June 14, 2026 · 8 items · 3 sources
Anthropic released Claude Fable 5, hitting 95% on SWE-bench — shattering the previous record (Opus 4.8 at 88%). 1M token context, 128K output, 91/100 Senior Engineer score. Priced at $10/$50 per million tokens. Simultaneously released Claude Mythos 5 for scientific reasoning.
OpenAI deployed a new memory system called "Dreaming" (June 4) that synthesizes context across chat sessions. Memory becomes product infrastructure, not a setting. Stale/contradictory context reduced.
May 2026 rankings: Kimi K2.6 and MiMo-V2.5-Pro tie at AA Index 54 — just 3 points below closed-source leaders. DeepSeek V4 Pro is #1 for agentic work (GDPval-AA Elo 1554, SWE-Bench 80.6%). 9 major models shipped in 6 weeks.
Microsoft shipped a dense wave around Copilot: app (expanded preview), CLI refresh, SDK GA, cloud/local sandboxes. AI coding moves from autocomplete to managed work sessions. Direct competition for Claude Code and Codex.
Non-standard transformer architecture targeting agentic workflows. Too early for production, but adds competitive pressure to the frontier.
30B total / 3B active (MoE), Apache 2.0 license, 256K context. Runs on modest hardware. Best option for self-hosted lightweight coding AI.
H Company published variants from 0.8B to 35B on Hugging Face. Screen-control agents that run locally — privacy and latency play.
Inference runs locally on Apple Silicon, Swift-native APIs. Privacy-first for health/finance apps. Apple ecosystem only.
The open-source vs. closed-source gap has never been smaller: 3 index points. Release cadence is accelerating (9 major models in 6 weeks). The battle is shifting from "best model" to "best agent ecosystem."