# The Latest AI Models and Model Rumors Source: https://www.youtube.com/watch?v=-plNBWEQhHU Recap page: https://rapidrecap.app/video/-plNBWEQhHU Generated: 2026-02-17T14:02:50.531+00:00 --- ## Quick Overview OpenAI announced GPT-5.3-Codex-Spark, an ultra-fast model for real-time coding achieving over 1000 tokens per second inference on low-latency hardware in partnership with Cerebras, while Google DeepMind released Gemini Deep Think with an upgraded version achieving gold medal performance on the International Mathematical Olympiad (IMO) ProofBench Advanced test, scoring 90% accuracy. **Key Points:** - OpenAI released GPT-5.3-Codex-Spark, a smaller, ultra-fast version of GPT-5.3-Codex optimized for real-time coding, delivering over 1000 tokens per second. - Codex-Spark was developed in partnership with Cerebras, utilizing their low-latency hardware for near-instantaneous coding responses. - The new coding model performs significantly faster (roughly 15 times faster) than the previous GPT-5.3 Codex model. - Google DeepMind upgraded Gemini Deep Think, achieving gold medal performance on the IMO ProofBench Advanced test with 90% accuracy. - The upgraded Deep Think model, Aletheia, is designed to autonomously generate, verify, and revise pure math proofs. - Anthropic announced a $30 billion Series G funding round at a $380 billion post-money valuation, and also donated $20 million to Public First Action to support AI policy education. - Anthropic's growth is strong, with 1 in 5 businesses using their models on Ramp now, and their annual recurring revenue (ARR) grew 7x in the past year to $1.4 billion. ![Screenshot at 00:00: The announcement page for OpenAI's GPT-5.3-Codex-Spark, highlighting its nature as an 'ultra-fast model for real-time coding in Codex'.](https://ss.rapidrecap.app/screens/-plNBWEQhHU/00-00-00.jpg) **Context:** The video covers recent major announcements from leading AI labs: OpenAI, Google DeepMind, and Anthropic, focusing heavily on model performance improvements, especially in specialized tasks like coding and mathematics, alongside significant corporate funding news. OpenAI is pushing for faster coding inference with Codex-Spark, Google is advancing mathematical reasoning with Gemini Deep Think's Aletheia agent, and Anthropic continues rapid growth supported by massive new funding and policy advocacy. ## Detailed Analysis The video reports on several key AI developments. OpenAI introduced GPT-5.3-Codex-Spark, a specialized, smaller version of GPT-5.3-Codex designed for real-time coding, achieving inference speeds of over 1000 tokens per second on low-latency hardware developed with Cerebras, making it about 15 times faster than the regular GPT-5.3 Codex. It excels at coding tasks but cannot handle multimodal inputs or long reasoning tasks. Simultaneously, Google DeepMind announced Gemini Deep Think's upgrade, featuring the Aletheia math research agent, which achieved gold medal performance on the IMO ProofBench Advanced test with 90% accuracy, demonstrating superior reasoning quality compared to its previous public release. This agent iteratively generates, verifies, and revises math proofs. Anthropic also made headlines by raising $30 billion in Series G funding at a $380 billion post-money valuation, with significant investment from Silicon Valley VCs and sovereign wealth funds. Furthermore, Anthropic donated $20 million to Public First Action to support AI policy education, aiming to keep US in the AI race. Data shows Anthropic's growth is strong, with 1 in 5 businesses using their models, and their ARR hit $1.4 billion, growing 7x in the past year, indicating strong enterprise adoption despite competition from OpenAI. ### OpenAI Codex-Spark - Ultra-fast model for real-time coding - Delivers over 1000 tokens/second on Cerebras hardware - 15x faster than GPT-5.3 Codex - Limited to text-only, 128k context window - Excels at coding edits, logic revisions, and frontend iteration ### Google DeepMind Gemini Deep Think - Upgraded version, Aletheia, shows gold medal performance on IMO ProofBench Advanced test (90% accuracy) - Aletheia autonomously generates, verifies, and revises math proofs in a loop - Outperforms previous best model (Opus 4.6) on specific benchmarks ### Anthropic Funding & Growth - Raised $30B in Series G at $380B valuation from VCs and sovereign wealth funds - Donated $20M to Public First Action to support AI policy education - ARR reached $1.4B, growing 7x in the past year; 1 in 5 businesses on Ramp pay for Anthropic models ### Claude Cowork Updates - Claude is now available on Windows with full feature parity with macOS, including file access, multi-step task execution, plugins, and MCP connectors - Microsoft is reportedly rushing to develop rival features for Copilot in response to Claude's feature parity. ![Screenshot at 00:00: The announcement page for OpenAI's GPT-5.3-Codex-Spark, highlighting its nature as an 'ultra-fast model for real-time coding in Codex'.](https://ss.rapidrecap.app/screens/-plNBWEQhHU/00-00-00.jpg) ![Screenshot at 00:53: A chart from the OpenAI announcement showing Codex-Spark's superior accuracy compared to other Codex versions on the SWE-Bench Pro benchmark.](https://ss.rapidrecap.app/screens/-plNBWEQhHU/00-00-53.jpg) ![Screenshot at 02:21: The availability and details page for Codex-Spark, noting it is text-only and available via API for a small set of design partners initially.](https://ss.rapidrecap.app/screens/-plNBWEQhHU/00-02-21.jpg) ![Screenshot at 03:22: A diagram illustrating the iterative process of the Aletheia math research agent powered by Gemini Deep Think, showing the Problem -\> Generator -\> Candidate Solution -\> Verifier -\> Final Output loop with a Reviser step.](https://ss.rapidrecap.app/screens/-plNBWEQhHU/00-03-22.jpg) ![Screenshot at 05:51: Anthropic's announcement graphic showing revenue growth, with the final bar \(Today\) significantly larger than previous years, representing $1.4B run-rate revenue.](https://ss.rapidrecap.app/screens/-plNBWEQhHU/00-05-51.jpg)