This brief covers AI news from 2026-09-26 UTC.
Tencent’s biggest open-weight release yet lands the same day Microsoft turns Copilot into a coding and agent platform.
Tencent open-sources Hy4 preview, a 770B MoE under Apache 2.0
What happened: Tencent released weights for Hy4 preview, a 770B-parameter MoE that activates 49B parameters per token and supports a one million token context. It ships under Apache 2.0 in BF16 and FP8 variants, with official deployment examples using 8 GPUs.
Why it matters: A permissively licensed frontier-scale model means teams can self-host or fine-tune instead of renting closed APIs, though LoRA fine-tuning needs at least 64 GPUs with 96GB each.
Microsoft adds code generation and an always-on agent to Copilot
What happened: Microsoft unveiled new Copilot capabilities including a coding tool and an always-on AI agent, and is embedding Word, Excel and PowerPoint directly inside the Copilot app.
Why it matters: Office workers get code generation and persistent agents inside the tool they already use, which changes what internal tooling teams are expected to build themselves.
Cognition crosses $1B annualized revenue on Devin demand
What happened: Cognition AI hit a $1 billion annualized revenue run rate on September 25, up from roughly $492 million in May and about $900 million in early September. The company raised over $2 billion in September at a $48 billion valuation.
Why it matters: The company behind Devin doubling revenue in four months is the clearest signal yet that enterprises are paying for autonomous coding agents, not just autocomplete.
Appeals court lets Pentagon keep Anthropic off-limits
What happened: The D.C. Circuit denied Anthropic’s petitions in a 2-1 ruling, leaving in place the Defense Department’s supply-chain risk designation that bars employees and contractors from using Anthropic products for DOD business. Anthropic can seek en banc review or petition the Supreme Court.
Why it matters: Defense contractors building on Claude for government work now have to plan around a designation that has survived appeal, and route around it or wait for further litigation.
LiteLLM launches LiteAgents to swap agent harnesses
What happened: LiteLLM released LiteAgents, an SDK that switches between Deep Agents, Pydantic AI, Claude Agent SDK, Codex and OpenCode by changing one harness field, keeping tools, MCP connections and model config. It follows the Claude Agent SDK query() pattern.
Why it matters: Teams can benchmark harnesses on their own tasks without rewriting the surrounding app, which makes agent-loop choice a config decision instead of an architecture commitment.
Replit acquires Atta to put analysis in chat
What happened: Replit announced it acquired Atta, bringing business analysis and charting into Replit. Interactive charts in Replit chat are available today, letting users ask for visualizations without writing SQL.
Why it matters: Data exploration moves into the same agent surface where the app is built, so teams can go from question to chart without a separate BI tool or analyst handoff.
Google says Gemini 4 is in post-training
What happened: Google DeepMind chief scientist Koray Kavukcuoglu said Gemini 4 has entered post-training, including safety testing and guardrails, and that Google hopes to ship it much earlier than the end of 2026. Internal testing already uses it to power Antigravity, Google’s agentic coding environment.
Why it matters: Developers planning on Antigravity or Gemini APIs should expect a model swap on an unspecified timeline, with no date committed yet.