AI News in March 2025: GPT-4o, Gemini, DeepSeek and MCP

In late March 2025, I review a series of AI releases. Artificial Analysis reported that the updated GPT-4o improved on its selected intelligence and coding indices, leading the non-reasoning models it compared for coding. Those scores are specific to the evaluator’s benchmarks and the models available then, not a permanent ranking across all tasks.

The displayed intelligence index rose from 41 for the November 2024 version to 50 for the March update, just behind the new DeepSeek V3 in that index. I also discuss high demand for compute and invoke Jevons’ paradox: cheaper or more efficient AI can encourage more use. That is a possible demand response, not proof that every efficiency improvement increases total spending.

OpenAI described improvements in detailed instruction following, technical and coding tasks, creativity and emoji use. I note that the updated model was being rolled out to users and that image-generation limits were introduced amid high demand. In my own experience, ordinary GPT-4o queries had also become slow; that anecdote is not a controlled latency comparison.

I am impressed by Gemini 2.5 Pro, particularly its coding performance and speed, and say it is the best coding model I have personally used at that point. The model was appearing in tools including Windsurf and Cursor. I want to test its one-million-token context window against my own codebases; a large context window alone does not prove that a model will use every part of it effectively.

DeepSeek also released the V3-0324 checkpoint. Its model card reports improvements over the earlier V3 and lists an MIT license; the weights are available, though the model’s size makes local use demanding. I highlight its reported coding and math results, including AIME 2024, while recognizing that developer-selected benchmarks and settings are not a complete measure of performance.

ARC Prize launched ARC-AGI-2 as a more difficult test of adapting to new visual patterns. The early table shown in the video lists o3 low at 4% on ARC-AGI-2 versus 75.7% on ARC-AGI-1, and a human panel at 100% and 98% respectively. These figures describe the specific early evaluation and panel, not every person or all future models.

The table shows an estimated $200 per task for o3 low and $17 for the human panel. I also misread the o1 high cost aloud as $445; the displayed table says $4.45. I like the benchmark because it asks systems to infer transformations from examples, but my description of it as a perfect AGI test is an opinion: no single benchmark establishes general intelligence.

An example uses colored grids: the solver examines input-output pairs, infers a pattern and applies it to a new input. ARC Prize reported that each ARC-AGI-2 task had been solved by at least two people within two attempts. That task-selection process is useful context for interpreting the human comparison, rather than a claim that every individual finds every task easy.

Zapier announced an MCP integration that could expose its connected apps to AI agents. OpenAI’s Agents SDK and Microsoft’s Copilot Studio also documented MCP support. I see a common protocol for connecting models to tools and context as a major development; the specific integrations, permissions and available tools depend on the service and configuration.

Image generation was another major topic that week. I show examples from Reve Image 1.0 and mention its position in an Artificial Analysis user-vote ranking. I also discuss Ideogram 3.0, whose company reported a strong Elo rating and offered features such as remixing, upscaling and styles. These are selected examples and dated evaluations, not proof that every generated image or text rendering is accurate.

CNBC reported that OpenAI expected revenue to reach $12.7 billion in 2025, roughly triple the previous year, according to sources. That was a forecast, not realized revenue. I am personally optimistic about the value of AI tools and acknowledge my own stake in the subject; my team and family’s experience is anecdotal, and I think education about practical uses matters.

I also discuss reported leadership changes in which Sam Altman would focus more on research and product while Brad Lightcap took a larger operational role. I mention the reported SoftBank investment plans and valuation as they stood at the time; a proposed deal or valuation should not be confused with a completed financing or a measure of profitability.

Qwen announced QVQ-Max as a visual reasoning model able to analyze images and video. Its team presented examples and benchmark claims, including a demonstration comparing two scenes. I describe it as open and downloadable, but the cited QVQ-Max announcement does not establish release of that specific model’s weights; the earlier QVQ-72B-Preview is a separate model.

I note access constraints I encountered with Qwen’s service and hope that additional providers will make models easier to use. Availability, regional requirements and whether weights can be downloaded depend on the particular product and date. The demonstration offers an example of visual reasoning, not a systematic assessment of all its capabilities.