Part III · 8 — State of the art (2026 snapshot)

Era 7 is open — it is being written at the speed at which you read. This doc is a dated snapshot of the frontier in mid-2026: the models, the scoreboards and the recent milestones. Unlike the earlier eras (history), here the content ages fast and must be revised with each generation of models.

🎨 Figure F-III.20 — The 2026 constellation of models. Brief: "starry sky"-type infographic with the main frontier models of 2025-2026 as stars grouped by lab (Anthropic, OpenAI, Google, DeepSeek, Alibaba, Moonshot, Meta, xAI, NVIDIA, Microsoft), star size ~ capability; thin lines connecting lineages. Compendium palette.

---

8.1 Recent timeline (2025 → mid-2026)

Milestones published after the Transformer turning point; release cadence of the current phase.

• 2024-11 — MCP · Anthropic launches the Model Context Protocol (a tool

  standard for agents); OpenAI adopts it in 2025-03; donated to the Linux Foundation (AAIF) in 2025-12.

• 2025 — Blackwell · NVIDIA B200 (2.5× H100, 25× more efficient) and

  B300/Blackwell Ultra arrive in the frontier clusters.

• 2025-05 — Veo 3 · Google · first video generator with native audio.
• 2025-09 — Sora 2 · OpenAI · realistic physics + synchronized audio.
• 2025-10 — Llama 4 Scout/Maverick · Meta · multimodal; 10M tokens of

  context (open record).

• 2025 — A2A protocol · Google · agent-to-agent communication (complementary

  to MCP); 150+ orgs, Linux Foundation.

• 2026-01 — Veo 3.1 · Google · 4K, vertical video, ingredients-to-video.
• 2026-02 — Gemini 3.1 Pro · Google · leads benchmarks (SWE-bench Verified

  78.8%; GPQA Diamond 94.3%; ARC-AGI-2 77.1%); native MCP.

• 2026-03 — Qwen 3/3.5 · Alibaba · MoE flagship 235B-A22B.
• 2026-04-02 — Gemma 4 · Google DeepMind · Apache 2.0 family (E2BE4B26B

  MoE/31B Dense); native multimodal, 140+ languages, 256K context.

• 2026-04-03 — Nemotron 3 Super · NVIDIA · LatentMoE (better accuracy per

  FLOP); 25T tokens of pretraining.

• 2026-04-20 — Kimi K2.6 · Moonshot AI · 1T-A32B MoE; Agent Swarm (300

  sub-agents, 4,000 steps); open-source leader on SWE-Bench Pro (58.6%).

• 2026-04-23 — GPT-5.5 · OpenAI · improved agentic/computer-use.
• 2026-04-24 — DeepSeek-V4 · MIT · V4-Flash (284B-A13B) and V4-Pro (1.6T-A49B);

  1M tokens; Compressed Sparse Attention (27% of the FLOPs, 10% of the KV cache of V3.2); Muon optimizer; price ~40-50% below competitors.

• 2026-04 — Grok 4.3 Beta · xAI · native video; 2M context; Grok 5 in training

  on Colossus 2 (1.5 GW).

• 2026-06-02 — Microsoft Build 2026 · the MAI family (7 in-house

  foundation models, from scratch, no distillation), Scout (always-on "autopilot" agent with governed identity in Entra), IQ/Web IQ (MCP-native grounding). External signal of the self-hosted-first thesis.

Claude / Fable family (Anthropic): frontier line in alignment and agency (Constitutional AI/RLAIF), the basis of Claude Code; current generations Claude 4.X (e.g.: Opus 4.8) and Fable 5.

---

8.2 Scoreboards (March/2026 — snapshot)

SWE-bench Verified (real GitHub issues — the most relevant benchmark for code agents):

Arena Elo (blind human preference):

Humanity's Last Exam (PhD-level questions): several frontier models already match or surpass the human average.

⚠️ Contamination: public benchmarks are partially or fully contaminated in the training of frontier models. To truly evaluate a domain, build private benchmarks (see ../05-building-ai/05-benchmarks-and-leaderboards.kmd).

---

8.4 Safety and regulation (2026)

• Constitutional Classifiers (Anthropic, 2026) — jailbreak success fell from

  86% → 4.4% (blocked 95% of attacks); 3,000+ hours of red teaming with no universal jailbreak.

• Mechanistic interpretability — named MIT Breakthrough Technology of 2026;

  "microscope" techniques integrated into deployment decisions; Gemma Scope 2 (DeepMind) is the largest open-source toolkit.

• EU AI Act — GPAI obligations in force since Aug/2025; high-risk systems

  must comply from Aug/2026 (HRAI may extend to Dec/2027 via the Digital Omnibus).

• Agent protocols: MCP (tools) and A2A (agent-to-agent), both under

  the Linux Foundation — the de facto standard for agentic systems.

8.5 Modalities — 2026 snapshot

• Native speech-to-speech replaces the ASR→LLM→TTS cascade: Moshi, GPT-4o

  Realtime, Gemini Live. Practical PT-BR: Whisper-v3 + LLM + F5-TTS/XTTS v2 (_{800ms) or a fine-tune of Moshi (}300ms).

• Image: the dominant one is Rectified Flow + DiT (SD3, Flux.1,

  AuraFlow).

• Audio: ASR Whisper v3/v4, Canary, SeamlessM4T; TTS F5-TTS, XTTS v2 (PT-BR),

  Fish Speech; codecs EnCodecMimiDAC; music MusicGen, AudioLDM 2.

• Video: Veo 3/3.1 (native audio, 4K), Sora 2; fine-tune of

  CogVideoX/HunyuanVideo viable (US$500–5,000).

• 3D and world models: NeRF, 3D Gaussian Splatting, SDF; 3D generation

  (Shap-E, TripoSR, Trellis, Hunyuan3D-2); interactive world models (DreamerV3, Genie 2, GameNGen).

Conceptual detail of each modality in the chapters of Part IV (02 diffusion, 07 ASR, 08 TTS, 09 music, 10 video).

8.6 How to read this snapshot

• Numbers age; the form persists: the frontier is defined by *ata +

  post-training* not by exotic architecture (they are all Transformers/MoE).

• Underlying trends of this phase: long context (1M–10M tokens), reasoning

  models (test-time compute), MoE to scale cheaply, agency (MCP/A2A, computer use), native multimodality and inference efficiency (KV cache quantization, speculative decoding).

Update this page with each generation. The rest of the Compendium (concepts, life cycle, types) is timeless; only this page is a portrait of the now.