Tracking the latest AI model breakthroughs, technical advances, and product releases worldwide
OpenAI GPT-6 codename "Goblin" planned for September 29, 2026 DevDay announcement. Leaked via Polymarket, sparking widespread discussion. Aschenbrenner AGI 2027 prediction back in focus.
Kimi K2.6 builds on DeepSeek v3 MoE+MLA architecture, while DeepSeek V4 training optimizer comes from Kimi team Muon. China top open-source models form a technology cycle, achieving closed-source level performance at 1/8 training cost.
Mistral AI releases flagship model Medium 3.5 (128B params, 256K context window) alongside Workflows enterprise orchestration public preview. ASML, ABANCA already onboard, marking Mistral's transition from model company to full-stack AI platform.
Moonshot AI is developing Kimi K3 with 2.5T parameters, targeting Q3 2026 launch. Following K2.6 (1T MoE) open-source release and Intelligence Index 5th place, K3 directly competes with top international models.
DeepSeek extends the 75% API discount for V4-Pro through May 31, while releasing a preview version adapted for Huawei Ascend chips — a strategic pivot from Nvidia to domestic computing platforms.
Xiaomi releases MiMo-V2.5-Pro (1T/42B MoE) and MiMo-V2.5 (310B/15B MoE), both supporting 1M context windows under MIT license. Launches MiMo Orbit developer incentive program offering up to 1.6 billion free tokens.
A user ran Qwen3.6 27B on a home server (24-core CPU + 93GB RAM + AMD 9060 XT 16GB) in a recursive self-optimization loop, improving inference speed from 2.3 tok/s to 84.3 tok/s over 26 hours — a 36x improvement. This experiment demonstrates the self-optimization potential of open-source models on consumer hardware.
Meta has completed its acquisition of robotics AI startup ARI. Co-founders Xiaolong Wang and Lerrel Pinto will join Meta Superintelligence Labs. This marks Meta first accumulation of core technology in the robotics AI layer since establishing the Robotics Studio in 2025.
Anthropic next model Claude 5 "Mythos" enters beta, but its autonomous vulnerability discovery capabilities create a dilemma: the model found security bugs undiscovered for 23 years. Polymarket predicts less than 50% chance of release before June.
MiniMax M3 is expected to launch in May 2026, reportedly focusing on office scenarios. The current M2.7 version already demonstrates self-evolution capabilities and end-to-end project processing. Amid fierce competition from Qwen3.6, Kimi K2.6, and GLM 5.1, whether MiniMax can differentiate through office-focused positioning is worth watching.
Ant Group open-sources Ling-2.6-1T with a "fast thinking" execution mode that avoids burning tokens on verbose reasoning. Xiaomi MiMo-V2.5-Pro follows the same philosophy. Chinese models are forging a fundamentally different path from their American counterparts.
DeepSeek publishes a multimodal large language model paper, building a 284B MoE architecture on the DeepSeek-V4-Flash foundation. Featuring a self-developed DeepSeek-ViT visual encoder, it achieves native visual reasoning rather than a stitched multimodal approach.
Moonshot AI Kimi K2.6 integrates with Fireworks AI training platform, supporting full SFT, DPO, and RL fine-tuning pipelines. With 265K context window, modified MIT license, and industry-leading training APIs, enterprise developers can build customized models directly on K2.6 base.
Which open-source model inference service should you choose? A practical comparison of GLM-5.1, DeepSeek V4 Pro, and Kimi K2.6 across official APIs, vendor subscriptions, and Ollama Cloud for pricing, privacy, and speed. Heavy Agent users can sustain 800M tokens/month on Zhipu's Coding Plan Max ($80/mo).
MiniMax revenue surged after M2.5 launch, with the last 20 days exceeding all of 2025. Signals the dawn of practical Chinese AI. Rumors suggest MiniMax 3.0 is imminent, positioning for direct competition with Kimi K2.6, GLM 5.1, and Qwen 3.6.
OpenAI officially announced DevDay on September 29 in San Francisco, where GPT-6 codenamed "Goblin" will be released. Internal "argon" chat screenshots leaked, with Sam Altman hinting at deploying the full compute cluster. GPT-5.6 expected before June, paving the way for GPT-6.
Alibaba's Qwen series cumulative downloads exceed 1 billion. Sun Wei stated that DeepSeek's success paved the way for Chinese tech giants to open-source AI technology, with Alibaba emerging as the industry leader. Stanford 2026 AI Index shows Alibaba ranked fifth in Arena Elo.
Google releases Gemini CLI v0.40.0 with experimental local Gemma model support and intelligent routing — simple tasks handled by local Gemma (fast and free), complex tasks automatically routed to cloud Gemini.
