March was the month the frontier labs stopped competing on raw benchmark scores and started competing on what you can actually build. Native image generation, models that reason before they answer, the first credible general-purpose agent, and the largest private funding round in history all landed in the same four weeks. For operators, the signal under the noise is clear: the capabilities are converging, the money is consolidating, and the question is no longer “is the model good enough” but “what do we wire it into.”
OpenAI shipped image generation that actually follows instructions
On March 25, OpenAI rolled native image generation into GPT-4o, replacing the DALL-E era with a model that renders legible text, accepts images as input, and stays steerable across a conversation. The jump that mattered wasn’t aesthetic — it was reliability. You can now ask for a diagram with correct labels, a mockup with real copy, or an edited photo, and get something usable on the first pass.
Why it matters for builders: This collapses a whole category of design and content tooling into a single API call. Before you add another Canva-style vendor or a stock-image subscription, ask whether 4o already covers the 80% case. The win is removing tools, not adding them.
Gemini 2.5 Pro made “thinking” the default
Also on March 25, Google released Gemini 2.5 Pro, a reasoning model that works through a problem before responding. It debuted at #1 on LMArena by a wide margin, with strong scores on hard math and science benchmarks and a genuinely large context window for whole-repository and long-document work.
Why it matters for operators: Thinking models trade latency and cost for accuracy. That’s the right trade for analysis, code review, and complex extraction — and the wrong trade for high-volume, latency-sensitive paths. The discipline is routing: use the expensive reasoning model where correctness pays for itself, and a cheaper fast model everywhere else.
Manus put a real general agent in front of users
On March 6, Chinese startup Butterfly Effect launched Manus, billed as a general-purpose autonomous agent that plans and executes multi-step tasks rather than just answering prompts. The demo went viral and the waitlist crashed. The product was uneven in practice, but it was a real preview of where agents are heading.
Why it matters for builders: Agents are moving from demo to default expectation, and the hard part isn’t the model — it’s giving it safe, scoped access to your actual systems. An agent with no reach into your data is theater. An agent with unscoped reach is a liability. Most teams need neither; they need a narrow agent that does one valuable workflow reliably.
The money consolidated around two labs
Two rounds bracketed the month. On March 3, Anthropic raised $3.5B at a $61.5B valuation. On March 31, OpenAI closed a $40B round led by SoftBank at a $300B valuation — the largest private funding round on record, paired with the Stargate compute buildout.
Why it matters: Capital is concentrating in a handful of model providers, which means the foundation layer will keep improving fast and staying cheap relative to its capability. That’s an argument against building your own models and for building durable, switchable systems on top of someone else’s. Your moat is your data, your workflows, and your judgment — not the weights.
The quiet story: everyone adopted MCP
In late March, both Google DeepMind and OpenAI signaled support for Anthropic’s Model Context Protocol — the open standard for connecting models to tools and data. A shared integration layer across the major labs is exactly the kind of boring infrastructure that pays off for years.
Why it matters: Standardized connections mean less rework when you switch models and less custom glue code. Build to the standard now and you protect yourself from lock-in later.
The throughline for March: the models are no longer the bottleneck. The bottleneck is integration, routing, and judgment about what to build versus what to skip. If you’re deciding where AI fits in your roadmap this quarter, let’s talk through it.