Model-Release

Breaking news from MiniMax: the company has officially released M2.5, the latest entry in its M2 model family — and the benchmarks are going to raise some eyebrows. SOTA on coding. Twice the speed of its predecessor. And priced so aggressively that the company’s own marketing frames it as “intelligence too cheap to meter.” At $1 per continuous hour of inference at 100 tokens per second, or $0.30 at 50 tokens/sec, MiniMax M2.5 is targeting a very specific pain point for anyone building with agentic AI at scale: cost. ...

The agentic AI landscape just shifted. OpenAI’s GPT-5.4 — launched March 5, 2026 — isn’t just a model update. It’s a direct bid to own the autonomous agent stack, arriving with native computer-use, a one-million-token context window, and a reworked tool-calling system that slashes token consumption by 47% on MCP benchmark tasks. If you’re building with agent pipelines, this is the model release worth paying attention to. What’s Actually New in GPT-5.4 Native Computer-Use This is the headline feature, and it’s genuinely significant. Rather than bolting computer-use on as a post-hoc capability, OpenAI has built it into GPT-5.4 at the architecture level. The model can observe screen states, click UI elements, type into fields, scroll, and navigate applications — autonomously, without requiring a separate vision model or operator middleware. ...

Most multi-agent AI systems are built by developers — frameworks assembled from components, with agents spawned programmatically, each given a role, each calling the others through APIs or queues. It’s architected software. What xAI shipped in mid-February is something structurally different: a model where the multi-agent council isn’t something you build around — it’s something that runs inside every response. Grok 4.20 Beta launched with four named agents — Grok, Harper, Benjamin, and Lucas — that execute a think-then-debate-then-consensus loop as part of the model’s native inference process. For queries below a complexity threshold, users may never notice the agents working. For hard problems, the loop is engaged automatically: agents independently reason about the problem, challenge each other’s conclusions, and surface a synthesized answer. You don’t configure this. It just runs. ...