Large language models

Core thesis

By mid-2026 the frontier is four to five labs – Anthropic, OpenAI, Google, xAI – trading the lead generation to generation. Meta abandoned open-weight Llama for the proprietary Muse Spark (April 2026), ceding the open lead to DeepSeek and Qwen, which anchor a capable open tier months behind the closed frontier. Frontier training runs now cost several billion dollars, not the ~$1bn of 2024. Value accrues to the model-and-platform layer and, increasingly, to whoever controls inference compute and agentic tool-use; the open tier is the ceiling on closed-API pricing.

State of the art (2026)

By mid-2026 the frontier is a handful of labs trading the lead, and the defining recent event is governmental. Anthropic launched Claude Fable 5 and the more capable Mythos 5 on 9 June 2026; three days later, on 12 June, a US export-control directive citing national security forced Anthropic to suspend all access to both, after a jailbreak of Fable 5's safeguards exposed Mythos's cyber capabilities. With Fable 5 and Mythos 5 disabled, Claude Opus 4.8 is again the available Anthropic frontier, alongside OpenAI's GPT-5.5, Google's Gemini 3.1 and xAI's Grok 4. Capability has shifted from raw scale to test-time reasoning and long-horizon agents. Open weights closed the gap faster than expected – DeepSeek V4 (MIT, April 2026) posts near-frontier agentic scores at a fraction of closed-API cost, and Meta – once the open-weight standard-bearer – pivoted to the proprietary Muse Spark in April 2026, leaving DeepSeek and Qwen as the open frontier. Anthropic's May 2026 round at a $965bn valuation pushed it past OpenAI, and it has filed confidentially for an IPO. The open question is no longer whether LLMs work but who captures value above the model layer – and, newly, which models are even allowed to ship across borders.

Open source vs closed — the moat question

Meta's Llama, Mistral, and DeepSeek demonstrate that capable open-source models can be trained at fraction of frontier cost. But the gap on the hardest tasks (complex reasoning, agentic multi-step planning) is widening. The moat is not in the model weights — it's in RLHF quality, safety tuning, enterprise distribution, and the data flywheel from production usage.