AI Roundup April 2025: Llama 4 Drops, OpenAI Ships o3 and GPT-4.1, Gemini 2.5 Flash Goes Hybrid

TL;DR

April 2025 was a model dump. Meta open-weighted Llama 4 (Scout and Maverick, both mixture-of-experts) on April 5, then spent the next week answering questions about its benchmark numbers. OpenAI shipped the API-only GPT-4.1 family on April 14 and the o3 and o4-mini reasoning models on April 16. Google used Cloud Next to launch Gemini 2.5 Flash, its first model with a tunable thinking budget. If you build, your defaults changed this month.

Meta Drops Llama 4 Scout and Maverick, Then Gets Grilled on the Benchmarks

On April 5, Meta released Llama 4 Scout and Llama 4 Maverick, its first natively multimodal, mixture-of-experts open-weight models. Both run 17B active parameters: Scout pairs that with 16 experts (109B total) and a headline 10M-token context window, while Maverick uses 128 experts (400B total) with a 1M context. A larger model, Behemoth, was teased but still training. The MoE move means you only light up a slice of the network per token, so you get big-model quality without paying full-dense inference costs.

The asterisk: the version of Maverick that posted those eye-catching LMArena results was an experimental chat-tuned build, not the weights you actually downloaded. That distinction set off a fair amount of noise about whether the public numbers matched the public model. For builders the lesson is the usual one: benchmark the weights you can run, not the leaderboard entry.

Why it matters for builders

Scout's long context and MoE efficiency make it genuinely interesting for local and self-hosted RAG. Just temper expectations against your own evals rather than the launch chart.

OpenAI Ships GPT-4.1, GPT-4.1 mini and nano (API Only)

On April 14, OpenAI launched the GPT-4.1 family: the full model plus mini and nano tiers, available only through the API. The pitch was sharper coding, better instruction following, and up to 1 million tokens of context. The nano tier is the headline for cost-sensitive builders: a cheap, fast model with a long context window is exactly the kind of workhorse that quietly powers a lot of production pipelines.

Worth noting these did not land in ChatGPT, they were API-first. That is a tell about who OpenAI was aiming at this round: people shipping software, not chat users.

OpenAI Follows With o3 and o4-mini, Its Tool-Using Reasoners

Two days later, on April 16, OpenAI released o3 and o4-mini, replacing o1 and o3-mini. The pitch was reasoning models that actually use tools mid-thought: searching the web, running Python, and reasoning over images inside a single chain of thought. They rolled out first to Pro, Plus and Team subscribers, with an o4-mini-high variant that thinks longer for higher reliability. By April 24, o4-mini reached free-tier ChatGPT users and the Chat Completions and Responses APIs.

The interesting bit is agentic by default: a reasoning model that decides on its own to call a tool, look something up, then keep reasoning. That is the shape a lot of us have been duct-taping together by hand, now baked in.

Google Launches Gemini 2.5 Flash With a Tunable Thinking Budget

At Google Cloud Next on April 9, Google unveiled Gemini 2.5 Flash in preview, billed as its first fully hybrid reasoning model. The headline feature is a thinking budget: you can turn reasoning on or off, or tune how much compute the model spends per query, trading latency and cost against accuracy on a per-call basis.

That is a quietly important idea. Most of us do not need a model to think hard about every request. A dial that lets you spend reasoning tokens only when a task actually warrants it is the kind of cost-control knob production teams have been asking for. It launched through AI Studio and Vertex AI.

Key Takeaways

• Open weights got a real MoE option. Llama 4 Scout and Maverick bring mixture-of-experts efficiency and long context to self-hosters, though you should trust your own evals over the launch benchmarks.
• OpenAI split its lineup by audience. GPT-4.1 went API-first for builders while o3 and o4-mini chased agentic reasoning, a clear signal of who each model is for.
• Tool use moved inside the reasoning loop. o3 and o4-mini calling search and Python mid-thought turns scaffolding you used to write yourself into a native capability.
• Reasoning became a dial, not a switch. Gemini 2.5 Flash's thinking budget lets you pay for deliberation only when a task needs it, which is exactly the cost lever production workloads want.
• The release cadence is brutal. Four major model launches in a single month means your model-selection assumptions have a shorter shelf life than ever, so keep your eval harness warm.