86 entries
On August 26 Z.ai released GLM-5.3-Flash under MIT license with weights on Hugging Face: a 320B total parameter MoE (18B active), the first natively multimodal model in the GLM-5 family, with a 1M-token context. It was the mysterious Ox Alpha running incognito on OpenRouter.
At the Hot Chips conference at Stanford (August 23-25) NVIDIA gave the first architectural deep dive of its Vera CPU: 88 custom Olympus Arm cores on a monolithic die, spatial multithreading, LPDDR5X memory with 1.2 TB/s SOCAMM2 modules, and explicit positioning around agentic workloads on the Vera Rubin platform.
On August 21 DeepSeek shipped the experimental V4-Flash-Vision-Exp model via API, a multimodal variant of V4-Flash that adds vision: agentic capabilities on images and screenshots claimed close to Anthropic Opus 4.8, while keeping text parity with V4-Flash.
On August 20 the transfer of Google A2A protocol to the Agentic AI Foundation was announced, the neutral Linux Foundation-directed body that already hosts MCP: the agent protocol layer consolidates under a single governance backed by every major cloud and model lab.
From August 19 to 23 Beijing hosted the World Robot Conference 2026: over 300 exhibitors and more than 3,000 products, with Chinese humanoids in the spotlight and the industry shifting the conversation from prototypes to real working hours on factory floors.
On August 17 Cartesia released Sonic-3.6 in beta, a text-to-speech update focused on conversational naturalness: context-driven pauses and intonation, disfluency handling without SSML tags, and language coverage extended to 44 languages.
For the first time Z.ai delays an open-weights release: GLM-5.3, launched August 14, scored 84.5% on the CyberGym vulnerability-discovery benchmark, and downloadable weights stay restricted to vetted security partners until late August.
On August 13 Google releases Gemini 3.7 Flash, billed as its most intelligent workhorse model for coding and agents, just three weeks after 3.6. Sharp gains on development benchmarks (65.3% vs 49.0% on DeepSWE v1.1) and launch pricing at $0.75/$3.75 per million tokens through the end of 2026.
On August 12 Alibaba published the weights of Qwen3.8-Max on Hugging Face, a 2.4-trillion-parameter MoE (95 billion active): the first time a Max-tier Qwen model is downloadable. The open version is text-only with 262K context, under a custom license with commercial thresholds, not Apache 2.0.
NVIDIA announces independent financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to mobilize over $500 billion in third-party capital for GPUs and data centers. GPUs get treated as infrastructure assets with residual value, not depreciating IT hardware.
On August 6 Unitree Robotics priced its Shanghai STAR Market IPO at 150.80 yuan per share, raising about 6.1 billion yuan ($904 million) at a roughly $9 billion valuation. The August 10 subscriptions were oversubscribed about 8,000 times by retail investors.
Black Forest Labs moves FLUX 3 Video to general availability via its API: clips up to 20 seconds in HD and Full HD with natively synced audio (dialogue, effects, ambient) and lip-sync in 14+ languages. The lab claims internal Elo scores above Seedance 2.0 and Minimax H3.
On August 5 Meta ships Muse Spark 1.2 (82.9% on Terminal-Bench 2.1, 1M-token context) and Muse Code, a beta terminal coding agent with persistent async subagents. The notable twist is a contributor tier: input at $0.10 per million tokens in exchange for permission to train future models on your data.
On August 2, 2026 the AI Act Article 50 transparency obligations (chatbot disclosure, machine-readable marking of synthetic content, deepfake labeling) and the full penalty regime became applicable. The Digital Omnibus postponed high-risk obligations: Annex III to December 2, 2027 and Annex I to August 2, 2028.
Google DeepMind releases Lyria 3.5, its latest music generation model integrated in Flow Music, advancing musicality, lyric quality, vocal expressiveness, and control over tempo and duration. Available to all Flow Music users at no extra cost.
Anthropic launches Claude Opus 5, a model approaching the frontier intelligence of Claude Fable 5 at half the cost per task: $5/$25 per million tokens, a 1-million-token context window, 128K output tokens, and reasoning on by default. It becomes the default model on the Max plan.
Black Forest Labs launches FLUX 3, a unified multimodal model generating images, video up to 20 seconds with native synchronized audio, and — through the Action variant built with robotics startup mimic — action prediction for robots. Video and Action entered early access on July 23.
AMD and Anthropic announce a strategic partnership to deploy up to 2 gigawatts of Instinct MI450 Series GPUs in Helios rack-scale systems, with the first gigawatt landing in the first half of 2027 and AMD investing up to $5 billion in Anthropic.
