Good Morning, Model Builders!
OpenAI's next flagship reportedly still needs alignment work, while Grok Voice 3.0 takes the top speech-to-speech spot.
OpenAI Flagship Still Needs Alignment Work

TL;DR: OpenAI will reportedly ship at least one model this year, but its next flagship still needs alignment work. It uses the same looped transformer technique as Astra to lift math and coding, which makes safety monitoring harder. Microsoft accidentally exposed details of the technique on a public page. Read more →
Grok Voice 3.0 Tops Speech Rankings

TL;DR: Grok Voice 3.0 debuts at number one on the Artificial Analysis Speech to Speech Index with 86.9, ahead of GPT-Live-1 at 83.2 and Gemini 3.8 Live at 82.6. It improves 5.6 points on Grok Voice 2.0 and is available through the SpaceXAI Realtime API at $0.08 per minute. Read more →
Liquid AI Opens 24B Local Model

TL;DR: Liquid AI released LFM2.5-24B-A2B, an open-weight 24B Mixture-of-Experts reasoning model with only 2B active parameters. Built for local agents, it scores 48.09 on LiveClaw and decodes 184 tokens per second on an Apple M5 Max, giving builders a capable model that runs without cloud inference costs. Read more →
Qwen Flash Model Hits 685 Tokens

TL;DR: TokenSpeed results show Qwen3.8-Flash-Next reaching 685 output tokens per second per user on Nvidia GPUs. For teams serving open models, that kind of per-user throughput means faster agent loops and lower serving cost per request, and it shows how much inference headroom community tooling can still unlock. Read more →
Bank of America Patches Mythos Flaws

TL;DR: Bank of America hired 1,000 contractors to patch software flaws exposed by Anthropic's Mythos model. The figure offers a rare look at what defending against frontier AI capabilities actually costs, and signals that security spending is becoming a real line item for enterprises as stronger models reach the market. Read more →
Amazon and Nvidia Step Up Bets

TL;DR: Amazon and Nvidia are reportedly stepping up their AI bets, with a $7 billion deal involving Decart and talks with Reflection AI. The moves show the biggest compute owners using acquisitions, investments and partnerships to lock in emerging model developers as competition for frontier capability intensifies. Read more →
OpenAI Infrastructure Chief Sachin Katti Exits

TL;DR: Sachin Katti is leaving his infrastructure role at OpenAI and will stay on as an advisor, saying he decided to move on to build something new. In his farewell post he said OpenAI secured double-digit gigawatts of compute during his year there, the capacity that underpins its model roadmap. Read more →
Intel Details Xeon Racks and GPUs

TL;DR: Intel unveiled Xeon rack designs and new details on its D-series GPUs aimed at agentic AI workloads. For teams planning inference capacity, the key questions are memory, power, software compatibility and deployment readiness, which will decide whether Intel becomes a credible alternative supplier for model serving. Read more →
Bitdeer Adds 67 MW Blackwell Site

TL;DR: Bitdeer is expanding its AI infrastructure in Malaysia with a new 67 MW facility designed for Nvidia Blackwell systems. No customer deal has been announced yet, so the capacity is still speculative, but it adds to the growing pool of Blackwell-ready power in Asia Pacific for model training and inference. Read more →

Top AI models. Lower costs.
Access GPT, Claude, Gemini, and more through one OpenAI-compatible API—and save up to 70% on AI API costs. Create your DIT account, get your API key, and start building.