Skip to content
Source · Daily Brief

AI Daily Brief — 17 July 2025

One of the loudest agent days of the year. OpenAI launched ChatGPT Agent, fusing Operator + Deep Research + ChatGPT into a single mode with its own virtual computer. Mistral answered the same day with a sweeping Le Chat update — Deep Research + Voxtral voice + image editing + Projects — explicitly targeting OpenAI and Google in enterprise. Perplexity raised $100M at $18B. Fortune broke that Grok 4 shipped with no system card.

Top stories

  • OpenAI launches ChatGPT Agent. General-purpose agentic mode that uses its own virtual computer to browse, click, fill forms, run code and edit spreadsheets. Merges Operator and Deep Research with ChatGPT’s conversational interface. Pro (400 messages/month), Plus and Team (40 messages/month); Enterprise and Edu to follow. New SOTA of 68.9% on BrowseComp, 17.4 points above Deep Research. via OpenAI
  • Mistral brings Deep Research, Voxtral voice and image editing to Le Chat. Deep Research agent that plans, searches and writes structured sourced reports; native multilingual reasoning; advanced image editing; Projects; voice mode powered by new Voxtral speech model. Free, Pro, Team and Enterprise tiers — framed as a challenge to OpenAI and Google. via Mistral
  • Voxtral hits Hugging Face: open-source speech in 24B and 3B. Voxtral-Small-24B-2507 and Voxtral-Mini-3B-2507 under Apache 2.0. 32K context, up to 30 min transcription / 40 min understanding, native function-calling from voice. Mistral claims it beats Whisper large-v3, GPT-4o mini Transcribe and Gemini 2.5 Flash on all tasks. via Hugging Face
  • Perplexity raises $100M at $18B valuation. Extension round up from $14B a few months earlier. Lands a week after Perplexity unveiled its agentic Comet browser to Max subscribers. via Bloomberg
  • Fortune: xAI’s Grok 4 ships without a safety report. No system card — the industry-standard safety/transparency document — drawing public criticism from OpenAI and Anthropic researchers. Anthropic’s Samuel Marks called it “reckless”. xAI’s Dan Hendrycks said internal “dangerous capability evaluations” were run but did not publish results. via Fortune

Who shipped

OpenAI shipped ChatGPT Agent. Mistral shipped Le Chat Deep Research + Voxtral. Perplexity shipped capital.

Open-source pulse

Voxtral is the new state-of-the-art open speech-understanding stack — multilingual, long-context, Apache-licensed — and the first credible open challenger to closed Whisper-class systems for transcription, summarization and voice-triggered function-calling.

Money, infra & hardware

ICML 2025 main conference days concluded in Vancouver — 3,339 papers accepted overall, with strong representation from Apple, Capital One, Google and Two Sigma research teams in LLM alignment, scaling and agent research.

Quiet corners

SiliconANGLE framed the Le Chat upgrade as Mistral’s pivot from raw model lab to a full ChatGPT/Gemini-style consumer-and-enterprise product, with European data residency as the wedge. MacRumors highlighted concrete ChatGPT Agent use cases — brief upcoming client meetings, plan and buy ingredients, analyze competitors and assemble slide decks.

By the numbers

  • 68.9% / +17.4 — ChatGPT Agent on BrowseComp / point gain over Deep Research
  • 400 / 40 — Pro / Plus/Team monthly Agent message allowances
  • 24B + 3B — Voxtral parameter sizes (both Apache 2.0)
  • $100M / $18B — Perplexity round / valuation
  • 3,339 — ICML 2025 accepted papers

Compiled by AI Feed’s editor from verified web sources for 17 July 2025.