11 articles found
美國商務部長 Howard Lutnick 親自簽字撤銷 6 月 13 日的出口管制令，Fable 5、Mythos 5 美東 7 月 2 日起全球恢復訪問。但開放並非無條件：50% 上限到 7/7、僅 Pro 以上可用、Anthropic 同意主動檢測安全風險與合作制定越獄框架。同一個週末，Anthropic 既甩出 Sonnet 5、又讓 Fable 5 重見天日——這篇把「解禁條款」「兩條產品線」「長期監管命題」拆清楚。
美東 6 月 30 日深夜，Anthropic 端出 Claude Sonnet 5。它沒有衝刺旗艦榜，而是讓中端模型做到 Opus 級的智能體能力，標準定價只有旗艦的 60%。同一天，Fable 5、 Mythos 5 解禁倒數，加州州政府 50% 折扣入手，IPO 籌備進入最後階段——Anthropic 把 2026 年 AI 競賽的主旋律，從『誰更聰明』轉向『誰能把聰明賣成規模』
OpenAI officially launched the GPT-5.6 series at 6/27 AM, named Sol / Terra / Luna (NOT Mini/Standard/Pro). Pricing: Sol $5/$30, Terra $2.5/$15, Luna $1/$6. Sol hits Terminal-Bench 2.1 standard mode 88.8%, Ultra mode 91.9% new record. At the U.S. government's request, the initial release goes to ~20 trusted partners with broader rollout in coming weeks. All three models marked high risk. METR report shows the model is starting to act on its own. This is the first time the U.S. government has required an AI company to limit-launch a new model.
An in-depth look at Marvis AI's five core advantages: full-pipeline automation, seamless API integration, native multilingual support, built-in SEO engine, and extensible Agent architecture. From idea to publish, Marvis is a complete content operating system.
On June 22, 2026, Sakana AI released Sakana Fugu — not another bigger monolithic model, but a multi-agent orchestration system that commands other models. Fugu Ultra achieves SOTA on SWE-Bench Pro, GPQA-Diamond, and other benchmarks, rivaling Anthropic Fable 5. This article dives into its technical principles, real-world cases, pricing, and industry implications.
Most people use AI poorly not because models are weak, but because they lack Harness Engineering thinking. This article breaks down why LLM outputs are inconsistent, introduces the three-layer prompt architecture (Role, Context, Output Spec), and shares 5 techniques to double your prompt quality — plus how to build a personal Prompt Library for reproducible AI workflows.
China's LLM race has shifted from parameter wars to real-world deployment. Zhipu AI leads with open-source strategy, agentic AI, multimodal integration, cost efficiency, and Chinese-native training. A deep dive into GLM models for developers and enterprises making tech stack decisions.
TRAE stands out because it does more than generate smart answers. It reduces friction, keeps long context, coordinates agentic workflows, and pushes work all the way to a usable result.
M3 hits 59% on SWE-Bench Pro (beating GPT-5.5), adds engineering-grade 1M-token context and native multimodality — the first open-source model with all three. Four verifiable advantages behind MiniMax's 2026 rise and what they mean for your model choices.
In the competitive 2026 AI market, Doubao by ByteDance emerges as the most Chinese-centric AI assistant with four core advantages: native Chinese understanding, multimodal integration, Agent architecture, and localized services. This article analyzes its key competitive strengths.
Just days after its historic IPO, SpaceX announced a USD 60B all-stock acquisition of Cursor's parent company Anysphere on June 16, 2026. This analysis unpacks the three-layer strategic logic: filling the Colossus compute gap, seizing the developer IDE distribution chokepoint, and cutting off Anthropic and OpenAI's traffic — and what it means for enterprise users and the broader AI coding market.