Category
The Qwen-Image-2.1 model from Alibaba natively generates and edits transparent images, creates images in 2K and processes up to ten reference images. Its weights are available, and it also runs on an NVIDIA 3090 GPU. Commercial use requires a separate license.
Cloudflare has released the Clef and Clef-flash models, which assign probabilities to predefined options. They support both text and images and are available on the Workers AI platform as well as under the Apache-2.0 license on the Hugging Face platform.
The AstaBrief 8B model generates scientific reports with citations and is available in the Fast mode of the Asta service. The authors are releasing both the weights and the training data. According to their measurements, the entire process takes an average of 51.1 seconds compared to 178.5 seconds in Thinking mode.
The company Black Forest Labs has released the Flux 3 Image model. According to the company, it handles multi-step edits without changing other parts of the image. It supports up to ten reference images and output of up to 4K. A free demo is available.
The company Tavus has introduced the Griffin model for live audiovisual conversations. According to its study, 48% of participants considered it human after a one-minute call. The Griffin-Lite model is available to selected testers; a more powerful version is to follow after safety questions have been resolved.
The Ideogram 4.5 model is available on the Ideogram platform and via API. According to Ideogram, it changes only the requested parts of an image. It offers four quality tiers at native 2K resolution for 0.8 to 22 cents per image.
According to an article by Živě.cz, the unreleased Gemini 4 Argon model by Google reached first place in the Arena.ai leaderboard. The article states a significant lead over the competition, but does not include a specific score.
NVIDIA has released the open foundation model Kumo Tabular (28–215M parameters) for classification and regression on tabular data. According to the company, the model predicts without training, tuning, or feature engineering, and leads on the TabArena, BeyondArena, TALENT, and ScoringBench benchmarks.
The research paper introduces InfiMed2 – medical multimodal foundation models in versions with 4B and 27B parameters. According to the authors, the 4B model achieves 66.73% accuracy after RLVR and outperforms the larger Qwen3.5-9B, while the 27B model achieves 73.72 % on five benchmarks.
OpenAI has made GPT-6.1 Sol generally available on Amazon Bedrock with enterprise data security controls. The model was released in late September as a cheaper replacement for the postponed GPT-6.1 Astra, with coding performance that, according to OpenAI, approaches that flagship version at one-fifth of the price.
New: GPT-6.1 Sol is available on Amazon Bedrock; It matches the performance of GPT-6 Astra on the DeepSWE v1.1 test set; On DeepSWE v1.1, it outperforms GPT-6 Sol by 6.4 % with lower reasoning effort; Availability is expanding to AWS infrastructure through a third-party provider
The company xAI has introduced Grok 4.7 on Amazon Bedrock: a context window of 500 000 tokens and four reasoning effort levels. According to Artificial Analysis, performance on agentic tasks and knowledge work improves compared with Grok 4.6, but at roughly twice the token consumption.
The Claude Sonnet 5.5 model is available on Amazon Bedrock and in Claude Platform on AWS – according to AWS, a cheaper and faster option for well-defined tasks such as bug fixes, SQL generation, or UI testing.
The new family of open-source agentic models Holo4 (27B dense, 35B-A3B MoE) handles GUI control, writing and running code, MCP, and API all in one model. On OSWorld 2.0, Holo4-27B achieves 61.7% compared to 81.8% for the Opus 5.5 model, but at significantly lower cost.
The Claude Sonnet 5.5 model has been released with the identifier claude-sonnet-5-5. It is available via the Claude API as well as in the offerings of Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry.
Boris Power, head of applied research at OpenAI, said that 80 to 90 percent of the company's research targets GPT-7, GPT-8 and further generations, because real value is created in the jump between generations, not in incremental versions such as GPT-5.1 to 5.2.
The Czech company BottleCap AI has released the second model in the ThinkingCap series, a modified version of Alibaba's Qwen 3.8-27B model. According to the manufacturer, it consumes 37% fewer thinking tokens with slightly higher accuracy, and it is free on Hugging Face.
From 29. 9. 2026, OpenAI is reopening the Pro plan at $200/month to new users and removing the 5-hour usage limit, but at the same time halving API credits per dollar. The company is thus moving from subsidized subscriptions toward payment based on actual usage. This follows the price reduction for the GPT-6 Sol and Luna models on 22. 9.
New: API credits per dollar halved; Pro plan at $200/month reopened to new users; 5-hour window for the weekly allotment removed; Strategic shift from subsidized subscription models to pay-per-use pricing; Microsoft is making similar changes with Copilot
Anthropic engineer Jackson Kernion explained why the writing of Claude models has gotten worse despite growing performance in math and code — according to him, the models have learned to write for other AI models, not for humans.
NVIDIA has published additional technical details on the open-weight model Nemotron 3 Diarization: according to the company, it is 41% more accurate than its predecessor and leads the Diarization-Bench leaderboard with an error rate of 14.72% (DER).
New: 41% better error rate than Streaming Sortformer (specific percentage); The model can be combined with Parakeet to create transcripts with speaker tags; Audio buffer adjustable to four levels (30.4–0.32 seconds); Error rate increases with more participants, strong noise, and hares of sound; The Voice Arena benchmark is strict – it counts overlapping speech and misalignments
According to its own announcement, OpenAI has improved prompt caching in the GPT-6 model - a higher cache hit rate, new diagnostics, explicit breakpoints in prompts, and controls to reduce latency and costs.
AI Radar monitors Czech and international sources every day, looking for changes that truly deserve attention.
The result is a quick overview of what has actually changed in the AI world, rather than another stream of articles.
Use the CS/EN switch to read the same Radar in Czech or English. English content is published after its translation has been checked, so new and older items may appear later.
Each row represents one event — not one article. At a glance, you can see its significance, credibility and main point.
Illustrative example, not a current news item.
Agent mode is available to all paying users
Until now, the mode was available only on the highest plan; it is now available on all paid tiers without a waitlist.
The detail page contains a fuller summary, its significance and original sources. Practical impact appears in the detail and the For individuals and For businesses views. An AI Flash item reaches the main selection only after it has been expanded and meets the publication rules.
We first summarize each event in the same way for everyone. Based on those same facts, we then explain what the change means for your own use and what it could mean for how a company operates.
Today’s briefing is the same for everyone. Pages For individuals and For businesses can be found in the main navigation — they select only events relevant to the given use case.
Reset history · Outage history · Subscribe in a feed reader (RSS) · Methodology and label explanations
