New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
Anthropic just shared a detailed breakdown of how its Safeguards team keeps Claude reliable across its entire lifecycle. The defense relies on a unified harm framework across five risk dimensions and pairs pre-launch stress tests with real-time fine-tuned classifiers that steer risky responses on the fly. Why it matters: Anthropic is showing how frontier labs can mitigate malicious use without blunt, frustrating model refusals.
Liquid AI just released LFM2-VL, a lightweight open-weight vision-language model family sized at 450M and 1.6B parameters. Built for resource-constrained setups like phones and wearables, it delivers up to 2x faster GPU inference than existing alternatives. The models are available on Hugging Face and free for commercial use for companies under $10M revenue.
OpenAI just significantly increased reasoning rate limits for ChatGPT Plus subscribers, noting that caps across all model classes will soon exceed pre-GPT-5 levels. The team is also rolling out a UI update to clearly show which model is answering each query. The adjustments follow a bumpy GPT-5 rollout plagued by model autoswitcher issues.
xAI just made its latest Grok 4 model freely accessible to all users worldwide, escalating the battle against OpenAI's GPT-5. The rollout introduces Auto and Expert modes with generous initial usage limits, though Grok 4 Heavy stays paywalled. Musk also confirmed that ads will soon arrive in the chat interface to offset mounting GPU costs.
OpenAI just restored GPT-4o as an option for paid ChatGPT Plus subscribers, only a day after replacing it with GPT-5. The abrupt switch and removal of the model picker sparked widespread complaints from users who missed 4o's warmer tone or had built specific workflows around it. CEO Sam Altman said the team will monitor usage to decide how long legacy models stick around.
OpenAI just launched GPT-5 on its API platform, designed specifically for heavy coding and autonomous agent tasks. It sets fresh benchmark records including 74.9% on SWE-bench Verified while slashing tool errors, introduces plaintext tool calling, and arrives in three distinct sizes to balance cost and speed.
OpenAI just released GPT-5, combining fast responses and deep reasoning through a real-time router that cuts hallucinations by up to 80%. It is rolling out to all ChatGPT users, with extended reasoning reserved for Pro subscribers.
OpenAI just teased a Thursday livestream with a not-so-subtle hint: replacing the 's' in livestream with a '5'. CEO Sam Altman had already posted a screenshot showing 'ChatGPT 5', and OpenAI's applied research lead teased its public debut. Why it matters: After months of anticipation, the next-gen flagship model looks finally ready to drop.
Google's Kaggle and DeepMind just kicked off their inaugural AI chess tournament, where general LLMs square off without dedicated chess engines. Grok 4, Gemini 2.5 Pro, o4-mini, and o3 all swept their matches 4-0 to reach the semifinals, with Elon Musk claiming xAI spent almost zero effort on chess. Why it matters: Unlike traditional engines, LLMs still hallucinate illegal moves once off-book, making board tracking a telling proxy for spatial reasoning.
Anthropic just released Claude Opus 4.1, a mid-cycle refresh focused on real-world coding, reasoning, and agentic workflows. It scored 74.5% on SWE-bench Verified without extended thinking, matching the leap seen between Sonnet 3.7 and Sonnet 4. Best of all, it keeps the same pricing and is already live for paid users, Claude Code, and major cloud APIs.
Google DeepMind just unveiled Genie 3, a general-purpose world model that generates interactive environments in real time at 24 fps and 720p from simple text prompts. It keeps scenes consistent for minutes and lets users prompt in-world events like weather shifts on the fly. It is currently rolling out as a limited research preview for select academics and creators.
Google's Kaggle just launched Game Arena, a benchmarking platform designed to test model reasoning through strategic gameplay. From Aug. 5-7, eight flagship models—including OpenAI's o3, Gemini 2.5 Pro, and DeepSeek-R1—will compete in a live chess tournament with zero access to external chess engines like Stockfish. Why it matters: Unlike easily saturated static benchmarks, game theory provides a dynamic test of planning, memory, and logic under incomplete prompt conditions.
Google just launched Deep Think in the Gemini app for Google AI Ultra subscribers, introducing parallel thinking techniques that let the model explore multiple hypotheses simultaneously. Built to tackle complex coding and research problems, it hits Bronze-medal performance on internal 2025 IMO benchmarks while remaining fast enough for daily use. Subscribers get access via a toggle, though it comes with a fixed daily prompt limit.
Z.ai just introduced its new flagship MoE models, GLM-4.5 and GLM-4.5-Air, prioritizing depth over width to enhance reasoning and autonomous agent capabilities. GLM-4.5 features 355B total parameters with only 32B active, launching with day-one support on SGLang alongside FP8 variants. Why it matters: it hits a 90.6% tool-calling success rate and outpaces Claude 4 Opus on web browsing, making it a formidable foundation for autonomous workflows.
OpenAI is preparing to roll out GPT-5 as early as next month, packing o3 reasoning capabilities directly into the main model to eliminate the friction of switching tools in ChatGPT. It reportedly plans to launch combined, mini, and nano versions, while also racing to release a new open-weight reasoning model before July wraps up.
Google just released the stable, generally available version of Gemini 2.5 Flash-Lite, its fastest and lowest-cost model in the 2.5 family. Priced at $0.10 per million input tokens with a 1M context window, it cuts audio pricing and brings native controllable reasoning. It is essentially built for developers who need rock-bottom latency and costs without giving up output quality.
Alibaba's Qwen team just scrapped hybrid thinking mode to train Instruct and Thinking models separately for peak quality. Along with the pivot, it rolled out Qwen3-235B-A22B-Instruct-2507 and an FP8 version, promising stronger general capabilities and better performance on agent tasks.
Google just introduced conversational image segmentation for Gemini 2.5. Instead of relying on single nouns, the model parses relational instructions like "the car farthest away" or abstract ideas like weather damage to return pixel-level masks via a single API call.
Google DeepMind just announced that an advanced version of Gemini with Deep Think achieved gold-medal performance at IMO 2025. Unlike last year's systems that took days and required translation into formal languages, Gemini worked end-to-end in natural language to solve 5 of 6 problems within the standard 4.5-hour limit. It will soon roll out to Google AI Ultra subscribers.
Hume AI just launched EVI 3, a new speech-to-speech model alongside updated TTS capabilities. Instead of stopping at voice cloning, the model captures speaking style and unique language patterns—what Hume calls "personality cloning." Why it matters: voice assistants are moving from just sounding like you to actually talking like you.
