Grok 4.3 regresses on benchmark, cuts costs

This is the kind of benchmark regression that makes “better and cheaper” claims look premature unless the product can defend the quality drop on real workloads.

• –The headline number is a major regression, not a marginal wobble: 93.4 down to 67.5 is large enough to be a meaningful change in capability.
• –The lower cost matters, but only if the cheaper run still meets the user’s bar; here, the benchmark suggests it may not.
• –Extended NYT Connections is a narrow benchmark, so it should be treated as one signal rather than a full product verdict.
• –If this reflects an actual model change rather than a prompt/eval artifact, it points to a release that optimized efficiency at the expense of reasoning consistency.
• –The GitHub benchmark referenced in the post makes this easy to reproduce or challenge, which increases the credibility of the comparison.