Blog

Requesty combines Frankfurt routing, zero retention, a DPA, and EU-region model deployments behind one API. That is a strong foundation for an EU LLM stack, provided you lock down region, models, fallbacks, logging, and your own workflow.

Kimi K3 targets the same territory as Claude Fable 5 and GPT‑5.6 Sol: 1M context, long-running agents, coding, and knowledge work. It is not a finished replacement, but its planned open weights make it worth a measured pilot.

Google DeepMind argues that agents may make hypothesis generation cheap, while validation remains slow and expensive. The same pattern applies inside companies: more AI output is only useful if you have review, audit, and decision capacity.

OpenAI is bringing Chat, Work, Codex, and Sites together, while Anthropic pushes Claude Cowork across desktop, web, and mobile. This is not just a UI change. AI is moving from chat into a work layer over documents, apps, files, and approvals.

Claude Code agent teams, /goal, hooks, Remote Control, and Slack show that agentic development is not only about the model. A new Microsoft rollout study adds the missing operational layer: adoption, visible use, measurement, and review.

Plaud is no longer just an AI recorder with transcription. Through MCP and CLI, recordings, transcripts, and AI summaries can flow into Claude, Codex, ChatGPT, or terminal workflows.

GPT-5.6 is no longer only a limited Sol preview. OpenAI has described Sol, Terra, Luna, ChatGPT Work, Programmatic Tool Calling, and ultra mode for parallel agentic work.

Fable 5 is available again after the June suspension. For companies, that does not mean using it everywhere. It belongs in a model router as an expensive specialist with budget control, fallback, and audit.

Meta has opened Muse Spark 1.1 through a paid Model API. The important part is not the benchmark story, but the combination of aggressive pricing, agentic positioning, custom chips, and possible cloud infrastructure.

Cloudflare is showing what a web for AI agents could look like: crawlers authenticate, follow site rules, and can pay programmatically for access.

Microsoft Scout shows where Microsoft 365 agents are heading. OpenClaw and ClawHub also show that the biggest risk is not chat, but the mix of skills, shell access, browser automation, and company data.

OpenAI is coming to Amazon Bedrock and is now also showing Jalapeño, its own inference chip. For companies, this is a signal that AI is moving into cloud governance and deeper infrastructure for cost, latency and capacity.

Anthropic published the first Project Glasswing update on May 22, 2026. Claude Mythos Preview helped partners find more than ten thousand high- or critical-severity vulnerabilities and reported 23,019 findings in open-source projects. The important shift is not only the number, but the new bottleneck: verification, disclosure and patching. Aktualizováno o OpenAI Daybreak, GPT-5.5-Cyber a Codex Security.

Z.ai has released GLM-5.2, an open model with 1M context, 128K output, and a focus on long coding-agent tasks. For companies, the important signal is that open models are moving from chat into project-scale engineering work.

Higgsfield expands MCP, Supercomputer and Marketing Studio with a more practical production layer: plugins directly inside Adobe Premiere Pro, After Effects and now Photoshop.

NotebookLM is not just another chatbot. It is an AI notebook over your own sources: PDFs, websites, Google documents, YouTube videos or notes. It is useful for research, company documentation and students who want to learn actively with citations.

Google DeepMind is showing an experimental text diffusion model. For companies, the point is not immediate deployment, but the direction: faster intermediate steps, editing, coding, and lower-latency agent workflows.

Google I/O 2026 was not only about Gemini 3.5 Flash. The bigger story is the stack around it: Gemini Omni, Spark, AI Mode in Search, Workspace Live, Antigravity 2.0, Universal Cart and SynthID show Google moving AI from chat into everyday infrastructure.

Apple gives developers a native path to on-device and Private Cloud Compute models. For companies, the practical value is privacy, latency, structured output, and app-level actions.

Anthropic updated its Privacy Policy for Claude Free, Pro, and Max. The practical impact is strongest in companies where people use personal Claude accounts, Claude Code, and connected apps outside managed work environments.

Google released Gemma 4 12B, a local multimodal model for text, images, audio, and video. The key idea is an encoder-free approach and easier deployment on your own hardware.

OpenAI is expanding Codex with Sites, role-specific plugins, and team workflows. This is less about generating a page and more about creating useful internal work artifacts.

Anthropic released Claude Opus 4.8. A practical look at agents, coding, research, 1M context, fast mode, benchmarks, and pricing.

OpenAI released GPT-5.5 to paid plans. It promises significantly better coding, computer use and deeper research. The model is another step toward a unified super-app combining ChatGPT, Codex and the browser.

