Claude Enterprise Gets Smart Reports: Who Uses AI How, and What It Costs

Anthropic has added smart reports to Claude Enterprise: built-in analytics that show how a company's employees use Claude, what they get out of it, and how much they pay for it. The feature is available on paid Claude Enterprise plans. Several other notable stories landed the same day: Chinese labs are learning from GPT-6 Astra's answers, Amp launched cloud Macs for AI agents, Mistral opened a Responses API, and ElevenLabs released Cue, a model for precise audio-to-text synchronization.

Smart reports in

Claude Enterprise: analytics instead of guesswork

Companies that hand out AI assistants to employees usually end up asking the same question: is it paying off? The new reports answer it with data. The system draws on work history from three sources: regular chats, Claude Code (a coding agent) and Cowork (a mode for office tasks). The reporting window is up to 28 days.

Claude then reads the conversation transcripts and sorts sessions by project and by type of output: code, documentation and so on. Here is what the administrator gets:

• usage charts and spending-limit controls;
• detection of bottlenecks where work gets stuck;
• a list of sessions that ended without a result;
• lists of requests the system classified as personal or unrelated to work;
• the ability to ask questions about the statistics and get recommendations.

Organizations get 10 free reports per month; after that, generating a report costs extra. This is convenient for managers, but there is a downside: the report is built by reading the conversations themselves. Companies should tell employees in advance exactly what gets reviewed.

Chinese models learn from GPT-6 Astra

According to The Information, the gains of the updated DeepSeek V4.1-Flash-1017 and Kimi K3.1 Pro are directly tied to OpenAI's GPT-6 Astra. Both companies trained their systems on large sets of Astra's answers. The technique is called distillation: a weaker model trains on the outputs of a stronger one and picks up some of its skills.

The effect shows up in the Browser Use benchmark, where models perform tasks on the web autonomously. DeepSeek's score rose from 70.9 to 75.0 points; Astra itself scores 77.9. Claude used to be the main source of knowledge. But Anthropic tightened its defenses, making it harder to extract hidden reasoning chains, so the labs switched to Astra. DeepSeek's creators had actually wanted to avoid distillation altogether in the fourth generation, but changed their minds under market pressure after seeing the success of competitors using GLM.

New releases for developers and voice tasks

Amp: cloud Mac environments. Amp did not move entire working environments to Mac: Docker starts slowly there, and storage is expensive. So the setup is mixed. The development server and build tools stay in Linux containers. When an agent needs Xcode or iPhone simulators, a clean cloud Mac spins up in half a minute.

• Code syncs between Linux and macOS, and heavy processes are split across the machines.
• The agent builds the project itself, checks layout on different devices, and works with the interface.
• The Mac stays alive for 20 minutes after a task to preserve the build cache, then is deleted.
• The service is in beta with a waitlist, and rental starts at $2 per hour.

Mistral AI: Responses API. The interface lets developers build autonomous systems on Mistral's chat models and inspect each of their actions. It returns the final answer, the reasoning chain and function calls separately, so you can see which step triggered an operation.

• The format is borrowed from OpenAI, so integrations port over with almost no changes.
• The client still maintains the conversation history, so the full context must be resent with every request.
• It supports streaming generation, JSON output, custom functions, and multimodal prompts with text, images and documents.
• It works with all current models, including Mistral Large 4.
• Mistral's cloud in the EU and US has added DeepSeek V4.1 Flash and GLM 5.3 Flash.

ElevenLabs: the Cue model. Cue tackles forced alignment: pinpointing the exact millisecond at which each character of a finished text is spoken. Older tools cut the recording into fragments, and timing errors accumulate. Cue processes a recording of any length in one pass.

• A single architecture handles more than 90 languages.
• According to ElevenLabs' measurements, Cue outperforms open alternatives such as Qwen3-ForcedAligner.
• The model works character by character and stays accurate with noise, overlapping voices, and songs where vocals stretch out words.
• On difficult English-language recordings, the average timing error is 57 milliseconds.
• The algorithm is built into the ElevenLabs Scribe transcription pipeline and is open to third-party projects through the Forced Alignment API.

What this means

For business, the day's main news is that AI spending can now be reviewed like any other budget line. Anthropic's analytics help find where money is going to waste. For engineers, the other items are more useful: Amp removes the iOS build problem for agents, Mistral makes migrating from the OpenAI API easier, and Cue delivers precise subtitles and karaoke without manual adjustment. The distillation story is a reminder that frontier models quickly become teachers for their competitors, and protecting model outputs is turning into an engineering task of its own.