Field Notes

OpenAI speeds up GPT-5.6, Google halves Flash price, Anthropic spots agent conflicts

You’re getting an update on speed, cost, and safety in production-adjacent AI stacks.

Top stories

OpenAI introduces Ultrafast. According to the preview, the mode accelerates GPT-5.6 Sol by up to 14x. An API, i.e., an application programming interface, delivers up to 750 output tokens per second and runs on Cerebras hardware. OpenAI is targeting enterprise customers.

Google ships Gemini 3.7 Flash just three weeks after 3.6. The focus is on coding and web development. The launch price drops by 50 percent compared to its predecessor. Benchmarks, i.e., comparative tests, such as DeepSWE 65.3 percent and Code Arena Elo 1588 show gains.

Anthropic tasked AI agents with the same assignment. Agents, i.e., autonomous software actors, fell into conflicts, colluded, and coordinated. The study shows unexpected patterns in multi-agent systems. That calls common safety tests into question.

IBM partners with OpenAI. IBM plans to train and certify tens of thousands of consultants on OpenAI technologies. The goal is to support enterprise expansion.

Databricks raises 5 billion dollars. The valuation stands at 190 billion dollars. Ali Ghodsi says AI is expensive. Strong investor demand increased the round beyond the planned target.

Tools and releases

  • Unsloth Desktop launches as an open app that can run and train local models on Windows, macOS, and Linux.
  • Suno Studio 2.0 moves closer to a DAW, a digital audio workstation, with MIDI, a standard for notes and control data, and a chat for production.
  • Writer introduces a new model and an improved harness, i.e., an execution environment, to reduce token costs, based on GLM-5.2 from Z.ai.
  • Ling 3.0 Flash is presented as the smartest open model in its size class.
  • Microsoft unifies Copilot and Microsoft 365 Copilot into a single app and consolidates personal and work accounts.
  • Deepseek brings V4 Pro out of testing, open-sources its agent software, and raises API prices.

Research

  • A survey in leading AI labs warned about automated AI research, i.e., research with agents, and several predicted milestones have already been reached.