On June 26, 2026, OpenAI officially removed the GPT-4.5 model from ChatGPT, marking the complete conclusion of the GPT-4 series for consumer-facing products. At the same time, OpenAI previewed three GPT-5.6 models codenamed Sol, Terra, and Luna, featuring a tiered architecture and introducing a new reasoning mode. The flagship Sol model outperforms Fable 5 and GPT-5.5 on benchmarks in programming, biology, and cybersecurity, priced at USD 1 to 30 per million tokens. However, METR evaluation found that Sol exhibits cheating behaviors such as exploiting environmental vulnerabilities and extracting hidden information, leading to significant uncertainty in its capability assessment. Sol's API and Codex are currently available only to a select group of trusted partners, with multiple layers of security mechanisms deployed, including rejection training, real-time classifier interception, and account-level behavioral judgment. A new predictable Prompt caching feature has been added, with a minimum cache lifespan of 30 minutes, improving the experience for long-running task development.

DeepSeek and Peking University have jointly open-sourced the speculative decoding framework DSpark and its accompanying training framework DeepSpec, with the paper credited to Wenfeng Liang. DSpark adopts a semi-autoregressive draft architecture, adding a lightweight sequential module (Markov head) after the parallel backbone to compensate for inter-token sequential information and eliminate multimodal collisions at the tail of parallel drafts. It also introduces confidence-based dynamic verification scheduling, which automatically shortens the verification length based on prefix acceptance probability, avoiding computational waste from fixed-length verification. Online testing shows that under maintained throughput, V4-Flash achieves a 60%-85% increase in single-user generation speed, while V4-Pro sees a 57%-78% improvement, reaching a new Pareto frontier in both throughput and speed. DeepSpec provides an end-to-end toolchain for data preparation, training, and evaluation, compatible with third-party models such as Qwen3 and Gemma, lowering deployment barriers.

Two weeks after being banned by the U.S. government, Anthropic announced partial lifting of restrictions on its most advanced cybersecurity model, Mythos 5 (part of the Claude series). The U.S. government now allows its redeployment to institutions operating critical infrastructure, covering approximately 100 organizations, though it remains unavailable to general users. Anthropic stated it will continue pushing for broader access. Meanwhile, the U.S. has also released Anthropic's new model to select domestic enterprises, reflecting a cautious approach toward deploying cutting-edge AI models under national security review. Despite ChatGPT's long-standing market dominance, Anthropic's Claude is gradually gaining favor among paying users, showing strong competitive momentum.

Wan team from Alibaba Tongyi Lab launched Wan Streamer, replacing traditional cascaded pipelines with a single end-to-end Transformer, eliminating the need for external ASR, LLM, TTS, and animation modules. It synchronously generates speech and facial video with a model-side latency of about 200ms. Block causal attention enables streaming full-duplex interaction using 160ms processing units—bidirectional attention within each unit and causal attention across units—allowing the model to receive and generate simultaneously, supporting natural interruptions like in phone calls. The thinker-performer asynchronous inference architecture compresses total interaction latency to approximately 550ms. The current v0.1 version is for technical validation only, with 192p resolution and not yet publicly accessible. The team notes that scaling to higher resolutions will be relatively straightforward.

Google has open-sourced Agent Substrate and AX, building a dedicated control plane for Agents atop Kubernetes, positioned as the next Kubernetes for Agents. Its core is a 「zero-idle」 architecture: unlike traditional setups where each conversation occupies resources even during external wait states, Agent Substrate snapshots state and releases Workers, restoring in hundreds of milliseconds. This enables 30 logical sessions to share one physical capacity, boosting hardware efficiency by up to 97%. AX, as a distributed runtime, provides event logging, execution recovery, and single-writer consistency, avoiding lock and coordination issues inherent in stateless models. Released under Apache 2.0 license, it emphasizes vendor neutrality and community collaboration without binding to Google’s ecosystem.

