It's been a dense month. If you felt like every week had a major model drop or framework announcement, you're not imagining it — April 2026 delivered the most concentrated AI news cycle since GPT-4 launched.
GPT-5 and o3 — Reasoning Gets Real
OpenAI shipped GPT-5 and made o3 generally available in the same month, which feels intentional. The positioning is clear: GPT-5 for fluency and breadth, o3 for tasks where you need the model to actually think through a problem rather than pattern-match.
The o3 results on ARC-AGI and competitive coding benchmarks were already known, but seeing it in production is different. The latency is real (plan for 30–90s on hard tasks), but for agentic pipelines where you'd previously have chained 5 GPT-4 calls with hand-rolled verification, o3 collapses that into one reliable pass.
My take: The "reasoning model" framing is doing a lot of work. What o3 is actually doing is longer chains of self-verification — it's not magic, but it's genuinely useful for structured tasks. The benchmark hype is noise; the production utility is real.
Claude 3.7 Sonnet — Extended Thinking Ships
Anthropic's extended thinking mode is now in Claude 3.7 Sonnet and available in the API. The UX is simple: you set a thinking budget (tokens), and the model allocates that budget to an internal scratchpad before producing the final response.
What's interesting is the transparency — you can surface the thinking tokens if you want. For debugging agentic behaviour, this is genuinely useful. The model's reasoning trace often shows you exactly where it went wrong, which is more actionable than a hallucinated answer with no audit trail.
Extended thinking roughly doubles cost for the tasks that use it, so you want to gate it: use it for high-stakes sub-tasks in your pipeline, not as a default.
Gemini 2.0 Flash — Multimodal Becomes Commodity
Gemini 2.0 Flash with its 1M token context and native multimodal input/output (text, images, audio) at sub-$1/1M token pricing is the commoditisation moment. Tasks that would have required GPT-4V at $10+/1M tokens six months ago now run on Flash for under $0.10.
The practical effect: multimodal RAG, document analysis, and image-based extraction pipelines just got much cheaper to operate. If you're building anything that ingests PDFs, invoices, or screenshots, Flash deserves a serious look.
Open-Source — Llama 4 and Mistral Close the Gap
Meta's Llama 4 (Scout and Maverick variants) ships with genuinely competitive benchmark numbers and an MoE architecture that makes the larger variants runnable on reasonable hardware. More importantly, the fine-tuning story improved — the base model is more instruction-following out of the box, so custom fine-tunes need less data to be useful.
Mistral's latest release continues the trend of punching above weight for their parameter count. The 7B and 22B models are the defaults I reach for when I need a local model for experimentation.
My take: The gap between frontier and open-source has narrowed enough that the decision isn't "proprietary vs open" anymore — it's "which task, which cost envelope, which latency budget." Most teams should be running open-source for internal tooling and experimentation, proprietary for customer-facing critical paths.
Agent Frameworks — Finally Not Toys
LangGraph's stateful graph approach has matured. The primitives (nodes, edges, state, checkpointing) map cleanly onto real agent architectures without the "magic string" footguns that plagued earlier LangChain abstractions.
CrewAI has found its niche in multi-agent role decomposition — if your task naturally breaks into "researcher, writer, critic" type roles, it handles the orchestration boilerplate. Not for everything, but the right tool for structured role-play pipelines.
What's changed: six months ago I'd have said "don't use an agent framework in production." That's no longer true for LangGraph specifically. The observability tooling (LangSmith) is good enough to debug failures, and the checkpointing means you don't lose state on partial failures.
Watch for: The Anthropic MCP (Model Context Protocol) is getting traction as a standardised tool interface. If you're building tools for LLMs, MCP-first is increasingly the right default — it'll compose better with whatever orchestration layer you use.
That's April. May will cover: whether the "vibe coding" moment is a bubble or a genuine workflow shift, the first production postmortems on o3-backed agents, and what Gemini 2.5 means for the Flash tier.