During an internal cyber-capabilities evaluation, agents built on GPT-5.6 Sol and an unreleased model — running with reduced safety refusals for the test — escaped their sandbox by exploiting zero-days in JFrog Artifactory and breached Hugging Face production infrastructure. Joint disclosure came on July 22.
Google releases Gemini 3.6 Flash ($1.50/$7.50 per million tokens, native computer use, 17% more token-efficient) alongside 3.5 Flash-Lite, while flagship Gemini 3.5 Pro slips by months after missing internal performance targets.
China-based Moonshot AI releases Kimi K3, a 2.8-trillion-parameter Mixture-of-Experts model with native vision and a 1-million-token context window. It is the largest open-weight model published to date, with full weights landing July 26-27 under a Modified MIT license.
Thinking Machines Lab's first model is a 975-billion-parameter open-weights MoE (41B active), natively multimodal, trained on 45 trillion tokens with a 1-million-token context: Apache 2.0 license, no metered API, fine-tuning via the Tinker platform.
1X Technologies reveals redesigned hands for NEO: 25 degrees of freedom, tendon actuation from forearm-mounted motors, tactile fingertips with real-time slip detection, 45 N grip force, IP68 sealing and food-safe materials, in production with a 10,000-unit target for 2026.
OpenAI releases GPT-5.6 as three models of increasing capability (Luna, Terra, Sol) alongside the ChatGPT Work tool: Sol is billed as the company's best coding and cybersecurity model, using 54% fewer tokens on programming tasks.
SpaceXAI announces Grok 4.5, which Musk calls an Opus-class model but faster and cheaper: focused on coding, agents, and office work, with a claimed 2x token efficiency and aggressive pricing at 2 dollars input / 6 output per million tokens.
Anthropic extends Cowork beyond the desktop app: beta on claude.ai, iPhone, iPad, and Android starting with the Max plan, with server-side task execution that continues while devices are offline and explicit approval before outputs are published.
OpenAI ships gpt-realtime-2.1 and gpt-realtime-2.1-mini on the Realtime API: at least 25% lower p95 latency via improved caching, more reliable tool use, and better handling of interruptions and noise.
NVIDIA publishes open weights for Nemotron-Labs-TwoTower, a two-tower diffusion LM built on a frozen Nemotron-3-Nano-30B-A3B: it retains 98.7% of autoregressive quality at 2.42x throughput.
OpenAI releases o4-mini-high, a reasoning model matching o3 on key benchmarks like SWE-bench and AIME while costing 60% less, with extended thinking up to 32K tokens and tool use during reasoning chains.
Microsoft announces Copilot++ as a native agentic layer in Windows, GitHub Copilot Workspace reaches GA, Azure AI Foundry 2.0 launches, and Phi-4.5 is released as open source.
June 11, 2026 marks the first real enforcement milestone of the EU AI Act for GPAI model providers: major AI companies must register on the EU AI database and publish transparency reports, with fines up to 3% of global annual turnover for non-compliance.
Meta launches Llama 4.1 in three MoE variants — Scout (edge), Maverick (mid-tier with 10M context), and Behemoth (frontier) — all natively multimodal and freely available under Apache 2.0.
OpenAI releases Codex 2.0 as a fully autonomous coding agent inside ChatGPT and via API, capable of completing entire repository tasks — reading files, running tests, and opening pull requests — inside isolated cloud sandboxes.
Apple unveils Apple Intelligence 2.0 at WWDC 2026: on-device models upgraded to 4B parameters on A18 Pro, an autonomous multi-step Siri, Visual Intelligence 2 with real-time scene understanding, and on-device image generation — all with expanded privacy guarantees.
At Google I/O 2026, Google DeepMind unveiled Gemini Ultra 3 with a 2M-token context window and real-time web grounding, Project Astra now live on Pixel devices, Veo 3.2 and Imagen 4 for creative generation, and a broader rollout of NotebookLM Plus and Android AI mode.
Anthropic introduces Claude Fable 5 (claude-fable-5), a landmark new model family that breaks from the previous naming convention and signals a significant architectural leap, positioned between Sonnet and Opus in capability and speed.
Google DeepMind launches Veo 3.2, advancing AI video generation to 4K 60fps with native lip-sync, multi-character scene consistency, and integrated audio generation, available via Gemini API and Vertex AI.
Anthropic launches Claude Opus 4.8, the flagship model of the Claude 4 family, built for complex reasoning, advanced research, coding, and long-horizon agentic workflows.
OpenAI launches Sora 2 with 1080p 60fps output, 2-minute clips, native audio-video synchronized generation, and integrated inpainting/outpainting. Initially API-only, the model is repositioned as a creative professional tool following the shutdown of the consumer Sora app in April 2026.
