AI Signal

The field's center of gravity has moved from model capability to control of the agent layer: who owns the harness that sits across enterprise software, who routes between models, and who enforces permissions the model itself cannot be trusted to respect. The frontier is multi-vendor again. Opus 5.5, GPT-6 Astra and Gemini 4 Carbon each lead different domains, and Carbon comes close to the leaders at about one-third the cost. Meanwhile labs are openly softening safety commitments and conceding they may ship models they cannot verify. Capability on realistic professional work is still low and uneven: about 12% top score on GDPx, and 0% on binary-graded finance and legal tasks. Value is therefore accruing to whoever builds verification, routing and enforcement around the models, not to the models alone.