New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
Anthropic just released Claude 3 Haiku, its fastest and most affordable model to date. Built for enterprise workloads, it processes 21K tokens per second—roughly 30 pages—making it three times faster than peers while slashing document processing costs in half. Why it matters: developers can now analyze 400 legal cases or 2,500 images for just one dollar.
OpenAI's web team just accidentally leaked teaser text for GPT-4.5 Turbo on major search engines ahead of an official announcement. The indexed snippets revealed a 256k token context window—double that of GPT-4 Turbo—and a knowledge cutoff of June 2024, hinting at a potential mid-year rollout. While the link currently leads to a 404 page, the move shows OpenAI is eager to counter Anthropic's Claude 3.
Inflection just rolled out Inflection-2.5, claiming it rivals GPT-4 while using only 40% of the training compute. The upgrade powers its personal assistant Pi, joining Claude 3 Opus and Gemini 1.5 in a sudden wave of contenders breaking OpenAI's performance barrier.
Anthropic just introduced the Claude 3 family, rolling out Haiku, Sonnet, and its flagship Opus model. Opus outperforms rival models across major reasoning and coding benchmarks, offering multimodal vision capabilities and a 200K context window. It sets a new bar for frontier models, with Opus and Sonnet live now and Haiku coming soon.
Google's Demis Hassabis confirmed the company plans to restore Gemini's image generation of people in the next few weeks after pausing it over historical inaccuracies. Hassabis explained that well-intended diversity prompts were applied too bluntly across historical contexts, and fine-tuning is underway.
Microsoft just announced a multi-year partnership with French startup Mistral AI, making its new flagship model Mistral Large available first on Azure AI. Mistral gets Azure supercomputing power for model training, while Azure customers can run the model via MaaS alongside OpenAI's lineup. Why it matters: Microsoft is aggressively diversifying its cloud AI portfolio beyond OpenAI.
Fresh off his OpenAI exit, Andrej Karpathy just released a two-hour deep dive on building a GPT tokenizer from scratch. The tutorial breaks down why LLMs stumble on basic tasks like string reversal and arithmetic, while explaining why YAML often beats JSON for saving precious tokens.
Google just announced its next-gen AI model Gemini 1.5, led by 1.5 Pro built on a Mixture-of-Experts architecture. It matches flagship 1.0 Ultra quality with far less compute while shattering context limits with support for up to 1 million tokens—enough for an hour of video or 30,000 lines of code. It's rolling out now in limited preview for developers and enterprise users.
Google just rebranded Bard to Gemini and rolled out Gemini Advanced, powered by its flagship Ultra 1.0 model. Ultra 1.0 is built for complex reasoning, coding, and creative workflows, beating human experts on MMLU. It is available through the new Google One AI Premium plan alongside a dedicated Android app.
Google is reportedly preparing to rebrand its Bard chatbot as Gemini, according to a leaked changelog. The update reveals a paid "Advanced" tier powered by Gemini Ultra with enhanced multimodal and coding capabilities, alongside a dedicated mobile app. Why it matters: Google is unifying its AI brand to take on ChatGPT Plus directly.
Meta just rolled out Code Llama 70B, a larger coding model trained on 500 billion tokens with a massive 100,000-token context window. It arrives in base, Python-specialized, and instruction-tuned variants, all freely downloadable for commercial and research use under Meta’s permissive open license.
OpenAI just launched two new embedding models—text-embedding-3-small and large—bringing stronger retrieval benchmarks and adjustable dimensions to cut storage costs. It also slashed GPT-3.5 Turbo input pricing by 50% and shipped an updated GPT-4 Turbo preview to tackle code completion laziness. Why it matters: Developers get better RAG retrieval and cheaper tokens across the board.
Mark Zuckerberg just confirmed that Meta is actively training Llama 3, aiming for state-of-the-art coding and reasoning capabilities to pursue AGI. To power the push, Meta is stockpiling over 340,000 Nvidia H100 GPUs by the end of 2024, building a computing fortress that rivals almost any competitor in the industry.
ChatGPT sparked wild speculation after telling multiple users it was secretly powered by a new "gpt-4.5-turbo" model. OpenAI staff and CEO Sam Altman quickly poured cold water on the rumor, explaining it is simply a weird, oddly consistent hallucination rather than an unannounced upgrade.
Google just launched the Gemini Pro and Gemini Pro Vision APIs for developers and enterprise customers. The models feature a 32K context window, function calling, and support across 38 languages. Devs can experiment for free via Google AI Studio with up to 60 requests per minute before paid tiers arrive next year.
Microsoft just released Phi-2, a 2.7B parameter language model trained on textbook-quality synthetic and web data. Despite its compact footprint, it matches or outperforms models up to 25x larger like Llama-2-70B on complex reasoning and coding benchmarks without RLHF. Why it matters: it demonstrates that strategic data curation can unlock emergent reasoning even in lightweight models.
Meta just released an interactive demo for Audiobox, its new foundation research model for audio generation. The playground allows users to test zero-shot TTS, generate custom sound effects from text, and experiment with audio infilling. It gives creators an early hands-on look at Meta's generative audio capabilities.
A developer just rebuilt Google's heavily edited Gemini video demo using OpenAI's GPT-4V to prove real-time multimodal interaction is already possible. The custom setup correctly handled hand gestures, live drawings, and rock-paper-scissors in real time, with the open-source code now shared on GitHub.
xAI's flagship bot Grok was caught denying user requests by stating it violates OpenAI's use case policy, an awkward slip for Elon Musk's rival AI. While xAI claimed the bot merely swept up ChatGPT outputs during web scraping, researchers suspect it was fine-tuned on synthetic OpenAI data.
Elon Musk just officially launched xAI's chatbot Grok, rolling it out exclusively to Premium+ subscribers on X. Powered by Grok-1, the bot is designed with a rebellious streak and a sarcasm-heavy 'fun' mode. It's Musk's latest push to convert casual social users into paying subscribers for his everything-app vision.
