juan_kok cv.pdf
← back to tracker
June 11, 2026

Baseline — the June 2026 model landscape

First pass of the Frontier Models tracker: 43 models and tools across closed frontier, open frontier, small/on-device, agentic frameworks, and classifiers/encoders — with benchmarks, trade-offs, and a pipeline view of what's next.

This is the inaugural snapshot of the Frontier Models tracker — a sibling to the Inference Engines tracker, but pointed at the models themselves (and the frameworks and classifiers around them) rather than the engines that serve them. It covers five categories with 43 entries, refreshed weekly.

The five categories

  • Frontier · Closed — the proprietary leaders: Claude (Opus 4.8 / new Fable 5 tier**)**, GPT-5.5, Gemini 3.5, Grok 4.3, and Amazon Nova 2.
  • Frontier · Open — the open-weight frontier, now decisively Chinese-led: DeepSeek V4, Kimi K2.6, GLM-5.1, Qwen3.6, plus Mistral Large 3, gpt-oss, NVIDIA Nemotron 3 (best US-open), Llama 4, and the asterisked MiniMax M3.
  • Small & On-device — the “punching above their weight” set: Phi-4, Gemma 4 (phone-to-server), Qwen3.5 small, SmolLM3, IBM Granite 4 Nano, Liquid LFM2, and Apple AFM 3.
  • Agentic Frameworks — from autonomous agents (OpenClaw, OpenManus, OpenHands) to orchestration libraries (LangGraph, CrewAI, Microsoft Agent Framework, OpenAI Agents SDK, Google ADK, smolagents, Claude Agent SDK) and platforms (Dify, n8n).
  • Classifiers & Encoders — the understanding side of NLP: ModernBERT, DeBERTa-v3, embedding leaders (Qwen3-Embedding, EmbeddingGemma, Gemini Embedding), rerankers (Qwen3-Reranker, Cohere Rerank 4), safety classifiers (gpt-oss-safeguard, Llama Guard 4), and GLiNER2 for zero-shot extraction.

What stands out this week

  • Anthropic leads on intelligence and coding. Opus 4.8 tops the Artificial Analysis Intelligence Index at 61.4, and a new Fable 5 tier landed Jun 9 above it (headline numbers still firming up). GPT-5.5 is #2 (~60.2) and SOTA on terminal/agentic coding (82.7% Terminal-Bench 2.0).
  • The open frontier is Chinese-led. DeepSeek V4 (MIT) posts 80.6% SWE-bench Verified and 93.5% LiveCodeBench; Kimi K2.6 leads open SWE-bench Pro (58.6). The best US-open model, NVIDIA Nemotron 3 Ultra (AA Index 48), trails the top Chinese open models by ~6 points.
  • Small models keep closing the gap. Phi-4-mini-reasoning (3.8B) rivals o1-mini on math; Gemma 4 runs natively multimodal in under 1GB on a phone; Qwen3.5-9B reportedly matches gpt-oss-120B on several evals.
  • Agents hit their “1.0 / GA” year. LangChain/LangGraph, Microsoft Agent Framework, Google ADK 2.0, and the OpenAI Agents SDK all reached stable releases. OpenClaw became the most-starred repo in GitHub history — and a security lightning rod. The proprietary Manus corporate model collapsed, leaving open OpenManus as the path forward.

In the pipeline

The most-watched unreleased items: GPT-5.6 (rumored late June), Gemini 3.5 Pro GA (2M context, Deep Think), Grok 5 (in training, rumored up to 10T params), DeepSeek R2 (uncertain), MiniMax M3 weights (pending verification), and Llama 5 (rumored ~2027 as Meta pivots toward closed models).

Process note: this is a fast-moving space and several figures here are vendor-reported or pre-GA. Rumored and unverified items are flagged inline and in each model’s detail card. Always confirm versions, prices, and benchmarks against the lab’s official page before relying on them.