Xiaomi MiMo-V2.6-Pro: 1T-A42B Open Weights Model Trained for $3MClaude Opus 5.5 and OpenAI GPT-6 Sol/Luna Launch Same Day, Sparking Price WarClaude Code v2.1.280 Adds Opus 5.5 as Default with 1M ContextOpenAI Introduces GPT-6 Sol and Luna with Different Cost-Capability TradeoffsContext Poisoning: Long-Context LLMs Lose Key Evidence as Irrelevant Text GrowsPentagon Says Overreliance on AI Contributed to Missile Strike on Iran SchoolStanford R&DE Used AI to Race-Swap Students for AdvertisingOpenAI AI Trainers Fired for Using AI to Train the AISitefire Trains Model to Detect AI Web Content from Structure AloneAI·rete·RAG: Rete Rule Engine Decides, RAG Explains WhyPlaceReasoner-Beta: Reasoning-Driven Macro Placement with New BenchmarkParallel Halves Research Time and Cost with GPT-6 AstraGPT-6 Improves Prompt Caching with Higher Hit Rates and Lower Latencyllm 0.36 Released with GPT-6 Sol and Luna Supportllm-anthropic 0.29 Adds Claude Opus 5.5 SupportAI-Written TikTok Scripts Are ObviousGoogle's Oscar-Winning Researcher John Platt on Automating Scientific Discovery with AIGeneralized Multimodal Foundation Model Handles New Modalities and Tasks After DeploymentAdaMem: Adaptive Memory Token Allocation for RAG Soft CompressionTreeSpark: Calibrated Load-Adaptive Draft Trees for Semi-Autoregressive Speculative DecodingXiaomi MiMo-V2.6-Pro: 1T-A42B Open Weights Model Trained for $3M

Xiaomi released MiMo-V2.6-Pro, a new top open-weights model with 1T total parameters and 42B active parameters, trained for just $3 million. It comes from a Chinese frontier lab, signaling intensifying competition in open-source models. Developers can freely obtain the weights for research and applications.

Anthropic released Claude Opus 5.5, and about an hour later OpenAI released GPT-6 Sol and GPT-6 Luna. A day earlier, Grok 4.7 and Xiaomi MiMo v2.6 Flash/Pro also launched. Simon Willison says it will take time to assess these models, but his early impression is fierce competition and a price war. Multiple frontier models launching simultaneously gives developers more choices, but time is needed to compare real-world performance.

Claude Code released v2.1.280, adding Claude Opus 5.5 as the default Opus model with 1M token context, priced at $4/$20 per million tokens with $0.20 per million cache reads. More lists in fullscreen mode now support mouse: the wheel scrolls the /skills list, and skill state options in /plugin are clickable. A new CLAUDE_ environment variable was added.

OpenAI announced GPT-6 Sol and GPT-6 Luna, two new models designed to bring frontier intelligence to everyday work with different balances of capability and cost. The announcement did not disclose specific parameters, pricing, or availability details. Developers should watch for follow-up updates on integration options to choose the right model for their budget and performance needs.

A new paper formalizes “context poisoning”: although large language models can process increasingly long prompts, their ability to locate and use decisive evidence degrades as irrelevant or confusable context is added. The authors model this as extreme-value attention interference, explaining the degradation at the attention level. The work is directly relevant to long-document QA, legal and medical retrieval, and suggests developers should look beyond context window size to how models actually retrieve evidence under noise.

The Pentagon acknowledged that overreliance on AI systems contributed to a missile strike on a school in Iran. The incident has raised concerns about risks in military AI applications. The HN discussion is active with 183 comments. Specific details and accountability are still under investigation.

Stanford Review reports that Stanford’s Residential & Dining Enterprises (R&DE) used AI to race-swap students in advertising materials. The story drew 81 points and 60 comments on Hacker News. The incident raises ethical concerns about AI-generated content in university marketing, including modifying student images without consent and the sensitivity of racial representation.

404 Media reports that people training OpenAI’s AI were fired for using AI to do the training work. The story drew 71 points and 53 comments on Hacker News. It highlights tensions between efficiency pressures and quality control in AI training: workers may use AI tools to speed up annotation or generation, conflicting with requirements for authentic training data.

