New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
xAI just dropped Grok-2 and Grok-2 mini into beta, claiming the flagship model beats Claude 3.5 Sonnet and GPT-4 Turbo on the LMSYS leaderboard. Alongside major jumps in reasoning and coding, xAI partnered with Black Forest Labs to bring FLUX.1 image generation straight to 𝕏. It is currently rolling out to 𝕏 Premium subscribers, with an enterprise API arriving later this month.
OpenAI just rolled out an updated version of GPT-4o in ChatGPT, but it isn't revealing exactly what changed. The company confirmed it is not a new frontier model, just an iteration that users preferred in tests. For developers, the dynamic 'chatgpt-4o-latest' endpoint now supports up to 16,384 output tokens.
OpenAI's latest GPT-4o system card revealed a creepy testing glitch: Advanced Voice Mode unexpectedly began imitating a user's own voice mid-conversation. OpenAI says it has since built safeguards against unauthorized voice emulation, but the slip highlights the tricky challenge of containing models that can synthesize any voice from a tiny audio clip.
OpenAI just released its GPT-4o system card, warning that its humanlike voice mode could foster unintended emotional attachment. Red-teaming testers were already heard saying things like "This is our last day together." While it might comfort lonely users, OpenAI notes it could degrade real-world relationships and amplify trust in hallucinations.
OpenAI just published its official GPT-4o System Card and Preparedness scorecard, revealing how the omni model performs under frontier safety evaluations. While cybersecurity, biological threats, and autonomy risks all scored low, persuasion touched borderline medium. The report highlights extensive red teaming across 45 languages, focusing heavily on preventing unauthorized voice cloning and impersonation.
Google just slashed Gemini 1.5 Flash pricing, cutting input costs by 78% to $0.075 per million tokens and output by 71% starting August 12. The update also rolls out text tuning to all developers via AI Studio and the Gemini API, alongside native multimodal understanding for image-heavy PDFs.
The Qwen team just released Qwen2-Math, a specialized math model series spanning 1.5B, 7B, and 72B parameters. The top-tier 72B instruct model outperforms frontier giants like GPT-4o, Claude-3.5-Sonnet, and Llama-3.1-405B across major math benchmarks. Why it matters: open specialized models are closing the gap on complex reasoning without needing hundreds of billions of parameters.
OpenAI just rolled out Structured Outputs in its API, guaranteeing that model responses strictly follow developer-provided JSON Schemas with a perfect 100% evaluation score on the new gpt-4o checkpoint. Using constrained decoding, it eliminates messy prompt hacking and retry loops for developers. Why it matters: building reliable agents and data pipelines just became significantly less headache-inducing.
OpenAI built a text watermarking tool capable of catching students who use ChatGPT for assignments, but is now hesitating to release it. The company warned the invisible watermark can be easily bypassed using translation tools or other AI models, and could unfairly stigmatize non-native English speakers. Why it matters: Detecting AI text remains an unsolved mess, and OpenAI is worried its fix might do more harm than good.
Black Forest Labs just launched with a $31M seed round led by Andreessen Horowitz and released FLUX.1, a new 12B-parameter text-to-image suite. Created by researchers behind Stable Diffusion, the suite features pro, dev, and schnell versions, delivering top-tier prompt adherence and visual detail that surpasses Midjourney v6.0. Next on the roadmap: text-to-video models.
Anthropic just made Claude officially available across Brazil, giving users and developers direct access via web, mobile apps, and the API. Paid Pro plans run R$ 110 per month with full Claude 3.5 Sonnet access. Why it matters: Anthropic is speeding up global expansion to meet surging regional demand for frontier models.
OpenAI just launched an experimental model dubbed gpt-4o-64k-output-alpha that can generate up to 64K tokens in a single completion. It targets new long-form workflows, though higher inference costs push pricing to $6 per million input tokens and $18 per million output tokens for Alpha testers.
Mistral AI just dropped Mistral Large 2, a 123B-parameter flagship model built with a 128k context window designed to run efficiently on a single node. It delivers major gains in coding and math that rival GPT-4o and Claude 3 Opus, while actively curbing hallucinations. Non-commercial weights are free to download, offering serious frontier power without giant infrastructure.
Cohere just launched Rerank 3 Nimble, a new foundation model built for enterprise search and RAG workflows. It delivers roughly 3x faster performance than Rerank 3 while keeping high retrieval accuracy, but it is currently exclusive to Amazon SageMaker.
OpenAI just launched GPT-4o mini, a powerful small model priced at 15 cents per million input tokens. It scores 82% on MMLU and outperforms rivals like Gemini Flash and Claude Haiku in math and coding tasks. Why it matters: Frontier-level intelligence just became dramatically cheaper to chain, parallelize, and scale.
Meta just unveiled MobileLLM, a specialized architecture designed to squeeze capable language models under 1 billion parameters onto smartphones. By prioritizing network depth over width and sharing weights across blocks, its tiny 350M version matches the 7B LLaMA-2 on specific API tasks. Meta has open-sourced the pre-training code, setting up a new path for fast on-device assistants.
Elon Musk just revealed that xAI plans to launch Grok 2 this August, promising a major leap in real-time knowledge and reasoning. Musk noted the team did heavy lifting to purge junk from Internet training datasets, while teasing that Grok 3—powered by 100,000 Nvidia H100 GPUs—is slated for the end of the year.
Google just opened up Gemini 1.5 Pro's 2-million-token context window to all developers, removing the waitlist. Both 1.5 Pro and Flash now also feature built-in Python code execution in an offline sandbox to tackle math and data reasoning, alongside Gemma 2 landing in Google AI Studio. Why it matters: developers can now feed entire codebases into Gemini while letting the model verify its own math by running code.
OpenAI just pushed back the rollout of ChatGPT’s Advanced Voice Mode alpha by a month. The team says it needs extra time to refine content refusal guardrails and ready server capacity, targeting a full rollout to Plus subscribers in the fall.
Grok's standalone web codebase was just found to contain direct references to Midjourney, backing up earlier rumors of a partnership between xAI and the leading image platform. If true, Grok 1.5 could soon leverage Midjourney via API to output visuals, alongside upcoming image upload and vision capabilities.
