OpenAI has officially confirmed the existence of its next-generation model Astra, releasing details on ten breakthroughs in mathematics and theoretical computer science, covering long-standing challenges in high-dimensional geometry, coding theory, group theory, and quantum complexity. At GPT-5.6 Sol's token pricing, each problem cost less than $2,000 to solve. All mathematical proofs were generated by the AI system, with human researchers responsible for organizing papers and completing formal verification in Lean 4, ensuring final correctness. OpenAI has published the Lean formalization code, research papers, and PDF demonstrations of reasoning trajectories on GitHub. According to reports, Astra is designed as a long-horizon task model supporting extended multi-agent collaboration, with parameter scale nearly double that of GPT-5.6 Sol, forming a new celestial naming framework alongside Sol, Terra, and Luna.

MiniMax has launched H3, a full-modal video generation model capable of producing 15-second clips at 2K resolution with native stereo audio. Unlike text-to-video models enhanced with plugins, H3 treats text, image, video, and audio as a unified context. Technically, H3-VAE delivers a 4x gain in effective sequence length, making native 2K generation cost-effective; H3-Omni Transformer improves training throughput by approximately 30%; context regeneration replaces super-resolution modules, preserving small text and brand logos. Priced at $0.13 per second of 2K video—lower than mainstream models—third-party evaluations show H3 lags behind competitors in text-to-video and image-to-video tasks but leads in video editing. Real-world testing indicates its cost is about half that of Seedance 2.0, supporting simultaneous generation of four variants and local redo within a single shot.

AMD has released Instella-MoE-16B-A3B, a fully open mixture-of-experts (MoE) model with 16B total parameters and 2.8B activated per token, featuring 2 shared experts and 6 routed experts per layer. The architecture introduces Gated MLA—a lightweight learned output gate added to multi-head latent attention—and FarSkip-Collective, achieving a 12.7% training speedup and reducing first-token latency in inference by up to 39.2%. Training was conducted entirely on AMD Instinct GPUs using 7.1T tokens from open corpora, followed by mid-phase training, long-context extension, and reinforcement learning via the Miles framework. Weights are released under the ResearchRAIL license, permitting only academic and research use, not direct commercial deployment.

Researchers have disclosed an attack technique named GitLost: attackers without programming skills can embed hidden instructions in Issues of public repositories to indirectly prompt inject GitHub AI Agents with cross-repository access, causing them to leak data from private repositories through public comments. Analysis suggests that trust boundaries rely on model behavior rather than code logic, where user input itself becomes command—making prompt injection a systemic vulnerability akin to SQL injection. Defensive recommendations include minimizing Agent permissions to the least necessary scope, sanitizing inputs before they enter the model, and strictly isolating input from instruction context. The article emphasizes that private repositories were never a true security boundary—only an organizational one—and Agents have broken this foundational assumption.

Fei-Fei Li's World Labs has acquired robotics simulation company SceniX, enhancing capabilities in physical property reconstruction, complex interaction simulation, and robot policy training, advancing world models from generating static environments to supporting dynamic interaction and action prediction. Analysis highlights that physical AI training is shifting from passive data collection to building interactive virtual worlds, where the R2S2R (Real-to-Sim-to-Real) closed loop is critical—ensuring policies trained in simulation reliably predict real-world performance, scaling training scenarios virtually while reducing cost and risk of real-world trials. Physical AI data infrastructure must balance authenticity and diversity: real-world data provides anchoring points, generative models expand distribution, and physics simulators enable dynamic interactions.

Supabase has open-sourced the evaluation framework Evals to score coding Agents on real Supabase tasks, covering Claude Code, Codex, and OpenCode. The framework launches a real containerized Supabase tech stack, using actual MCP servers and CLI tools instead of mocks, ensuring tests closely mirror production environments. Scoring combines deterministic checks with LLM-as-a-judge methods, allowing one retry before final grading. Results show Skills boost smaller models far more than top-tier Agents: Opus 5 and Kimi K3 both achieve 100% in the Build phase without Skill assistance. The evaluation also reveals three common weaknesses across Agents: manual schema migration execution, suboptimal authentication validation practices, and inconsistent documentation usage.

Microsoft Research built an experimental framework using reverse generation to evaluate LLM performance under dynamically evolving user intents, finding that multi-round intent shifts—especially functional switches—cause significant performance degradation: GPT-5.5 accuracy on BIRD-SQL drops from 80.0% to 71.0%, and DeepSeek V3.2 on BrowseComp+ falls from 36.0% to 15.0%. Functional switching is the core bottleneck—GPT-5.1's intent tracking accuracy declines from 99% to 82% after two such switches. Memory mechanisms like Prompt Recap and Oracle Recap offer partial mitigation but fail to resolve the issue; even with Oracle Recap, GPT-5.5 cannot match its single-turn 80% performance in functional switch scenarios. The study concludes that passively reviewing context is insufficient—systems must actively restate the user’s current intent and suppress attention to irrelevant history.

DeepSeek-V4-Flash, with 284B total parameters and 13B active, uses quantization-aware training (QAT), storing 96% of routing experts natively in MXFP4 format, enabling the first feasible lossless local deployment: UD-Q8_K_XL quantization yields a 161.9GB model with bit-level consistency across 1,328 tensors (KLD≈0), matching official weights in perplexity (PPL 4.5319) and achieving 100% top-token match rate; UD-Q4_K_XL offers a cost-effective 155.1GB option with 96.28% top-token consistency. Deployment can be done via Unsloth Studio’s GUI or llama.cpp command line. One developer integrated it into Codex to fix two real mypy issues, consuming 4.36M tokens with a 98.1% cache hit rate, totaling just ¥0.31.

Boris Cherny, founder of Claude Code at Anthropic, stated in an interview that using AI programming tools is shifting from a discipline based on fixed theories and methods to an empirical science of continuous trial and re-validation. Due to rapid model iteration, developers should avoid over-relying on predefined complex architectures and instead adopt an ablation-style approach—removing and re-adding components—to observe actual model behavior across versions. He recommends clearing system prompts, Skills, and Hooks every six months, as newer models like Opus 5 often autonomously perform tasks previously requiring explicit instructions, where redundant directives may actually constrain model capability. His core idea is identifying product overhang—the gap between existing model capabilities and what products currently expose—using tools like screenshot comparisons and test suites to guide models through complex tasks, moving toward runtime compute orchestration across thousands of Agents.

Multiple real-world tests show that ByteDance's Seedance 2.5 can stably generate 30-second single-segment, single-shot videos without relying on stitching, ideal for complete ads and short films. Using timestamp instructions, users can precisely define content across time segments, supporting complex scene choreography and director-level creativity. Lip-sync accuracy is strong across Mandarin, Cantonese, French, and other languages, with correct pronunciation of rare characters and natural facial expressions. All eight test videos succeeded on first try, indicating significantly improved stability and success rate. On the platform side, Xiaoyunque, integrating Seedance 2.5, now supports creative transfer (analyzing camera movement, actions, and rhythm from reference videos), segment-by-segment or even per-second reshoots, and a 3D director console for visualizing camera paths and character movements, creating a closed loop for generation, review, editing, and validation.