UPDATED JULY 27 — K3 WEIGHTS NOW LIVE
● AA Intelligence Index: K3 #4 (score 57) — Opus 4.8 score ~52 (K3 leads)
● Price: K3 $3/$15/M — Opus 4.8 $5/$25/M (K3 40% cheaper)
● Agentic coding: K3 SWE Marathon 42.0% #1 — Opus 4.8 not published
● Self-hostable: K3 (weights live) — Opus 4.8 closed
● Hallucination: K3 51% (independent) — Opus 4.8 lower (Anthropic discloses Opus 5 higher, implying 4.8 baseline is better)
● Data residency: K3 hosted = China NI Law. K3 self-hosted = your cloud. Opus 4.8 = Anthropic US.
● Production maturity: Opus 4.8 — months of production use. K3 weights released today.
Kimi K3 (self-hosted) for: Teams with GPU infrastructure who need frontier-adjacent intelligence at lower cost, full data sovereignty, and no vendor dependency. The 51% hallucination finding means test on your specific tasks first — K3 is clearly superior for agentic coding and frontend; factual tasks need validation.
Claude Opus 4.8 for: Teams that need proven production stability (months of battle-testing), better hallucination profile, and US data residency without self-hosting overhead. Note: Opus 5 (launched July 24 at same price) supersedes Opus 4.8 for new workloads — upgrade to Opus 5 rather than staying on 4.8.
Last updated July 27, 2026. Related: K3 download guide → · Claude Opus 5 (replaces Opus 4.8) →
⚖ Our Verdict
Kimi K3 wins on intelligence (AA Index #4, score 57 vs Opus 4.8 ~52), agentic coding (SWE Marathon #1), price (40% cheaper), and now self-hostable (weights live July 27). Opus 4.8 wins on hallucination profile (K3 has 51% independent rate), production maturity, and US data residency on hosted API. Note: Opus 5 (July 24) supersedes Opus 4.8 at same price — upgrade to Opus 5 for new Anthropic workloads.