Anthropic: Claude Opus 4.6 (Fast)

Claude Opus 4.6 has strong base-model evidence: Anthropic provides launch, API, and system-card documentation; Terminal-Bench 2.0 reports 58.0%±2.9 for Anthropic’s Claude Code configuration; and the supplied LMArena evidence places the named Fast variant first at 1500 Elo. This supports high capability, reasoning, and public-signal scores, though it remains below the published Opus 4.8 anchor (93 technical, 95 coding) because there is no equivalent direct benchmark package here. It is materially stronger than GPT-5.6 Terra on reasoning evidence and Pegasus 1.5 on general and coding scope. The principal issue is variant validity. Anthropic’s current Fast-mode documentation explicitly says `claude-opus-4-6` requests with `speed:"fast"` run and bill at standard speed. Consequently, the listed Fast configuration should not receive a premium speed score despite the marketplace naming. Terminal-Bench evidence applies to Opus 4.6, not demonstrably to a distinct Fast-serving path. No direct DeepSWE, SWE-bench, LiveCodeBench, Artificial Analysis, Aider, or Arena-Hard result was supplied for this variant. Developer experience remains solid because the base API, model identifier, cloud availability, and documented control are established, but preview status and the effective standard-speed fallback reduce operational certainty. At $30/M input and $150/M output, it is substantially less cost-effective than the mid-priced published anchors and offers no verified fast-mode benefit. Pricing documentation is reasonably clear, but the discrepancy between advertised Fast pricing and current fallback behavior requires validation before deployment.