Datacurve published DeepSWE on May 26, 2026: a benchmark for long-horizon coding agents. The important part is not just the leaderboard, but the measurement method: original tasks, larger repository changes, and behavioral verification.

Google released Gemini 3.5 Flash at I/O 2026. Independent benchmarks show excellent speed and agentic performance, but also a higher cost than the older Flash model and weaker spots in hard coding, instruction following and sensitive decisions.

At Google I/O 2026, Google introduced Gemini Spark: a personal AI agent designed to run in the cloud, work across Gmail, documents, calendar and tasks, and keep working outside an open chat window. That is a bigger shift than another chatbot.

AI s rozumem showed a practical Claude plugin for meal planning connected to Rohlik through MCP. The interesting part is not food planning itself. It is the pattern: the plugin keeps the workflow, MCP exposes a tool, and the human stays in the approval loop.

Jirka Herník gave six local models and one cloud baseline the same task: write Tetris as a single HTML file. The result is a practical map of where local coding agents already make sense and when a stronger API model is still the better fallback.

The original Claude ladder is useful in subscriptions, but API automation needs a different map: guardrail/PII filters, cheap repeated calls, a main worker, and a top specialist.

On April 23, 2026 Google DeepMind showed Decoupled DiLoCo, a way to train large models across separate compute clusters. For companies, this is less about training their own GPT and more about where AI infrastructure is going.

Xai added custom voices and OpenAI followed on May 7, 2026 with new Realtime API voice models. For companies, voice agents are becoming a product layer with reasoning, tools, translation, transcription, safety and audit.

Xai launched Grok Build, an early beta coding agent for the terminal. The important part is not another CLI tool, but the shift toward agents that plan work, edit code, and show diffs before approval.

On May 14, 2026 OpenAI announced Codex in the ChatGPT mobile app. The important part is not coding on a phone, but unblocking long-running agent work at the exact moment it needs human judgment.

On 14 May 2026, Meta opened developer access to the display in Meta Ray-Ban Display. The important part is not only the HUD itself, but the two build paths: a mobile SDK and Web Apps built with HTML, CSS and JavaScript.

IBM released Granite 4.1, an open model family under Apache 2.0 for text, vision, speech, embeddings and guardrails. For companies, it is most interesting as a practical component for documents, RAG and controlled workflows.

NVIDIA released Nemotron 3 Nano Omni, an open model for text, image, video and audio. For companies, it is interesting as a multimodal sub-agent that can simplify pipelines made from OCR, vision, ASR and LLM steps.

On April 27, 2026 GitHub announced usage-based billing for premium Copilot AI requests. For companies, this is a clear signal: AI coding is no longer only a per-seat subscription, but an operating cost that needs limits and audit.

Anthropic vydal Opus 4.7, OpenAI GPT-5.5, Alibaba Qwen 3.6-Max, DeepSeek V4. Za 14 dnů se v AI stalo víc než za celý loňský kvartál. Co z toho plyne pro vaši firmu?

Anthropic vydal 16. dubna 2026 novou verzi Opus. Lepší kódování, task budgets, 3,75 MP obrázky. Tady je praktický přehled—co tomu svěřit dnes a kde si ještě počkat.

Anthropic a OpenAI vydali nové vlajkové modely v týdnu po sobě. Praktický pohled na to, který se hodí na co—bez marketingu, jen z reálného nasazení.

China’s DeepSeek introduced preview versions of V4 Flash and V4 Pro. They bring a new Hybrid Attention architecture, a 1 million token context window and top-tier coding benchmark scores.

AI nástroje mohou dramaticky zvýšit produktivitu vašeho týmu. Podívejte se na 5 konkrétních způsobů.

Alibaba unveiled Qwen 3.6-Max-Preview—its most capable model so far. Significant gains in coding, reasoning and agentic tasks, plus a large context window.

Anthropic přidal funkci, kterou jsem si v prosinci napsal jako wishlist. Task budgets dají AI agentovi finanční mantinely—a vám klid v noci.

OpenAI introduced ChatGPT Images 2.0 on April 21, 2026. The important part is not only better image quality, but a workflow where AI keeps context, iterates visuals and helps with product, marketing and design drafts.

OpenAI released GPT-5.4 on March 5, 2026 in three variants: Standard, Thinking (reasoning-first) and Pro (maximum capability). The API supports a context window of up to 1.05 million tokens—the largest OpenAI has ever offered commercially.