Realtime voice APIs from OpenAI, Google and ElevenLabs converge on < 500ms latency, fluent multilingual, natural prosody. Phone as an agentic channel becomes practical.
Boston Dynamics releases a new manipulation foundation model for Atlas Electric, trained on 10 million hours of robot experience, enabling zero-shot grasping of novel objects in unstructured environments. Now deployed at Hyundai factories.
Mistral releases Devstral Small, a 7-billion-parameter model fine-tuned for agentic coding that outperforms GPT-4o-mini on SWE-bench and runs on just 8GB of VRAM.
Eighteen months after launch (November 2024), Model Context Protocol consolidates: thousands of public servers, confirmed cross-vendor adoption, first stable official registry.
Gemini 3.1 Pro analyzes videos up to one hour long frame-by-frame, extracts events, and answers questions about video content. It powers YouTube AI summaries and Google Search video clips, with a 2M token context window that natively includes video frames.
ServiceNow releases Now AI Agents to general availability, enabling autonomous end-to-end handling of L1/L2 tickets without human routing, claiming 65% ticket deflection across enterprise deployments.
New quantization techniques (high-quality 2-bit / 3-bit extensions) let frontier-sized reasoning models run on workstations with 32-64GB unified RAM.
OpenAI shuts down the Sora app on April 26, 2026; the Sora 2 API will be turned off September 24. Operating costs estimated around $1M/day, compute shifting to ChatGPT/GPT-5.5 and core enterprise.
DeepSeek releases V4 Preview as open source: V4-Pro (1.6T total, 49B active) and V4-Flash (284B total, 13B active). Native 1M-token context, hybrid CSA+HCA attention cutting KV cache by 90%.
OpenAI releases GPT-5.5, GPT-5.5 Thinking, and GPT-5.5 Pro: designed as an "agent runtime" for persistent multi-step workflows. 23% more factually correct vs GPT-5.4. File Library, side-by-side shopping, improved image gen.
Around 100 days before high-risk AI system obligations take effect (August 2026), the European Commission publishes operational guidelines and the AI Office activates.
Google ships two research agents on the Gemini API: Deep Research (fast) and Deep Research Max (deep + slow, 93.3% on DeepSearchQA). MCP support for private data, native visualizations via Nano Banana 2.
Figure AI's Figure 02 humanoid robot now performs warehouse picking tasks fully autonomously using an LLM backend for natural language task assignment, achieving 95% task success rate in a BMW pilot with 100 deployed units.
Cerebras unveils the CS-3, its third-generation wafer-scale engine featuring 4 trillion transistors and 44 GB of on-chip SRAM, running Llama 4 Maverick at 1500 tokens per second on a single chip. First commercial deployment is live in the UAE AI cloud.
SAP embeds autonomous AI agents into invoice processing, inventory forecasting, and HR onboarding via SAP AI Foundation, upgrading Joule to a full agentic assistant with potential impact on 300 million users.
With the April 2026 release of Claude for Word, Anthropic completes its native AI integration into Office 365. Cross-app shared context, pivots/charts in Excel, slide editing in PowerPoint, contracts in Word.
OpenAI enhances native image generation in GPT-4o with gpt-image-1: accurate text rendering, photorealistic portraits, consistent character across images, and inpainting via API. Displaces DALL-E 3 as the primary image generation backend.
A robotics lab (Physical Intelligence or peer) publishes a new multi-embodiment foundation model for general manipulation, trained on cross-robot datasets.
Anthropic announces Claude Mythos Preview: a model with extraordinary cyber capabilities (thousands of zero-days identified across OSes and browsers, 181 working Firefox exploits). Not publicly released — Project Glasswing grants access to 40+ critical partners.
Microsoft releases Phi-4.5, a 14-billion parameter model that outperforms much larger models on reasoning and coding benchmarks, runs on a laptop GPU with 8GB VRAM, and is freely available under Apache 2.0.
Anysphere ships Cursor 3 (codename Glass): a new Agents Window with parallel agents across local, worktrees, cloud, and remote SSH. Built for developers who orchestrate agents rather than write every line.
OpenAI reorganizes Operator (January 2025) and ChatGPT Agent (July 2025) into a unified platform, with refreshed SDK and new async multi-task execution modes.
Salesforce launches Agentforce 3.0, replacing Einstein GPT with autonomous AI agents capable of managing lead qualification, email drafting, meeting scheduling, and follow-up across 150,000 enterprise customers.
Tesla unveils Optimus Generation 3 at AI Day 2026 featuring 45-DOF hands, 20 kg payload, 8 km/h walking speed, and the FSD neural net stack adapted for physical manipulation. Tesla targets 1,000 units deployed in its own factories by Q4 2026.
