Astra is now live in Codex and ChatGPT Work for paying users; OpenAI urged builders to ship demos on the new model ahead of next week's San Francisco event.
Largest equity round ever by a European tech company, led by Samsung Electronics with EQT Scaleup Europe Fund and PSG Equity; backers include a16z, Nvidia, Salesforce Ventures, Advent, BlackRock and Luxembourg.
$2B raised at nearly double the $26B mark of four months earlier, led by Andreessen Horowitz with Accel, Founders Fund, General Catalyst and Avenir; annualized run-rate revenue $900M, projected $4-5B by year-end.
Apache 2.0 weights live on Hugging Face and ModelScope; Mini (35B multimodal), Pro (397B computer-use leader, 56.4 OSWorld-2) and Max (1.6T MoE, 50.2 on AutomationBench v1.0.6, 0.1 behind Claude Opus 5).
Follow-up to the Nex-N2.5 family drop: Pro lands on OpenRouter with a free tier on September 9, extending the open-weights plus hosted-access dual strategy.
The viral Ox Alpha model is Zai's GLM-5.3-Flash, a 320B-parameter hybrid linear-plus-sparse MoE (18B active), topping GDPval AA v2 at roughly $0.09 per task and trained on China-made hardware.
Hermes Agent is now a first-class option inside Omarchy, the Arch-based agentic Linux distribution from DHH, reinforcing Nous Research's pivot from base-LLM fine-tunes toward an agent platform that runs third-party open-weights models locally.
MoE architecture grows multimodal embedding capacity without proportional compute, achieving strong retrieval across vision-language and image-text benchmarks while keeping inference efficient.
Rectified DPO formulation for safety alignment that addresses over-refusal and under-refusal trade-offs in open-weight models, claiming safety comparable to proprietary systems while preserving helpfulness.
Extends RLVR for large reasoning models with a tree-structured policy that adapts difficulty per branch, showing measurable gains in pass@k and single-sample accuracy over GRPO on math and code benchmarks.
Argues enriching environment-side feedback (rather than only warming up the agent) mitigates reward sparsity in long-horizon RL training, with ablations on agentic benchmarks showing faster convergence.
Two new API models — GPT-Image-2.5 Flare (faster, default) and GPT-Image-2.5 Sunburst (premium precision) — roll out across ChatGPT, ChatGPT Work, Codex and the API with up to 50% lower latency than Images 2.0; Sketch drawing, templates and comment-on-image features land in ChatGPT.
1-petabyte catalogue of precomputed AlphaGenome predictions for every possible single-nucleotide variant in the human genome, plus the new AlphaGenome Variant Impact (AVI) score; free for academic research via a portal, API and Google Antigravity.
Muse ships with a personal file system, terminal and browser, powered by Muse Spark 1.3; $20 and $100 paid tiers announced alongside a free tier — the launch comes days after Meta's $18B multistate settlement over consumer harms.
Long-horizon benchmark that scores whether AI coding agents can build a working customer-service agent; even the strongest automated configuration scored under 24% on the paper.
Anthropic now lets business customers keep Claude traffic on their own servers for 30 days while OpenAI pitches Zero Data Retention via a new safety processing system; both flag automated misuse detection with limited public technical detail.
Firmus announced OpenAI as anchor customer for two Malaysia AI Factory sites, pushing the Australian operator past 900 MW contracted capacity and bringing Nvidia Vera Rubin NVL72 rack-scale systems to Southeast Asia.
Licensable datacenter compute subsystem with up to 128 cores per die, LPDDR6 memory and PCIe Gen 7, alongside an IDC-cited claim that Arm rack-scale servers now lead x86 in accelerated computing.
Joint initiative to migrate High-NA EUV from current 6-inch masks to 12-inch masks, claiming a path to higher fab productivity and lower chipmaking costs; Samsung plans High-NA EUV adoption from 2028, TSMC from 2030.
Intel says it has processed more than one million 300-mm wafers on High-NA EUV scanners, more than the rest of the industry combined, and is pioneering 6x12-inch photomasks to remove the stitching constraint on next-generation chips.
US startup began invite-only humanoid cleaning service using modified Unitree G1 units teleoperated from a control room; the company reports more than 50 homes serviced in roughly a month.
Italian legal AI startup Lexroom acquired 100% of France's Query Juriste and Bulgaria's Praven Intelekt, expanding its data-first legal AI platform to five European markets; financial terms undisclosed.
Italian VC firm exceeded its initial target; the fund will back Italian and European pre-seed and seed startups applying AI to B2B software and aims to raise €100 million by 2027.
Founder-focused summit opens September 9 with more than 1,000 expected participants from 21 countries, connecting startups, investors, companies and researchers around AI, deep tech, venture capital, sustainability and technology transfer.
Italy's eCampus University kicked off the NeuroLearning Interface project, combining AI, neuroscience and educational technology to develop adaptive and personalized learning models.
Series C led by Bain Capital Ventures with Thrive Capital, General Catalyst and Accel; AI agents handle operational workflow between prescription and treatment for nine of the top 15 biopharma; total raised now >$300M.
Co-led by Fin Capital and SMBC Fin Atlas Beyond Fund with Hannah Grey VC, National Bank of Canada, Y Combinator and Penny Jar Capital; enterprise AI compliance platform for legal and brand teams reviewing AI-generated content, with SoFi and Expedia Group as customers.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