The team at Sitefire (YC W26) trained a model to detect AI-generated web content using only HTML structural features, without analyzing text semantics. The founders have backgrounds in RL/optimization from Stanford and software engineering from TU Munich. Sitefire helps marketing teams get recommended by AI search engines. For content platforms and SEO practitioners, this structure-based detection offers a new signal for content moderation or quality assessment that does not rely on text-based classifiers.

A developer launched ai·rete·rag, addressing teams that put LLMs in charge of auditable decisions (lending, fraud, clinical triage) then bolt on guardrails. The project runs a pure-Python Rete engine evaluating YAML rules against facts first, then uses RAG to explain the decision. This makes decisions auditable and traceable while retaining natural-language explanations. Suited for developers building AI applications requiring compliance and explainability.

Automated macro placement remains a core challenge in VLSI physical design, where existing methods mostly optimize hand-crafted proxy objectives such as estimated wirelength. Researchers released PlaceReasoner-Beta, a reasoning-driven approach to macro placement, along with a benchmarking suite for evaluating it under realistic layout constraints. Unlike prior proxy-only optimization, the work emphasizes interpretable reasoning and reproducible evaluation. For chip design teams, it offers both a comparable method and a benchmark to test whether reasoning-based placement beats traditional proxy optimization.

Parallel used OpenAI’s GPT-6 Astra to let its agents research and synthesize labor-market data in half the time and at half the cost compared to prior models. GPT-6 Astra demonstrated higher efficiency in agent workflows, enabling Parallel to complete data-intensive research tasks at lower expense. The case shows GPT-6 Astra’s practical cost-benefit advantage in agent scenarios requiring heavy information processing and synthesis.

OpenAI upgraded prompt caching for GPT-6, improving cache hit rates and adding new diagnostics, explicit breakpoints, and controls to reduce latency and costs. Developers can manage cache behavior more precisely and avoid redundant computation. The practical impact is cheaper API calls and faster responses, especially for high-frequency repeated prompts.

Simon Willison released llm-anthropic 0.29, adding support for Anthropic’s Claude Opus 5.5. Users can invoke it via command line: llm -m claude-opus-5.5 “prompt goes here”. The update lets llm tool users easily run inference and testing with Anthropic’s latest flagship model.

Simon Willison quotes @therealcornpop noting that AI-written TikTok and YouTube scripts are super obvious. Tells include the ‘it’s not X, it’s Y’ construction, the rule of three, and a weird broken staccato style. These AI-isms have become signals audiences use to spot AI-generated content. Creators who want scripts to avoid detection need to steer clear of these formulaic patterns.

Google researcher John Platt, an Oscar technical achievement winner for Spider-Man 2, now focuses on AI for Science. In an interview, he discussed using machine learning to automate scientific discovery, tackle climate change, and how future generations can contribute to science in the age of superintelligent AI. Platt believes AI can accelerate materials discovery and climate modeling, but humans must set research direction. He noted that future scientists will shift to asking questions and validating AI findings rather than manual computation.

Most existing multimodal fusion models handle only predefined modalities (vision, text, audio) and a single task once deployed, making adaptation slow. A new paper proposes a “generalized multimodal foundation model” that aims to support new modalities and multiple tasks after deployment. If matured, this could cut the cost of retraining for each new modality or task, benefiting teams that need to quickly integrate new data types. The work remains at the research stage with no open-source release or productization announced.

A new arXiv paper introduces AdaMem for soft compression in retrieval-augmented generation (RAG). Processing many long passages is costly and can introduce distracting information. AdaMem encodes passages as compact sequences and adaptively allocates memory tokens to better balance compression rate and information retention. It aims to reduce RAG inference overhead while minimizing irrelevant content, offering value for RAG applications handling long documents.

A new arXiv paper introduces TreeSpark for semi-autoregressive speculative decoding. Existing block drafters make drafting nearly free by emitting an entire block of draft tokens in a single backbone pass. TreeSpark adds calibrated, load-adaptive draft trees to improve verification efficiency and inference speedup. The method aims to reduce language model inference latency, with potential value for high-throughput LLM deployment scenarios.