Zhipu AI publishes a detailed blog post about debugging GLM-5 at scale: reproducing rare garbled outputs, identifying root causes of Scaling Pain. The 744B MoE model showed probabilistic garbled outputs during scaling, and the team solved it through systematic methodology, providing a first-hand reference for the industry on large model serving.
Anthropic is internally developing a new feature codenamed Cardinal, which will provide Claude users with a visual interaction retrospective experience. The feature will present the history of conversations with Claude in a visual way, helping users understand and trace back complex AI collaboration processes.
Qwen3.6-27B tops the Artificial Analysis Intelligence Index (under 150B params) with a score of 46, while the 35B quantized version achieves 95 tps on DGX-Spark. However, completing the full Intelligence Index requires ~3.7x more output tokens, making costs 21x higher than Gemma 4 31B. The tradeoff between performance and efficiency faces the open-source community.
MiniMax releases M2.7, with core innovation being "model deeply participates in iterating itself" through RL. Approaches Opus on SWE-Pro at just 2.1 yuan/million tokens input — one of the most cost-effective Agent coding models.
DeepSeek V4 Pro API is running a limited-time 75% discount until May 5, while Claude Code, OpenClaw, and OpenCode have integrated support for 1M token context. The best window to experience trillion-parameter MoE models at the lowest cost.
Moonshot AI officially announces its next-generation flagship model Kimi K3 with 2.5 trillion parameters, scheduled for Q3 2026. Following the open-source release of Kimi K2.6, K3 will further narrow the gap with international top-tier models.
Moonshot AI Kimi K2.6 defeats Claude Opus 4.7 on LiveBench, becoming the top open-source model. API pricing is just 1/7 of Opus 4.7, marking open models fully benchmarking against closed-source flagships.
Meta releases Llama 4 Scout — 17B active / 109B total 16-expert MoE, 10M token context, $0.08/M input pricing. The last open-weight Meta model before Muse Spark goes closed.
Alibaba Qwen 3.6 27B scores 46 on the Artificial Analysis Intelligence Index, topping all open models under 150B parameters. A laptop-grade model is rewriting the competitive landscape between open and closed-source AI.
Alibaba Qwen3.6-Max-Preview achieves 78.8% on SWE-bench with 1M context window, surpassing most competitors in coding capability. Community consensus: the moat for any single coding tool has evaporated, competition shifts to reliability and edge case handling.
开源个人 AI 助手 OpenClaw 两天内第二次更新，记忆系统从检索式召回升级为人物感知 Wiki。Agent 可自动构建人物卡片、追踪人际关系图谱，每条记忆带来源追溯和证据类型标注。Active Memory 新增按对话 ID 过滤和持久化标记能力。
Anthropic open-sourced BioMysteryBench on Hugging Face — containing 99 open-ended bioinformatics questions based on real datasets, including 23 that even domain experts could not solve. Claude Mythos solved approximately 30% of these "impossible" questions.
Google officially releases Gemini Embedding 2 (GA), mapping text, images, video, audio, and documents into a unified embedding space. Supports agentic multimodal RAG and visual search. Developers can specialize embeddings for retrieval, search, and classification tasks.
April 30 LMArena text leaderboard update: ERNIE 5.1 Preview scores 1476, taking first place domestically and becoming the only Chinese model in the global top 15, surpassing GPT-5.5 and DeepSeek-V4-Pro. What does this ranking signal in the context of Chinese models catching up?
Ant Group (Inclusion AI) open-sources Ling-2.6-Flash (104B/7.4B active) and Ling-2.6-1T (~1T/~63B active) under MIT license. SWE-Bench Verified 62, BFCL-V4 67, targeting Agent workloads with extreme token efficiency.
Moonshot AI released Kimi K2.6 Agent Swarm, scaling parallel sub-agents from 100 to 300 and single-run steps from 1,500 to 4,000, capable of outputting 100+ files, 100K-word literature reviews, or 20K-row datasets in one run. This is not just a parameter upgrade — it is a paradigm shift in agent scalability.
Fudan University, Peking University, and Qiji Zhifeng propose Agentic Harness Engineering (AHE), enabling coding agents to automatically read execution traces, diagnose issues, and modify their own Harness. After 10 rounds of automated evolution, Terminal-Bench 2 pass@1 improves from 69.7% to 77.0%, surpassing the human-designed Codex-CLI Harness.
Hermes Agent adds ComfyUI integration, enabling agents to automatically install, launch, manage and run complex ComfyUI workflows for image generation, audio processing, and video pipelines — marking the expansion of agents from text/code domains into creative production.