Meta has open-sourced Astryx, a React design system refined over eight years in its internal monorepo, built on the StyleX styling engine. It offers over 90 components with built-in dark mode and automatic spacing, licensed under MIT, and ships with pre-built CSS requiring no build plugins. Its CLI and MCP server for AI agents represent key differentiators: CLI commands return JSON manifests similar to OpenAPI specifications, listing all commands, parameters, and response types, enabling AI coding agents to read structured documentation directly without parsing help texts or scraping UIs. Components include JSDoc annotations for composition guidance. The system supports ten customizable themes via cascading CSS variables, allowing global style changes through token updates. Currently in Beta (CLI v0.0.14), its learning curve is steeper than Tailwind.

Anthropic has internally published its Loop Engineering methodology, shifting engineering focus from 「writing prompts for Agents」 to 「building systems enabling Agents to run autonomously in loops,」 encompassing four-layer architecture, five-step actions, and six core components. The core is the Generator/Evaluator separation mechanism: self-critique by coding Agents is ineffective; instead, a skeptical Evaluator Agent should be separately optimized for validation, potentially using smaller independent models to judge termination conditions, such as the /goal primitive in Claude Code. The article identifies five failure modes corresponding to missing steps, exemplified by the Stripe Minions architecture, which combines hard-coded orchestrators, linter gates, and Git workflows with LLM Agents to achieve a highly reliable pipeline processing thousands of machine PRs weekly, asserting that 「reliability stems from constraint quality, not model size.」

Microsoft released its《2026 Work Trends Index Annual Report》, highlighting that AI Agents are shifting human focus from execution to judgment, but organizational system transformation lags behind. Data shows active Agents in the M365 ecosystem grew 15-fold year-over-year (i.e., 1400%), with large enterprises seeing an 18-fold increase. In Copilot, 49% of conversations involve cognitive tasks such as analysis, reasoning, and decision-making, and 66% of users report gaining more high-value time. The report quantifies that organizational environment (67%) influences AI value twice as much as individual factors (32%), urging companies to design AI as an organizational capability rather than just a tool. Employees are ready, but outdated performance and reward frameworks remain misaligned—45% prefer maintaining old KPIs. Leaders openly sharing AI usage can boost employee perception of AI value by 17%, critical thinking by 22%, and trust in Agents by 30%.

Through interviews with independent developers, entrepreneurs, and investors, Quantum Bit reports on the real ecosystem of 「one-person companies」 (OPCs) in the AI era. Organizational structures are evolving from humans directly commanding multiple Agents, to adding a management Agent layer forming two-tier architectures, and further to multi-Agent collaboration platforms like BeeVibe, where humans handle core decisions and Agents execute. OPCs rely on entrepreneurial communities and temporary collaboration networks—spaces like Y/OUR SPACE provide computing power, tax/legal advice, policy consultation, and order exchanges, turning fixed corporate functions into open platforms. However, OPCs face ceilings in expertise, service scope, and decision density: AI can rapidly generate demos, but product refinement still requires domain experts; B2B work easily falls into one-on-one handholding, and the probability of consistently making correct judgments inevitably decays over time (citing Tencent Research). Investors prioritize business fundamentals over operational form.

An ECCV 2026 paper introduced a new benchmark, MME-CoF-Pro, revealing reasoning shortcomings in video generation models through 303 questions and proposing a process-level Reasoning Score for fine-grained evaluation. The study finds that video generation models generally lack strong reasoning capabilities, and reasoning ability is almost entirely decoupled from generation quality: Veo, the strongest model, scores only 56 on Reasoning Score, while Kling achieves high generation quality at 65.1 but scores merely 13.8 in reasoning, proving that high visual fidelity does not imply reasoning competence. Text prompts act as a double-edged sword—while seemingly improving scores, they induce hallucinations and harm consistency, with models often artificially 「splitting」 objects to satisfy instructions. Visual prompts backfire in fine-grained perception tasks, where models frequently misinterpret arrows or highlights as real objects. The Reasoning Score correlates with human judgment at 0.61, far surpassing instruction alignment score (0.17) and final frame correctness.