Anthropic released Opus 4.7, and OpenAI turned Codex into a computer work agent

This week showed that the AI race has shifted again, from chatbots to full-fledged work agents. Anthropic strengthened its flagship Opus 4.7 and immediately created a new design product for it, OpenAI taught Codex to see the screen and control the computer, and Google and Baidu expanded their toolset for voice and images.

Against this backdrop, it's particularly striking how AI is leaving laboratories and entering everyday work, the software market, and even stock market speculation. The main release of the week is Claude Opus 4.7.

Anthropic calls the model the company's strongest public release to date for long code pipelines, multi-step reasoning, and agentic tasks. On SWE-bench Pro, the result rose from 53.4% to 64.

3% compared to Opus 4.6. The model works better with vision and reads denser screenshots and diagrams with resolution up to 3.

75 megapixels. For complex scenarios, a new level of reasoning, xhigh, has appeared; in Claude Code it's now enabled by default. Additionally, the company introduced Task Budgets in beta to limit token spending per task and improved memory in long multi-session scenarios.

API pricing remained unchanged, but the new tokenizer can consume up to 35% more tokens on the same text. In parallel, Anthropic launched Claude Design — a separate tool that transforms a brief into several interface, landing page, and presentation variants, then provides export to Canva, PDF, PPTX, and HTML. OpenAI responded not with a new model, but with a new operating mode for Codex.

After the update, the agent gained computer use: it sees the screen, moves the cursor, and performs actions through the graphical interface. At launch, the feature is available on macOS, with multiple agents able to work in parallel without intercepting user focus. Within the product, an embedded browser appeared for working with localhost, image generation directly in the task flow, memory between sessions, and over 90 integrations with popular services like Jira, GitLab, Microsoft 365, Notion, and Slack.

Scheduled automations were also added, where the agent itself raises context and prepares tasks for the day. This is an important shift: competition is increasingly less about chat response quality and increasingly about how deeply a model can embed itself into the real work environment. Google and Baidu strengthened their applied layer this same week.

Gemini 3.1 Flash TTS supports over 70 languages, 30 ready-made voices, and over 200 audio tags that can be inserted directly into text so the model can whisper, cough, or change intonation according to script. Google offers preview through AI Studio and Vertex AI and automatically marks the result with a SynthID watermark.

Baidu, in turn, opened ERNIE Image — an image generator with 8 billion parameters that can run on 24 GB of video memory. The model's strong point is rendering text within images: from posters to interfaces and storyboards. But almost simultaneously, a researcher demonstrated reverse-SynthID — an open way to remove Google's invisible watermark from generated images with claimed accuracy around 91%.

It turns out to be an illustrative fork in the road: companies are quickly making generation more accessible, but protective mechanisms around AI content remain fragile. The most telling stories of the week came not from laboratories. Shoe brand Allbirds sold its assets and name for roughly $39 million, then announced a pivot to GPU-as-a-Service under the new name NewBird AI, and on a wave of frenzy saw its stock rise more than sixfold in a day.

Simultaneously, an AI cover of "Grey Night" 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 — the fictional "bisonmania" invented by a Swedish researcher: a fabricated diagnosis 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 interferes in music, finance, media, and even scientific citation.

This week's conclusion is simple: the AI market is entering a phase where value is determined not by abstract claims that a model has become smarter, but by its ability to take on a piece of real work, act within interfaces, and influence processes beyond the chat. The more confidently agents transition from text to action, the more important control, fact-checking, rights to output, and the robustness of protective mechanisms become. And it's precisely here that the coming months will be no less important than the next benchmark race.