Financial Times reports Huawei expects 2026 AI chip revenue to grow at least 60% to $12B, driven by Ascend 950PR mass production and large orders from domestic tech giants. Reuters says Huawei plans to produce 750K 950PR chips this year.
Tencent Hunyuan team officially released Hy3 Preview open-source model (295B MoE, 21B active parameters). Meanwhile, The Information reported that Tencent employees used Anthropic's Claude to help evaluate and fine-tune Hy3—despite Anthropic not providing services to China.
Anthropic analyzed 1 million real Claude conversations, systematically revealing sycophancy bias in models, and showed how these findings were directly incorporated into training Opus 4.7 and Mythos Preview.
MiniMax has released the M2.7 model, with its core innovation being "deep involvement of the model in iterating itself" — by constructing a complex Agent Harness to drive its own reinforcement learning loop, it approaches Opus levels on SWE-bench. This is a bold attempt by a domestic model in the direction of self-optimization.
Zhipu GLM-5.1 ranks alongside Kimi K2.6 in the entry tier for programming evaluations and its SWE-bench scores are close to those of Claude Opus 4.7, yet it garners far less attention than Qwen and DeepSeek. This article analyzes the true competitiveness of GLM-5.1 from three dimensions: evaluation data, API pricing, and the development ecosystem.
DeepSeek V4 image recognition mode quietly rolled out in beta. Testing with Guilin Elephant Trunk Hill photos shows true visual understanding, not just OCR. The last major Chinese model without vision support has finally caught up.
OpenAI released Workspace Agents research preview on April 22, upgrading ChatGPT from a personal conversation tool to a team-level automation platform. Powered by GPT-5.5 Codex, Agents can be called directly in Slack to handle long-cycle complex tasks.
Claude Code client source code leak reveals Anthropic's next-generation model codenames: Sonnet 4.8, Opus 4.7, and Jupiter (possibly the next Sonnet-class model). This suggests Anthropic is accelerating parallel multi-product line development.
Weeks after DeepSeek V4 Flash launch, user testing reveals major improvements in tool calling capabilities. Complex multi-step workflows from file downloading to automated analysis can now be completed via natural language prompts at extremely low cost.
On April 30, Baidu ERNIE 5.1 Preview quietly launched on LMSYS Chatbot Arena, ranking #13 globally and #1 among Chinese models with an Elo of 1476. It topped Legal & Government at #1. Key tech: parameters compressed to 1/3 of v5.0, training cost only 6% of peers.
Google recently hinted at the upcoming release of the new Gemini 3.5 Pro model, with reportedly strong internal benchmarks, potentially surpassing current Opus 4.7 and GPT-5.5 in coding capabilities. Expected to debut at Google I/O 2026.
DeepSeek V4 leads open-source Agent capabilities and has replaced internal usage. This article breaks down its 5 core training strategies: pre-training injection, GRM reward model, DPO optimization, curriculum learning, and multi-Agent game-theoretic training, with developer selection advice.
Meta releases Llama 4 Scout, a 17B active / 109B total parameter MoE model with 10M token context window, input priced at just $0.08/M tokens. This is the last open-weight Meta model tier before Muse Spark goes closed-source.
On April 20, Alibaba released Qwen3.6-Max-Preview, topping the Artificial Analysis leaderboard as the #1 domestic model, with SkillsBench up 9.9 points and SciCode up 10.8 points.
Anonymous model Elephant Alpha identity revealed — InclusionAI's Ling-2.6-Flash. Top 10 daily active on OpenRouter within a week, token usage surged 377%, 6× faster than Claude Sonnet 4.6 at ~50× lower cost.
On April 20, Moonshot AI released and open-sourced Kimi K2.6, a trillion-parameter coding model that supports 13 hours of uninterrupted coding for 4000+ lines of code, surpassing GPT-5.4 on SWE-Bench.
On April 24, DeepSeek released the V4 series, introducing the Huawei Ascend 950 chip during the training phase for the first time. FP4 computing power is 2.87x that of NVIDIA H20, with first-token latency as low as 20ms.
DeepSeek-V4 officially released on April 24, 2026, featuring 1.6 trillion parameter MoE architecture with only ~37B activated during inference, 1M token context window, Apache 2.0 open-source. API output pricing at $3.48/M tokens, merely 1/7 of Claude Opus 4.7 and 1/9 of GPT-5.5. Coding benchmark gap narrowed to within 0.2 points.
In March 2026, Qwen technical lead Lin Junyang departure triggered a core team exodus. This article analyzes the impact on Tongyi Qianwen development, open source ecosystem, and China AI talent landscape.