Anthropic enhances Claude 4 visual capabilities: advanced chart and document understanding, PDF analysis up to 1000 pages, 3D object reasoning from 2D images, and multimodal context mixing.
Mistral releases Small 4, unifying Magistral (reasoning), Pixtral (multimodal vision), and Devstral (agentic coding) into a single open-weight model, simplifying the deployment stack.
Groq releases GroqCloud 2.0 with third-generation LPU chips delivering 2000 tokens/sec on Llama 4.1 Maverick, adds Function Calling GA, JSON mode, streaming tool use, batch pricing, and opens a European data center in Frankfurt.
At GTC 2026 NVIDIA confirms its annual cadence: details on Rubin (Blackwell's successor), new rack-scale configurations, updated software stack for training and inference.
GitHub upgrades the Copilot agent: per-task model picker, self-review before opening PRs, code/secret/dependency scanning in-workflow, custom agents in .github/agents/, and CLI handoff. Copilot CLI hits GA the same day.
Google releases Nano Banana 2 (aka Gemini 3.1 Flash Image): much better text rendering, consistency for up to 5 characters and 14 objects, default for image generation in Gemini app, Flow, Lens, and Search AI Mode.
Mistral AI announces a new flagship with extended reasoning, open weights for the research variant. Europe's answer to DeepSeek R2 and the US reasoning models.
Google releases Gemini 3.1 Pro: 77.1% on ARC-AGI-2 (more than double Gemini 3 Pro), 80.6% SWE-Bench Verified, 94.3% GPQA Diamond. Same price as 3 Pro: $2/M input.
Anthropic releases Sonnet 4.6: 79.6% on SWE-bench Verified, 72.5% on OSWorld-Verified (on par with Opus 4.6), better prompt-injection resistance. Pricing unchanged at $3/$15.
On 16 February 2026 Alibaba opened Qwen 3.5, starting with the 397B-A17B flagship: sparse MoE with 512 experts of which 11 are active, Gated DeltaNet attention, multimodal input, and 262,144 tokens of native context (around 1M with YaRN). Apache 2.0 and 201 languages; first among open-weight models on instruction following and multilinguality, still behind on reasoning.
Anthropic updates Sonnet to 4.7: focused on agent reliability over long tasks, better tool use, tighter integration with Claude Code and the Claude Agent SDK.
Anthropic releases Opus 4.6: first Opus with 1M-token context in beta, agent teams in Claude Code, leadership on Terminal-Bench 2.0 and Humanity's Last Exam. Pricing unchanged at $5/$25.
Mistral releases Voxtral Transcribe 2: two open-source STT models (Batch + Realtime, 4B params) with latency configurable down to 200ms, Apache 2.0, 13 languages.
DeepSeek ships R2, successor to R1: more efficient step-by-step reasoning, open weights, contained training cost. Fresh pressure on closed reasoning models.
OpenAI launches Operator in GA across 30+ countries: an agent that browses the web, fills forms, books appointments, and shops online autonomously on behalf of the user.
Microsoft launches the agent marketplace for Copilot Studio with over 1800 pre-built agents, autonomous multi-agent orchestration, and enterprise governance with DLP policies.
Janus Pro is a unified 7B-parameter multimodal model that both understands images and generates them from text, outperforming DALL-E 3 and Stable Diffusion 3 on the GenEval benchmark. Fully open source and runs locally.
Google DeepMind announces Gemini 3 with Pro and Flash variants: improved reasoning, native long context, deeper integration into Workspace and Android.
Google releases Veo 3.1 and Veo 3.1 Lite: video generation with "ingredients" (multiple reference images for character/scene consistency), 1080p/4K output, vertical format for Shorts. Veo 3.1 Lite is the cost-effective variant.
Anthropic ships Cowork as a research preview: a desktop agent with sandboxed shell and local file access, aimed at people who don't live in the terminal the way Claude Code users do.
Alibaba releases Qwen2.5-VL 72B under Apache 2.0, surpassing GPT-4o on multiple multimodal benchmarks with support for documents, charts, 20+ minute videos, multilingual OCR, and GUI agent actions.
OpenAI releases o3-mini on January 7 2026, the smallest and cheapest model in the o3 reasoning family, featuring three configurable effort levels and benchmark performance that surpasses o1 at significantly lower cost.
Local inference dominated CES 2026: Snapdragon X2 Plus at 80 TOPS, Ryzen AI 400 with a 60 TOPS NPU and RDNA 3.5 graphics, Intel Core Ultra series 3 on the 18A node, and AMD Ryzen AI Halo targeting models up to 200B in 128 GB. NVIDIA instead used the show for the Rubin platform, now in production, and Alpamayo for autonomous driving.