Skip to content
Source · Daily Brief

AI Morning Brief — 14 August 2026

DeepSeek made the most aggressive simultaneous move of the cycle — shipping its full V4-Pro model, open-sourcing its agent framework under an MIT licence, and raising API prices, all in one announcement. Elsewhere, OpenAI previewed 14× faster inference via Cerebras, xAI launched Grok 4.6 with strong benchmark claims, and Google shipped yet another Flash revision.

Top stories

  • DeepSeek: new model, new framework, higher prices. DeepSeek released the full V4-Pro-0813 (out of preview), simultaneously open-sourced DeepSeek Harness v0.1 — a MIT-licensed agent framework built on the Cordis plugin meta-framework — and raised API cache-hit prices by roughly 6×. Community reaction to the model’s benchmark performance versus V4-Flash was mixed. via THE DECODER
  • OpenAI previews Ultrafast mode. A new API service tier runs GPT-5.6 Sol at up to 14× normal speed — reaching up to 750 output tokens per second — powered by Cerebras hardware. Rolling out initially to a limited group of enterprise customers with broader access planned. via TechCrunch – AI
  • Grok 4.6 launches with benchmark leadership claims. xAI’s new release reportedly scores 95% on GPQA Diamond and matches Claude Fable 5 on Perplexity’s WANDR benchmark at over 60% lower cost. The model is now available inside Perplexity and Warp. via X · @perplexity_ai
  • Gemini 3.7 Flash arrives three weeks after its predecessor. Google shipped a new workhorse model for coding and agentic workflows at half the price of Gemini 3.6 Flash, claiming improvements over competing models on several coding benchmarks. Gemini 3.7 Flash is already available inside GitHub Copilot. via THE DECODER
  • OpenAI’s leadership reshuffle continues. Chief Revenue Officer Denise Dresser is departing in the coming weeks; Dali Rajic, previously president of Wiz, is appointed as her replacement — the second senior executive exit in a single week. via The Verge – AI
  • Anthropic research finds multi-agent systems clash and collude. Researchers placed multiple AI agents on the same task and observed unexpected confrontation, coordination, and collusion, raising questions about whether existing safety evaluations capture multi-agent risks. via TechCrunch – AI

Who shipped

DeepSeek dominated discussion with a coordinated triple release — model, framework, and revised pricing — though the model’s relative benchmark standing versus V4-Flash drew scepticism. OpenAI had the broadest footprint: Ultrafast inference, ChatGPT Computer History (activity awareness across desktop apps on Mac), a new CRO hire, and an IBM consulting partnership. Google DeepMind shipped Gemini 3.7 Flash and a sign-language-to-text model (SL2T). xAI shipped Grok 4.6 and introduced Grok Bot, an early-beta autonomous agent that signs into external tools and returns completed work.

Open-source pulse

DeepSeek Harness v0.1 is the most significant open-source release of the cycle — an MIT-licensed agent harness built around a plugin architecture (Cordis), released alongside the new V4-Pro model. MiniMax-Music3 also dropped as open weights: an 8B LLM combined with a 2.7B DiT for full-song generation from lyrics and prompts, with a Hugging Face demo available. Suno released Studio 2.0 with MIDI import and a conversational interface, moving its platform closer to a full digital audio workstation.

Money, infra & hardware

Databricks closed a $5B round at a $190B valuation after investor demand pushed well past the company’s initial $1B target. The OpenAI Foundation announced a $100M commitment to an AI for Civil Society and Philanthropy programme targeting nonprofits and community organisations. Cerebras is the inference provider powering OpenAI’s Ultrafast tier.

Quiet corners

Meta was entirely absent — no Llama releases, no product announcements, and no significant community activity during the window. Cohere did not appear at all; Mistral surfaced only in a brief Hacker News link to an OCR 4.1 documentation update.

By the numbers

  • ~300 stories in 24 h across 25+ sources
  • Most-mentioned model: Grok
  • Most-mentioned lab: OpenAI
  • Notable absences: Meta, Cohere, Mistral

Compiled by AI Feed’s editor from all ~300 headlines published on 13–14 August 2026.