Claude Duet opens in preview, Mistral ships Voxtral 3, DeepSeek open-sources DeepInfer, OpenAI adds a 12-hour prompt cache

Claude Duet is a separate app where Claude reaches out first. Every morning it sends a short brief on what matters across your connected apps (Gmail, Slack, QuickBooks and others). Tasks you hand it keep running on its own virtual machine and browser, so your laptop can be off. It runs on Opus 5.5 and pulls in the memory, skills and connectors from your existing Claude account. It's on iOS, Android and claude.ai.

Before it sends a message, pays, signs or deletes anything, it shows you exactly what it will do and waits for approval. Trading securities and moving money between accounts are off-limits even if you approve. Credentials follow the Managed Agents design, so it never sees your passwords, and a dedicated classifier blocks card numbers and government IDs from leaking. If you're building an assistant-style agent, the safeguards are worth reading as a reference design. Note that its background checks, which run every few hours, count against your plan's usage limits.

Voxtral 3 is a family of speech models: three transcription (STT) models (Small Transcribe 3, the 4B Mini Transcribe 3, and Mini Transcribe Realtime 3) and one text-to-speech model, Mini TTS 3. A voice agent SDK built on LiveKit ties listening, reasoning and speaking into one stream. Transcription covers 50+ languages, with speaker diarization and hints for domain terms.

It's cheap: batch transcription runs $0.0035 to $0.004 per minute, realtime $0.008 per minute, and TTS $13 per million characters. Mistral says Small Transcribe 3 beats Whisper, Scribe v2 and Gemini 3.5 Transcribe on multilingual FLEURSVendor claim, with realtime latency under 500msVendor claim. If speech costs dominate your product, as in call centers or meeting notes, it's worth a comparison. EU endpoints and on-prem deployment help teams with data residency rules.

DeepInfer is the inference framework DeepSeek runs behind its own API. It started from vLLM and was heavily reworked, adding DeepSeek's own kernels (FlashMLA, DeepGEMM, DeepEP), expert parallelism (spreading an MoE model's experts across GPUs), structured outputs and speculative decoding. For now it supports only DeepSeek-V4.1-Flash and runs on NVIDIA SM100 (Blackwell) and Huawei Ascend 950.

export DEEPINFER_ACCELERATOR=nvidia
./build.sh
python examples/offline_inference.py

You get to read, in code, how a large MoE model is served at production scale. DeepSeek says a single cluster handles 155 trillion tokens a dayVendor claim. KV cache management, prefill/decode disaggregation and load balancing across replicas are left out, though, and Hopper and older GPUs aren't supported. Treat it as a design reference for teams running their own inference stack rather than something to drop into production.

In the OpenAI Responses API you can now pick a 12-hour minimum lifetime for prompt caches. On GPT-5.6 and later, set prompt_cache_options.ttl to "12h"; the default is "30m". You can mix both in one request, for example 12 hours for shared instructions and 30 minutes for recent turns.

Agents that run all day often go hours between requests, so their caches kept expiring. A 12-hour cache write costs 2.5x normal input (1.25x for 30 minutes), but reads cost 0.1x (0.05x on GPT-6.1 Sol). If long instructions and tool definitions get reused several times a day, you save. Put a 12-hour TTL on content you only reuse once or twice and you lose money.