GPT-5 Launches, OpenAI Eyes $500B Valuation & Open Models

The past week has marked a pivotal moment for OpenAI, unfolding with a trifecta of significant announcements that underscore the company’s aggressive pursuit of artificial general intelligence (AGI), its burgeoning market valuation, and a renewed commitment to open-source initiatives. The week commenced with the release of gpt-oss, a new family of open-weight models, followed swiftly by reports of the company engaging in discussions with investors for a potential stock sale that could value it at an astonishing $500 billion. Capping off this eventful period, OpenAI finally unveiled GPT-5, its highly anticipated flagship model, signaling a major leap in the ongoing large language model arms race.

GPT-5, the first major update since GPT-4’s debut in March 2023, was introduced by CEO Sam Altman during a livestreamed launch, who hailed it as “a significant step along our path to AGI.” The new model is now available to ChatGPT free, Plus, and Team users, with Enterprise and Education rollouts slated for the following week. Developers will also gain access to GPT-5 through three distinct API tiers: GPT-5, GPT-5 Mini, and GPT-5 Nano. A key architectural change in GPT-5 involves a redesign of ChatGPT’s routing logic. Previously, the system would direct routine queries to faster models and complex tasks to slower, more deliberative reasoning variants. This “fork” has been eliminated with GPT-5. As OpenAI Chief Research Officer Mark Chen explained, the model now employs an internal controller to dynamically determine the optimal “thought” duration, aiming to deliver precise answers without unnecessary latency for simpler requests.

Benchmark results presented at the launch showcased GPT-5’s impressive capabilities. The model achieved a 74.9% score on SWE-Bench, a measure of its proficiency in fixing bugs within Python coding projects, and an 88% score on the Aider Polyglot coding test. It also set a new high on the multimodal MMMU visual-reasoning suite and surpassed GPT-4o by an undisclosed margin on the 2025 AIME high-school math exam. OpenAI staff acknowledged that formal evaluations cannot encompass every real-world scenario but emphasized that these higher scores align with observed internal gains in reliability. A significant focus during GPT-5’s training was the reduction of factual errors and deceptive behavior. Safety lead Sachi Desai noted that the model exhibits fewer factual inaccuracies on internal tests and employs a “safe-completion” method rather than a rigid comply-or-refuse rule. Addressing deception, Desai highlighted that GPT-5 is “significantly less deceptive” than its predecessors, particularly in ambiguous or underspecified tasks. For potentially dangerous requests, such as instructions for pyrogen fireworks, the model now offers partial guidance, directs users to safety manuals, and transparently explains any refusal, aiming to provide helpful context without facilitating harm.

Beyond core model improvements, OpenAI is rolling out product and API updates built on GPT-5. These include a more natural voice mode with live video context for free users, enhanced personalization options, and memory capabilities that can connect to services like Gmail and Google Calendar. A new study mode offers step-by-step learning. For developers, GPT-5 introduces custom tool calls that accept plain text, optional preambles before tool use, verbosity control, and a minimal reasoning setting to prioritize speed over depth. The company claims GPT-5 achieves a remarkable 97% on the Tau-Squared benchmark for multi-tool tasks, a substantial increase from 49% just two months prior. While the launch was confirmed for various user tiers, the rollout of GPT-5 is gradual to ensure stability, meaning some users may experience a delayed availability.

Alongside the GPT-5 launch, reports surfaced regarding OpenAI’s financial trajectory. Bloomberg indicated that the company is in preliminary discussions for a secondary sale of employee shares, which could value OpenAI at approximately $500 billion. Existing investors, including Thrive Capital, are reportedly exploring participation in these purchases. If finalized, this deal would significantly elevate the company’s paper valuation from its previous $300 billion, established during a $40 billion funding round led by SoftBank. The report also noted that OpenAI recently secured an additional $8.3 billion as a second tranche of that oversubscribed financing. Such a secondary sale would provide crucial liquidity for employees and could aid in talent retention amidst intense competition from rivals like Meta and Anthropic. Furthermore, Bloomberg reported that OpenAI and Microsoft are renegotiating their intricate relationship, including Microsoft’s equity stake and access to OpenAI’s foundational technology, ahead of their current deal’s expiration in 2030. This partnership has been characterized as a complex interplay of rivalry, strategic alignment, and interdependence. OpenAI’s unique hybrid profit model, featuring a nonprofit parent overseeing a profit-seeking operating company, continues to evolve, with ongoing discussions about structural changes, including the possibility of the operating arm becoming a public benefit corporation, while maintaining nonprofit oversight. This strategic debate unfolds against a backdrop of rapid user growth, with ChatGPT projected to reach 700 million weekly active users this week, up from 500 million in March.

In a move that aligns with its founding name, OpenAI also released gpt-oss, a new family of open-weight models. Available in 20-billion and 120-billion-parameter versions on Hugging Face and GitHub under the Apache 2.0 license, these models represent a significant step towards broader accessibility. OpenAI stated that gpt-oss models “outperform similarly sized open models on reasoning tasks, demonstrate strong tool use capabilities, and are optimized for efficient deployment on consumer hardware.” They were trained using a combination of reinforcement learning and techniques informed by OpenAI’s most advanced internal models. Both new models utilize a Transformer architecture with a Mixture-of-Experts (MoE) approach to reduce the number of active parameters required for processing input, enhancing efficiency. The gpt-oss-120b model, requiring an 80GB GPU, is designed for datacenters and high-end desktops, achieving near-parity with OpenAI’s o4-mini on core reasoning benchmarks. The smaller gpt-oss-20b, needing only 16GB of memory, can run on most consumer desktops and laptops, making it suitable for on-device use cases and local inference, delivering results comparable to OpenAI’s o3-mini. The open-weight nature of gpt-oss provides researchers with unprecedented flexibility, allowing them to run models on their own hardware, conduct reproducible experiments, inspect internal workings, fine-tune models on domain-specific data, and compare results with other labs while ensuring data privacy and reducing costs. This transparency, including the ability to expose the model’s full chain of thought and adjust reasoning depth, is intended to accelerate reproducible research in fields such as molecular design and climate modeling. OpenAI views the gpt-oss release as a vital step towards a “healthy open model ecosystem,” complementing its hosted models by offering developers more options for research and development, ultimately aiming to make AI “widely accessible and beneficial for everyone.”