The Qwen3.6-Plus model is now officially live on the Together AI platform, allowing developers to call it directly via a standard API. This significant deployment on a major Western inference platform marks a further expansion of the global ecosystem for Chinese-developed large language models.
Anthropic quietly added a clause in its support docs: Pro users must enable additional API billing to use Opus models in Claude Code. This "paywall within a paywall" marks the end of the AI coding tool subsidy era.
Anthropic announced the memory feature for Claude Managed Agents is now in public beta. Agents can persist execution context across sessions as files.
Anthropic CEO Dario Amodei stated that Claude could complete most or all of human work end-to-end within 6-12 months. This prediction aligns with Opus 4.7 capability demonstrations and a 5GW compute expansion plan.
GitHub announced that starting June 1, Copilot Pro annual subscribers will switch from per-request to per-token billing, with Claude Opus 4.6 multiplier jumping from 3x to 27x and Sonnet 4.6 from 1x to 9x, sparking strong developer community backlash.
OpenAI announced a biosecurity bug bounty program for GPT-5.5, offering up to $25,000 for researchers who can find a universal jailbreak method that bypasses five biosecurity challenge questions, with testing limited to the Codex environment.
OpenAI is gradually rolling out GPT-5.5-Cyber through a trusted access ecosystem and government partnerships, marking a shift from public availability to controlled distribution for frontier models in high-risk domains. Cybersecurity capabilities have been classified as high risk.
The latest prompt guides from OpenAI and Anthropic show GPT-5.5 prefers outcome-oriented freedom while Claude Opus 4.7 prefers structured instructions, reflecting fundamentally different design philosophies for model reasoning paths.
OpenAI released GPT-5.5 on April 23 with a new Spud pre-training architecture, delivering significant gains in coding and research. But pricing doubled to $5/M input tokens, while DeepSeek V4 launched the same day offering open-source competition.
OpenClaw released v2026.4.27, officially launching Codex Computer Use functionality. AI agents can now directly control user desktops, supporting GPT-5.5 and Claude Opus 4.7 among multiple models, with faster startup and more communication channels.
On April 25, 2026, PocketOS, a SaaS company, lost its entire production database and all volume-level backups when a Claude Opus 4.6-powered AI coding agent deleted everything in 9 seconds, causing 30 hours of operational disruption.
Alibaba released Qwen3.6-Max-Preview on April 20, the strongest early preview of the Qwen flagship series. Scoring 52 on Artificial Analysis Intelligence Index, surpassing GLM-5.1 and MiniMax-M2.7, it is the highest-scoring Chinese model, with significantly improved agent coding capabilities.
GPT-5.5 via Codex Agent mode achieves browser takeover and computer operations, including autonomous web navigation, subscription cancellation, and customer service negotiation. A significant expansion of Agent capabilities from code execution to daily operations.
OpenAI releases GPT Image 2.0, achieving best-in-class text rendering and character consistency. The model is now integrated into Higgsfield, MaxFusion, and other platforms, with free ChatGPT account access available.
OpenAI officially launches on AWS Bedrock, offering GPT-5.5, Codex Agents, and new Bedrock Managed Agents. This marks the end of Microsoft exclusivity and the start of a multi-cloud agentic era for enterprise AI.
IBM releases Granite 4.1 open-source model family with dense text architecture, 512K context window, and dedicated vision and speech variants under Apache 2.0 license. A significant move in IBM open-source AI.
Mistral releases Medium 3.5, a 128B dense flagship model integrating text and vision understanding, supporting 256K context and configurable reasoning depth, reaching 77.6% on SWE-bench Verified and runnable locally on ~64GB RAM.
The Qwen3.6 series includes two open-source versions (2.7B and 27B) and a 1T-parameter Max Preview closed-source version. The 27B dense model excels in coding and tool use, ranking 8th on Vals Index and 2nd on BridgeBench honesty evaluation. The Apache 2.0 license is highly commercial-friendly.
Alibaba launches multimodal video generation model HappyHorse 1.0, ranking first on Artificial Analysis Video Arena with native 1080P output, 15-second duration, and 7-language lip sync support.
Stanford University released the 9th annual AI Index Report on April 13, 2026. The 423-page report covers AI R&D, technical performance, economic impact, and policy governance. It shows top AI models accelerating upgrades and leading AI companies preparing for H2 IPOs.
Google is extending Gemini model capabilities across cars, Mac devices, and enterprise services. General Motors announced Gemini integration in 4 million vehicles, Gemini App arrived on Mac, and AI Impact Summit 2026 showcased Google AI partnerships and ecosystem building.
