Anthropic ships Opus 4.7 as OpenAI turns Codex into a computer-using agent

The week showed that the AI race has again shifted from chatbots to full-fledged working agents. Anthropic strengthened its flagship with Opus 4.7 and launched a new design product on top of it, OpenAI taught Codex to see the screen and control a computer, and Google and Baidu widened the set of available tools for voice and images. Against this backdrop, it is increasingly visible how AI is moving out of labs and into everyday work, the software market and even stock-market speculation.

The main release of the week is Claude Opus 4.7. Anthropic calls it the company's strongest public release for long coding pipelines, multi-step reasoning and agentic tasks. On SWE-bench Pro, the score rose from 53.4% to 64.3% compared with Opus 4.6. The model handles vision better and reads denser screenshots and diagrams at resolutions up to 3.75 megapixels. A new reasoning level, xhigh, was added for complex scenarios, and it is now the default in Claude Code. The company also introduced Task Budgets in beta to cap token spend per task, and improved memory in long multi-session scenarios. API pricing has not changed, but the new tokenizer may use up to 35% more tokens on the same text.

In parallel, Anthropic launched Claude Design, a separate tool that turns a brief into several variants of interfaces, landing pages and presentations, then exports to Canva, PDF, PPTX and HTML.

OpenAI answered not with a new model but with a new operating mode for Codex. After the update, the agent gets computer use: it sees the screen, moves the cursor and performs actions through the graphical interface. At launch the feature is available on macOS, and several agents can work in parallel without taking over the user's focus. The product also gained a built-in browser for working with localhost, image generation directly in the task flow, memory across sessions, and more than 90 integrations with popular services such as Jira, GitLab, Microsoft 365, Notion and Slack. Scheduled automations were added too, in which the agent gathers context itself and prepares the day's tasks.

This is an important shift: competition is less and less about the quality of chat answers and more about how deeply a model can embed itself in a real working environment.

Google and Baidu strengthened the application layer the same week. Gemini 3.1 Flash TTS supports more than 70 languages, 30 preset voices and over 200 audio tags that can be inserted directly into text so the model whispers, coughs or changes intonation according to the script. Google offers a preview through AI Studio and Vertex AI and automatically marks the output with a SynthID watermark. Baidu, for its part, open-sourced ERNIE Image, an image generator with 8 billion parameters that can run on 24 GB of video memory. The model's strength is said to be text rendering inside images, from posters to interfaces and storyboards.

But almost at the same time, a researcher demonstrated reverse-SynthID, an open method for removing Google's invisible mark from generated images with a claimed accuracy of about 91%. It is a telling fork in the road: companies are making generation more accessible quickly, but the safeguards around AI content remain fragile.

The week's most telling stories came not from labs. The shoe brand Allbirds sold its assets and its name for roughly $39 million, then announced a pivot to GPU-as-a-Service under the new name NewBird AI, and amid the frenzy its shares rose more than sixfold in a day. At the same time, an AI cover of "Sedaya Noch" with a synthetic voice of a Western superstar reached the top of the global Shazam chart and immediately turned into a dispute over rights, authorship and royalties.

Another case is "bixonimania", invented by a Swedish researcher: a fictitious diagnosis that convinced not only several popular LLMs but also the authors of a real scientific publication, who cited the fake preprint as a genuine source.

Such stories hit harder than any benchmark: they show how AI is intruding into music, finance, media and even scientific citation.

The takeaway for the week is simple: the AI market is entering a phase where value is defined not by abstract claims that a model has gotten smarter, but by its ability to take over a piece of real work, act inside interfaces and affect processes beyond the chat. The more confidently agents move from text to action, the more important oversight, fact-checking, rights to outputs and the robustness of safeguards become. And here the coming months will matter no less than the next benchmark race.