State of LLMs – April 2026
2026-05-15
Frontier models under pressure, Chinese open-source alternatives closing the gap, new developments in memory and autonomous agents, and a striking survey that reveals just how disconnected executives and workers are on AI adoption.

LLM

GPT-5.4 and Claude Opus 4.6 are the top models right now, but both come with caveats.

Anthropic is struggling with compute capacity and has quietly started serving quantized, degraded versions of Claude to keep up with demand. The performance drop is significant, and the lack of transparency makes them unreliable.

Google is doing something similar with Gemini: models launch strong on benchmarks but degrade quickly. This is a symptom of an industry selling its products at a loss. That said, when inference cost isn't the bottleneck, Google shows real strength on the edge LLM side with their new Gemma 4 release.

GPT-5.4 — along with Codex and the Codex CLI — doesn't seem to have the same compute issues. OpenAI is clearly focused on winning over developers and enterprise customers.

For a fraction of the cost and roughly 90% of frontier model capability, there are now strong open-source alternatives from China: GLM-5.1, Kimi K2.5, MiniMax 2.7, and Qwen 3.6. Qwen also has a solid edge LLM. These models are catching up fast, with good tooling and coding harnesses (MiniMax CLI, Qwen Code, etc.). They're also expected to integrate most Claude Code features, following the leak of its source map.

Harness: software built around an LLM that runs a loop and provides tools and memory capabilities.*

Memory

Two things made noise this month on the memory front: Andrej Karpathy's "LLM Wiki" and MemPalace, a project from actress Milla Jovovich that got a lot of attention.

Autonomous Agents

Claude Code is actively pulling in Open Claw features, and so are Codex and other harnesses.

Hermes Agent keeps popping up in my feeds too, but I'm not sure whether it's genuinely good or just being pushed. I haven't tried it yet.

Enterprise AI Adoption

This quote from the AI Daily Brief podcast stood out to me:

"61% of executives trust AI for complex, business-critical decisions, versus just 9% of workers. On tooling, 88% of executives say their employees have adequate AI tools. Only 21% of workers agree — a 67-point gap."

Not surprising. Workers doing the actual job need far more precision from these tools than executives, who operate at a higher level of abstraction where AI tends to perform well.

In my view, the biggest barrier to adoption right now is the "deploy a chatbot and call it done" mindset. The real value comes from giving AI governed access to your data and systems — which is much harder, but far more impactful.