Model wire: three flagships in six days

The last week of October 2026 was absurd even by this industry’s standards: three flagship-class models in six days, plus a speed merchant worth knowing. A field guide, from the roundups linked below:

Kimi K3.1 Pro (Moonshot AI, Oct 23). A 3T-class open-weight flagship with a 1M-token context, thinking-only operation and agent training on top of the K3 base. API-first at $2/$10 per million tokens, weights promised "by Nov 4" — the open-weight launch pattern is now: API today, download soon.

GPT-6.1 Terra (OpenAI, Oct 26). The new mid-tier: visual reasoning, precise computer use, natural writing, priced at half its predecessor’s input cost ($1/$5). The flagship fights make the headlines; the mid-tier price cuts change the budgets.

Gemini 4 Carbon (Google, Oct 28). Announced as Google’s strongest model for cybersecurity and long-horizon work — and not actually available: access starts with "a subset of cyber defenders", with no public model ID yet. Release no longer means access.

Mercury 3 (Inception, Oct 19). The odd one out: a diffusion LLM running at a reported 1,068 tokens/second with a 1M-token window. When generation is that fast, interfaces built around waiting — streaming cursors, progress bars — start to look like legacy UI.

The other story of the week: both major independent leaderboards changed their scoring methods within days, so October’s numbers and November’s are not comparable. Read rankings accordingly.