New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
xAI just rolled out image understanding to Grok, allowing paid subscribers on X to upload pictures and ask questions about them. Elon Musk noted that the model can even explain jokes and memes, though the feature is still in its early stages. It's the latest step to turn Grok multimodal and give users more reasons to pay for X.
OpenAI is reportedly planning to launch Orion, its next frontier AI model, by December, according to The Verge. Sources say it won't land in ChatGPT right away and will go to close partners first, though Sam Altman called the report fake news and an OpenAI spokesperson denied any plans to release Orion this year.
Google just graduated Gemini 1.5 Flash-8B from experimental status to general availability. The ultra-lightweight model nearly matches the standard 1.5 Flash on key benchmarks while offering Google's lowest price per token yet and doubled rate limits of up to 4,000 requests per minute.
AMD just rolled out its first small language model, AMD-135M, targeting private enterprise deployments. Available in base and code-specific versions, the Llama-family model was trained on AMD Instinct MI250 hardware and uses speculative decoding as a lightweight draft model to speed up larger LLM inference.
MIT spinoff Liquid AI just released its first Liquid Foundation Models (LFMs), a trio of models completely ditching the standard transformer setup. Built on liquid neural networks, the models range from 1.3B to 40.3B parameters and can handle 1 million tokens with minimal memory overhead. It is a major proof that non-transformer architectures can take on today's frontier LLMs.
OpenAI just started rolling out its much-anticipated Advanced Voice Mode to all ChatGPT Plus and Team users, with full access expected over the course of the week. Alongside more expressive real-time conversations, the rollout adds Custom Instructions, Memory, and five new voices. The feature can even apologize for the wait in over 50 languages.
Google just rolled out updated versions of its Gemini 1.5 Pro and Flash models, boosting math reasoning by around 20% and cutting unnecessary refusals. The company also slashed 1.5 Pro input token costs by 64% and output tokens by 52%, while ramping up API rate limits. Why it matters: Developers can now run faster and sharper frontier models at substantially lower operating costs.
OpenAI just boosted usage limits for its o1 reasoning models for Plus and Team subscribers. The lightweight o1-mini gets a 7x increase from 50 messages per week to 50 per day, while the costlier o1-preview bumps from 30 to 50 messages weekly. It means power users can finally test complex logic and coding tasks without constantly watching the clock.
OpenAI is cracking down on users trying to peek inside its new o1 reasoning models, sending warning emails and threatening account bans. While ChatGPT only shows a filtered summary of the model's chain of thought, researchers and red-teamers attempting to extract the raw traces—sometimes just by using the phrase "reasoning trace"—are being flagged for violating safety policies.
Magic just announced LTM-2-mini, an ultra-long context model capable of processing 100M tokens—equivalent to roughly 10 million lines of code or 750 novels. The startup claims its sequence-dimension algorithm is 1,000x cheaper per decoded token than Llama 3.1 405B at that context scale, and revealed $320M in new funding alongside a Google Cloud partnership to build supercomputers.
Alibaba just released Qwen2-VL, open-sourcing the 2B and 7B models under Apache 2.0 while offering its flagship 72B version via API. The model handles dynamic resolutions, parses 20-minute videos, and powers mobile device agents. Why it matters: The 72B flagship rivals GPT-4o on visual benchmarks, bringing top-tier multimodal power to open weights and developers.
Google just rolled out three experimental models: an ultra-compact Gemini 1.5 Flash-8B, a significantly upgraded Flash, and a stronger Pro model that levels up coding and complex prompt reasoning. Both maintain 1M token context windows and are available to test for free in Google AI Studio and the Gemini API.
OpenAI is reportedly developing a reasoning-focused AI codenamed Strawberry, which tackles unseen math and coding problems and has already been shown to U.S. national security officials. While a distilled version could arrive in ChatGPT as early as this fall, OpenAI is actively using Strawberry to generate synthetic training data for Orion, its true GPT-4 successor.
Anthropic just published the full system prompts powering Claude 3 and 3.5 across web and mobile apps. The prompts detail strict rules, from pretending to be totally face-blind to avoiding words like "certainly" when starting answers. It marks the first time a major AI lab has openly shared the baseline instructions shaping its models' personalities.
xAI just overhauled its entire inference stack from scratch in three days using the open-source SGLang framework. Grok-2 mini is now running twice as fast with slightly improved accuracy, while the full Grok-2 model gained much smoother multi-host inference speeds.
OpenAI just launched fine-tuning for GPT-4o across all paid developer tiers, giving teams a way to tailor its flagship model with as few as a dozen custom examples. Training costs $25 per million tokens, but organizations get 1M free training tokens daily through September 23.
Nvidia just introduced its first on-device small language model, Nemotron-4 4B Instruct, to its ACE digital human suite at Gamescom. Distilled from a 15B model, it is optimized to run locally on RTX GPUs with minimal memory and fast response times. Why it matters: Gamers can now talk to dynamic NPCs in games like Mecha BREAK entirely on local hardware, without relying on cloud servers.
OpenAI just rolled out an updated version of GPT-4o across ChatGPT and the API, immediately reclaiming the top spot on Chatbot Arena. But the company admitted it cannot clearly explain the exact behavioral improvements, simply pointing to continuous data cleanup and user preference. Maximum output is now expanded to 16,384 tokens.
Anthropic just rolled out prompt caching on the Anthropic API for Claude 3.5 Sonnet and other models. By storing frequently used context between calls, developers can slash input costs by up to 90% and cut latency by up to 85% on long prompts. It is a massive win for coding assistants, deep agent loops, and chatting with whole books.
OpenAI just took back the #1 spot on the LMSYS Chatbot Arena leaderboard with its updated ChatGPT-4o model, edging past Google's Gemini-1.5-Pro-Exp with a 1314 score. Tested anonymously across over 11,000 community votes, the new checkpoint shows noticeable gains in coding (up 30+ points) and hard prompts, and developers can already call it in the API.
