Big Tech unveils new AI models: OpenAI, Anthropic, Google release updates
The artificial intelligence landscape witnessed a flurry of major advancements this past week, as industry titans OpenAI, Anthropic, and Google each unveiled significant upgrades to their foundational models, pushing the boundaries of AI reasoning, coding, and problem-solving capabilities. These releases underscore the accelerating pace of innovation and the intense competition at the forefront of the AI race.
OpenAI made a notable return to its “open-source principles” with the release of two new “open-weight” reasoning models, gpt-oss-120b and gpt-oss-20b, on August 5, 2025. While not fully open-source as the training data remains proprietary, these models offer publicly accessible parameters, allowing developers to customize and deploy them on their own infrastructure. The larger gpt-oss-120b, with 117 billion parameters, is designed for high-performance tasks and can run efficiently on a single 80 GB GPU, achieving near-parity with OpenAI’s proprietary o4-mini on core reasoning benchmarks. The more compact gpt-oss-20b, at 21 billion parameters, is optimized for edge devices and personal computers with just 16 GB of memory, delivering performance comparable to o3-mini. Both models, released under the Apache 2.0 license, excel in advanced reasoning, coding, competitive mathematics, and health-related queries, further supporting tool use and adjustable reasoning effort. Their Mixture-of-Experts (MoE) architecture contributes to fast and cost-efficient inference, making them versatile tools for research, development, and enterprise applications.
Meanwhile, Anthropic unveiled Claude Opus 4.1 on August 5, 2025, an incremental yet impactful upgrade to its flagship Claude Opus 4 model. This new iteration significantly enhances coding performance, achieving an impressive 74.5% on the SWE-bench Verified benchmark, an increase from Opus 4’s 72.5%. Opus 4.1 also boasts advanced reasoning and agentic capabilities, proving adept at in-depth research, data analysis, and tackling complex, multi-step problems with heightened precision. Its ability to handle long-horizon tasks and synthesize insights from vast datasets positions it as a powerful virtual collaborator for strategic decision-making across various fields. The model is available to paid Claude users and via Anthropic’s API, Amazon Bedrock, and Google Cloud Vertex AI, maintaining the same pricing as its predecessor.
Not to be outdone, Google rolled out Gemini 2.5 Deep Think AI, an advanced reasoning mode for its Gemini 2.5 Ultra model, starting August 1, 2025. Deep Think introduces a groundbreaking “parallel thinking” architecture, allowing Gemini to generate and simultaneously evaluate multiple ideas, much like human brainstorming. This innovative approach provides Gemini with extended “thinking time,” significantly boosting its ability to solve complex problems requiring creativity, strategic planning, iterative development, and advanced coding. The model, a variation of which achieved a gold-medal standard at the 2025 International Mathematical Olympiad, has demonstrated superior performance against rivals like OpenAI’s o3 and xAI’s Grok 4 on key benchmarks such as Humanity’s Last Exam and LiveCodeBench V6. Currently, Gemini 2.5 Deep Think is exclusively available to Google AI Ultra subscribers, with plans for broader API access for trusted testers in the near future.
These simultaneous launches highlight a pivotal moment in AI development, with leading companies pushing the envelope in reasoning, efficiency, and accessibility. As models become more capable and specialized, the focus shifts towards practical deployment and the nuanced balance between open accessibility and proprietary advantage. The continuous evolution of these AI powerhouses promises to reshape industries and redefine human-computer interaction in the years to come.