release marks the entry of large model competition into a new phase centered on inference efficiency. From a technical perspective, the combination of MoE architecture and adaptive reasoning duration resolves the long-standing cost-capability contradiction. GPT-5's 3.5 trillion parameters with only 35 billion activated reduces per-inference compute consumption by approximately 90% while maintaining near-dense model capability. Liu Gong notes that the adaptive reasoning mechanism, which responds in 0.3 seconds for simple tasks and extends to 5 minutes for complex ones, presents significant engineering challenges. OpenAI implements this through a dedicated routing network and compute budget scheduler, creating technical barriers that competitors cannot easily replicate in the short term. From a commercial perspective, GPT-5's pricing directly disrupts the economic model of AI applications. The input token price drop from $6 to $1.25 means a mid-sized AI application processing 100 million tokens daily sees monthly costs fall from $18 million to $3.75 million, a 79% reduction. Liu Gong predicts this price level will catalyze numerous AI application scenarios previously unviable due to cost, including real-time voice interaction, large-scale document processing, and personalized education. From an industry impact perspective, GPT-5's release will force Google and Anthropic to accelerate their inference efficiency efforts. Liu Gong believes Google's Gemini 2.5 Pro already approaches GPT-4 in capability, but GPT-5's cost advantage may give OpenAI a dominant position in the developer ecosystem. If Anthropic cannot deliver a comparably efficient model by Q1 2026, its enterprise customers may accelerate attrition. Liu Gong also cautions that GPT-5's MoE architecture demands higher inference infrastructure standards, potentially marginalizing smaller cloud providers unable to match OpenAI's specialized chip investments.
Data Points
- GPT-5 vs GPT-4 performance: MMLU-Pro 92.3% vs 84.1%, +8.2 points; GPQA 88.7% vs 78.5%, +10.2 points; SWE-bench Verified 82.4% vs 67.0%, +15.4 points. Cost comparison: input token price from $6 to $1.25, -79%; output token price from $18 to $10, -44%. Architecture comparison: GPT-5 has 3.5T total parameters with 35B activated (1% activation ratio); GPT-4 is a dense model with ~1.8T parameters all activated. Competitor comparison: GPT-5's MMLU-Pro 92.3% exceeds Google Gemini 2.5 Pro's 89.1% and Anthropic Claude 4 Opus's 87.6%.
Impact & Outlook
For the industry, within 3 months, GPT-5's low cost will trigger an AI application development boom, with new AI startup registrations expected to grow over 40% and existing developers' inference costs dropping an average of 60%, pushing more products from prototype to production. Within 1 year, GPT-5's MoE architecture may become the industry standard, forcing Google and Anthropic to release similarly architected models by Q2 2026 or lose developer market share. For ordinary users, within 3 months, ChatGPT Plus subscribers will gain GPT-5 access with ro
Reactions
Google DeepMind VP of Research posted on X that GPT-5's inference efficiency is impressive, but Google will adopt a more aggressive sparsification approach in Gemini 3. Anthropic CEO stated that Claude 4.5 will launch in Q1 2026 with inference costs below GPT-5. Microsoft Azure AI confirmed GPT-5 is fully available, noting record first-day enterprise call volumes. Nvidia shares rose 4.2% after the release, with analysts attributing increased demand for H200 and B200 chips to GPT-5's MoE architecture. Multiple AI developer community leaders noted that GPT
Angles to Explore
Points to consider
- GPT-5's MoE architecture impact on small and mid-sized AI chip makers
- How adaptive reasoning duration changes AI application UX design
- Long-term effects of OpenAI's pricing strategy on AI application business models
Source & Original
This item comes from Hugging Face Blog (published 2026-09-24 14:08:57). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.