OpenAI Rolls Out GPT-4.5 Turbo and GPT-4.5 Instant Updates
OpenAI has released two significant updates to its GPT-4 model family this week, aiming to refine user experience and enhance core capabilities. The first, GPT-4.5 Instant, launched on March 3rd, is described as a ‘vibes update.’ This iteration focuses on subtle improvements to tone, relevance, and conversational flow, directly addressing user feedback to make ChatGPT feel more helpful and less frustrating. Key changes include reducing unnecessary refusals and toning down overly defensive or moralizing preambles, allowing the model to provide direct answers when appropriate. This update is now available to all ChatGPT users and via the API as GPT-4.5-chat-latest.
Just two days later, OpenAI unveiled GPT-4.5, a more substantial upgrade. While the immediate impact on everyday users for tasks like brainstorming or general queries might be marginal, GPT-4.5 brings notable advancements. It offers slight improvements in coding capabilities over its predecessor, GPT-4.5 CodeX, and demonstrates enhanced performance in using computer tools and searching the internet. A major focus for GPT-4.5 is its native computer use capability, allowing it to navigate desktop environments through screenshots and simulated keyboard/mouse actions, a feature previously requiring a separate model. This advancement is attributed to improved visual perception and reasoning capabilities.
The new model also introduces ‘tool search,’ a feature designed to make the use of tools within chat more efficient. Previously, all tool definitions were included in every prompt, increasing costs and slowing responses for systems with many tools. Tool search allows the model to look up tool definitions only when needed, significantly improving speed and token efficiency, particularly beneficial for API users.
Furthermore, GPT-4.5 boasts enhanced web search capabilities, enabling it to synthesize information from multiple sources more effectively, especially for complex ‘needle in a haystack’ queries. For API users, a major upgrade is the introduction of a 1 million token context window, allowing for much larger codebases and conversation histories to be processed simultaneously.
GPT-4.5 is available to ChatGPT Plus, Team, and Enterprise users starting March 5th, replacing the GPT-4.2 Turbo model. While the free and $8/month Plus plans do not include it, paid subscribers can access it via the ‘Turbo 4.5’ and ‘Pro 4.5’ selectors.
OpenAI showcased several impressive demos of GPT-4.5’s capabilities, including creating and editing spreadsheets and documents with improved design, simulating computer actions like starring emails and adding labels in Gmail, and performing bulk data entry. More advanced examples featured the creation of a playable theme park game and an RPG from single prompts, along with a coded Golden Gate Bridge flyover simulation with decent graphics. While these demos are likely cherry-picked, they highlight the model’s potential, particularly for agent-like applications.
Google’s Gemini Enhances NotebookLM and Introduces Flash
Google has also made strides with its Gemini models. The company released Gemini 3.1 Flash, a model optimized for speed and cost-efficiency, making it ideal for API integrations and building applications. The creator of this transcript utilizes Flash for a personal tool called ‘YouTube Thumbnail Swipe File,’ which quickly describes uploaded thumbnails.
More significantly, Google enhanced its NotebookLM platform with a new ‘cinematic video overviews’ feature. This upgrade, powered by Gemini 3 Nano, V3, and potentially other advanced models, generates more dynamic and animated video summaries compared to the previous slideshow-like format. While currently exclusive to the ultra-tier plan ($250/month), this feature integrates motion graphics and animations, offering a glimpse into AI-assisted content creation that could rival tools like Adobe After Effects for quick animations.
Additionally, Google has integrated ‘Canvas in AI mode’ into Search, allowing users to generate content, such as simple HTML and JavaScript games like tic-tac-toe, directly within the search interface, similar to features found in Claude and ChatGPT.
Anthropic-Pentagon Drama and User Exodus from OpenAI
The ongoing saga between Anthropic and the Pentagon has seen significant developments. Following a dispute over Anthropic’s refusal to allow its AI models to be used for US citizen surveillance or fully autonomous weapons, the Pentagon initially designated Anthropic as a ‘supply chain risk,’ effectively blacklisting the company.
In a swift move, OpenAI stepped in to secure a contract with the Pentagon, publicly stating they adhere to similar, and even stricter, red lines, including prohibiting mass domestic surveillance and high-stakes automated decisions. This decision has been met with skepticism, with many questioning whether OpenAI’s public stance aligns with its private agreements with the Pentagon.
The fallout for OpenAI has been substantial. Anthropic capitalized on the situation by launching an easy way for users to switch to Claude, including importing chat history and preferences from other AI providers. Claude also made its memory features available to free users. This has reportedly led to a significant exodus of users from ChatGPT, with reports of a 295% surge in uninstalls and Claude becoming the number one most downloaded app in the App Store.
Internal memos from Anthropic, leaked to The Information, suggest the company believes its designation as a supply chain risk is politically motivated, citing a lack of political donations and support for AI regulation as reasons for the Pentagon’s disfavor, in contrast to OpenAI’s perceived alignment with the administration.
Despite the controversy, Anthropic’s revenue has reportedly surged, nearing a $20 billion run rate, while OpenAI faces a significant drop in business users. This shift indicates a growing preference for Anthropic among businesses, potentially driven by its strong coding models like Opus 4.5 and 4.6, in addition to the ethical considerations surrounding the Pentagon deal.
Source: AI News: Everyone's Leaving ChatGPT! (YouTube)
