Anthropic's Opus 5 matches Fable 5 performance at half the token cost

The headline buries the more consequential detail: Anthropic is competing on margin compression, not just benchmark rankings. Pricing at half the token cost of a near-equivalent model is a structural bet that volume and enterprise lock-in outweigh near-term revenue per call, a strategy that puts pressure on every competitor's pricing sheet regardless of raw performance.

Modelwire has no prior coverage to anchor this to directly, so this story lands largely without local context. It belongs to a broader pattern that has been building across the frontier-model space through mid-2026, where the competitive axis has visibly shifted from who scores highest to who scores well enough at the lowest cost. The ARC-AGI-3 result is notable because that benchmark was specifically designed to resist memorization, which makes the 30.2 percent figure harder to dismiss as eval overfitting than most leaderboard claims. GPT-5.6 Sol's weaker showing on the same benchmark gives Anthropic a rare apples-to-apples comparison that doesn't rely on self-selected evals.

Watch whether enterprise cloud providers (AWS Bedrock, Google Vertex) reprice or tier Claude Opus 5 differently than Anthropic's listed rate within the next 60 days. If they hold the discount, the cost story is real; if they absorb margin and list it higher, Anthropic's pricing is a direct-sales play, not a market-wide shift.