AI Industry Daily News
A roundup of the AI industry's day, centered on Codex Windows support, grok-build-0.1, Claude Opus 4.8, Command A+, and Rosalind Biodefense.
Today's highlights
Key topics and reactions
Samsung Reports 19-Fold Q2 Profit Rise but Shares Drop, Triggering Sell-off in US Chip Stocks
Samsung's preliminary April-June results showed revenue of 171 trillion won, up 129% year over year, and operating profit of 89.4 trillion won (about $58.4 billion), up from 4.7 trillion won a year earlier. The figure topped analyst estimates by about 6% and marked a third straight record quarter, driven by tight memory demand from the generative AI boom. Citi data showed DRAM average selling prices rose 44% quarter over quarter and NAND 53%.
Despite the beat, Samsung shares fell as much as 10.1% in Seoul and closed down 6.9%, while SK Hynix dropped about 6%. The KOSPI index fell 4.9%, triggering a circuit breaker. Having already risen more than 150% year to date on AI-investment expectations, the stock saw profit-taking on a result seen as strong but short of hopes.
The pullback spread to US markets on July 7, hitting memory and semiconductor names including Nvidia, AMD and Micron. Pre-market moves included Astera Labs down 14% and Intel down 10%, renewing concern over the durability of this year's AI rally.
OpenAI Releases GPT-5.6 Sol in Limited Preview Amid Speculation of Imminent Wider Launch
OpenAI unveiled the GPT-5.6 family on June 26, comprising the flagship Sol, the balanced low-cost Terra, and the fast, low-priced Luna. Sol remains in limited preview, with general availability planned for a few weeks later after a US government AI safety review.
According to the official System Card, Sol scored 88.8% on the TerminalBench 2.1 coding-agent benchmark and 91.9% in "ultra" mode, ahead of Anthropic's Claude Mythos 5. It reported 98% on Network Attack Simulation and claimed comparable performance to Mythos 5 using roughly one-third of the output tokens. New features include a "max" reasoning-effort mode and an "ultra" mode that coordinates sub-agents.
A series of vague posts from OpenAI staff has raised expectations that a GPT-5.6 announcement could come as soon as the following day. The release positions the family against Anthropic's Claude Fable 5 and Mythos 5, with OpenAI emphasizing agentic workflows, coding and cybersecurity, backed by its large-scale partnership with Cerebras announced in January 2026.
Anthropic Extends Free Paid-Plan Access to Claude Fable 5 Until July 12
Announced July 7, the measure grants five additional days of access. An interim allowance offering up to 50% of weekly usage limits for free was originally set to end July 7, with a move to usage credits from July 8. Because the extension coincided with the day weekly limits reset, some users reacted with irony.
Fable 5, positioned by Anthropic as a "Mythos-class" frontier model, was released June 9 for long-horizon agentic work and complex reasoning. After being temporarily restricted under US export controls on June 12, it was redeployed globally on July 1 with strengthened cybersecurity measures following the controls' removal on June 30.
The model is priced at $10 per million input tokens and $50 per million output tokens, with a 1 million-token context window and 128,000-token maximum output. In an Every evaluation for senior engineers it scored 91/100, ahead of Opus 4.8's 63 and GPT-5.5's 62.
Google DeepMind Publishes Gemma 4 Technical Report Detailing Encoder-Free 12B Design
Gemma 4, the latest lightweight open model line built on Gemini technology, natively handles text, image and audio. The lineup includes dense E2B, E4B, 12B and 31B models plus an MoE variant, 26B-A4B, all under Apache 2.0 and available on Hugging Face, Kaggle, Google AI Studio and Ollama. It supports up to 256K context, 140-plus languages, native tool use and a thinking mode.
The report highlights the 12B's encoder-free design. Instead of a separate vision encoder of about 550M parameters, the 12B projects 48x48 RGB patches and 40ms, 16kHz audio chunks directly into the LLM embedding space via a lightweight 35M projection. It is described as an early example of a mid-size open model unifying image and audio, simplifying the pipeline and lowering latency.
Other sizes use frozen ViT-based vision encoders (150M or 550M) and a USM-based audio encoder (305M). Efficiency techniques include p-RoP, which rotates only some dimensions in positional encoding. The report appears on arXiv, and the models were published in April 2026.
Category highlights
Higgsfield Demonstrates Fable 5 + Seedance 2.0 4K Game Workflow
Higgsfield showcased a 4K workflow combining Fable 5 and Seedance 2.0 that can generate an entire game, including story, logic and visuals, from a single prompt. Developers shared workflows such as producing multiple UI drafts at once, simultaneous App Store metadata audits and Japanese localization, and rebuilding a game that took four months in two days. Users cited heavy token consumption and the need for fine-tuning on complex game logic, sharing tips like combining /plan mode with Sonnet to save tokens.
GLM 5.2 on Together AI Ranks #1 for Output Speed and Latency on Artificial Analysis
GLM 5.2 running on Together AI took the top spot on Artificial Analysis for both output speed and latency, with the company highlighting low-level optimizations including ATLAS, Blackwell and TensorRT-LLM. Separately, Anthropic's Claude Sonnet 5 (Thinking) debuted at #6 on the new Agent Arena, and Baidu introduced "Unlimited OCR" featuring memory processing likened to human forgetting.
Cohere, BytePlus, ElevenLabs and OpenAI Ship Audio Updates
Cohere released the Apache 2.0 Arabic speech recognition model Cohere Transcribe Arabic, claiming top accuracy in the language. BytePlus launched Seed Audio 1.0, which generates voice, emotion, sound effects and background music from a single prompt. ElevenLabs formally entered Canada and plans a Toronto office, and OpenAI began offering GPT-Realtime-2.1-mini with reasoning via API.
NVIDIA and Hugging Face Integrate Isaac GR00T 1.7 into LeRobot 0.6
NVIDIA and Hugging Face integrated Isaac GR00T 1.7 and Isaac Teleop into LeRobot 0.6, adding world-model policies, VLA models and simulation benchmarks to strengthen the evaluation loop. NVIDIA also released Isaac Sim 6.0 with Newton physics and MCP integration, while xAI announced a new model focused on understanding and manipulating the physical world.
Anthropic, Databricks, Google and AWS Expand Developer Tooling
Anthropic published the development history of Claude Code, tracing it back to a 2021 internal tool called "clide." Databricks announced Lakeflow integration (Designer, Genie ZeroOps, 100+ connectors) and expanded Genie Code's agent capabilities. Google AI Studio added Android vibe coding, AWS SageMaker Studio enabled one-click deployment of Hugging Face models, and Chrome 150 reached stable.
Key trends
ICML 2026 Opens in Seoul with Over 20,000 In-Person Attendees
ICML 2026 in Seoul drew more than 20,000 in-person attendees, with Microsoft presenting over 100 papers including Fara 1.5. NVIDIA also announced an AI in Production Day for SIGGRAPH 2026.
Vidu, PixVerse and Stability AI Add Video Generation Features
Vidu S1 demonstrated real-time video calls with pets, PixVerse announced a "Team Ultra" plan and a real-time conversational AI live feature, and Stability AI shipped an update improving the realism and speed of its video synthesis tools.