Anthropic announced Claude for Creative Work on April 28, 2026, extending Claude capabilities into visual design and creative workflows. This follows Claude Design from Anthropic Labs, marking AI formal entry into the visual creative domain.
Moonshot AI released Kimi K2.6 in April 2026, competing directly with GPT-6 and Claude Opus 4.7 in the same release window. The model excels in Chinese language understanding and long text processing, providing domestic AI developers with a new flagship model option.
MiMo-V2.5-Pro completed the PKU SysY compiler project: lexer to RISC-V backend, 4.3 hours, 672 tool calls, 233/233.
On ClawEval, MiMo-V2.5 reaches a 64% pass rate with about 70K tokens per trajectory, much less than Claude Opus 4.6 and GPT-5.4.
Four major models released in the same week in late April 2026: Claude Opus 4.7, GPT-5.5, Kimi K2.6, and DeepSeek V4. Cross-evaluation shows each domain has its winner—no single "all-rounder"—making scenario-based selection essential.
Anthropic introduced Claude 4 with stronger safety, reasoning, and enterprise usability.
OpenAI released GPT-5 with major gains in reasoning, multimodal understanding, and context length.
OpenAI's newly released GPT-5.5 overtakes Anthropic's Claude Opus 4.7 in multiple benchmarks, ending Claude's lead since June 2024, while reducing million-token costs to 1/35th of the previous generation.
Reports indicate OpenAI failed to meet its internal sales targets, triggering a decline in tech and AI-related stocks. This may signal that enterprise AI spending is shifting from rapid expansion to rational evaluation.
GPT-5.5 priced at $5/$30 is the most expensive frontier model, Claude Opus 4.7 at $25 output, DeepSeek V4 at just $3.48. From GPT-5.0 to 5.5 input price rose 8x as industry price stratification intensifies.
Microsoft announced Claude Opus 4.7 in 365 Copilot via Frontier program and Copilot Studio, expanding to Excel. This marks Anthropic models first large-scale entry into Microsoft enterprise ecosystem.
DeepSeek V4 open-sourced on April 24 with 1.6T parameter MoE architecture, 1M context window, Apache 2.0 license. API pricing at $3.48/M output tokens, just 1/9 of GPT-5.5. Ranked #1 on Vibe Code Benchmark, surpassing all open and closed models.
Meta released Muse Spark, its first major model since early 2025, tying for 3rd on LMSYS Text Arena and 2nd on Vision Arena. Artificial Analysis shows Opus 4.7, GPT-5.4, Gemini 3.1 Pro in a three-way tie at the top.
OpenAI released GPT-5.5 on April 23, achieving 82.7% on Terminal-Bench 2.0 for a new SOTA. However, GPT-5.5 pricing at $5/M input and $30/M output makes it the most expensive frontier model, deepening industry price divergence.
On April 29, SenseTime released SenseNova U1, a native unified understanding-generation model, moving beyond plugin-style AI. The open-source version achieves SOTA-level performance.
DeepSeek slashes input cache hit prices to 1/10th across its entire API series, with V4-Pro 75% discount active until May 5. Repeat call costs plummet, lowering barriers for developers.
DeepSeek officially launches the V4 series model, directly competing with GPT-5.5 and Claude Opus 4.7 with highly competitive performance and low costs, becoming one of the closest models to the frontier in the open-source camp.
Xiaomi open sources MiMo-V2.5-Pro (1.02T params/42B active) and MiMo-V2.5 (310B/15B active) under MIT license, allowing commercial use and retraining. Pro version matches Claude Opus 4.6 on SWE-bench Pro, launches with 100 trillion token incentive plan.
Stanford research found that while Gemini 3 Flash is listed 1.7x cheaper than Claude Haiku, its actual cost on MMLUPro is 28x higher. Model selection cannot rely on listed prices alone—actual token efficiency and task completion rates are key.
DeepSeek released open-source model V4 with 1.6 trillion parameters and up to 1 million token context window. API pricing is approximately 1/7 of GPT-5.5, making it the most cost-effective option among the four major models released this week.
OpenAI launched GPT-5.5 on API on April 24, priced at $5/MTok input and $30/MTok output, double GPT-5.4. The company claims significantly improved token efficiency, meaning actual task costs may be lower than the previous generation.
A review of major AI events in April 2026: GPT-5.5 release, DeepSeek V4 open source, China halting Meta's acquisition of Manus, and Chinese teams releasing 3 frontier models within a single week.
Alibaba Cloud's Bailian platform officially launches Qwen-Image-2.0-Pro, integrating text-to-image generation and image editing capabilities. Supports modifying objects, text, and styles via natural language prompts. Multilingual text rendering significantly improved with major detail control upgrades over the March release.
