rationale:

Target and Status Quo

The forecast estimates the ratio of Anthropic's AI R&D uplift in value to Google DeepMind's (GDM) on December 31, 2028. "Uplift in value" measures the rate increase in research output that the developer itself judges valuable; it is explicitly distinct from raw coding throughput or speedups on old tasks metr.org. As of August 2026, neither lab appears to have crossed its critical thresholds for automated AI R&D. Anthropic's models, including Claude Mythos Preview and Opus 4.8, have not met its 2× capability-progress acceleration threshold or demonstrated the ability to substitute for senior researchers 5anthropic.com. Similarly, Google's Gemini models have not triggered internal ML R&D Critical Capability Levels, lacking the independence to chain tasks into end-to-end research workflows without human intervention 55 sources.

Anthropic's Near-Term Edge

Anthropic currently benefits from aggressive internal dogfooding and a tight, LLM-focused organizational structure. As of mid-2026, Anthropic reported that Claude authored >80% of merged code, with internal surveys suggesting a median ~4× self-reported output multiplier anthropic.com. However, coding throughput does not directly equal research uplift. The translation from labor acceleration to actual research progress is heavily damped by compute limits and experiment cycles, an effect often modeled with a labor exponent around 0.55 2 sources. While pure serial labor uplift may sit robustly above 2× metr.org, the actual in-value research uplift is substantially lower. Nonetheless, Anthropic's monolithic focus allows it to iterate quickly on embedding agentic workflows across its core research and infrastructure anthropic.com.

Google DeepMind's Structural Advantages

GDM possesses formidable countervailing advantages that limit Anthropic's lead in overall value creation. Google's massive compute fleet allows AI labor to convert into experiments far more efficiently than at compute-constrained labs. Crucially, GDM has established automated-research assets like AlphaEvolve, which produce verifiable value uplift independent of chat or coding adoption. These systems have already delivered concrete AI-stack improvements, including up to 32.5% FlashAttention kernel speedups, a 23% speedup of a Gemini architecture kernel, and the compression of kernel-optimization tasks from weeks to days deepmind.google. Furthermore, GDM reports broad internal AI assistance across code, design, and ideation metr.org, alongside broad scientific tooling like Co-Scientist and Gemini Deep Think deepmind.google.

Convergence and Tails

A median ratio of 1.1 reflects a modest but real advantage for Anthropic by end-2028, driven by its explicit strategic focus on the R&D feedback loop and dogfooding. However, this expectation sits close to parity to reflect long-term convergence in lab capabilities and compute constraints. Evaluating this alongside absolute R&D automation timelines confirms our baseline view that Anthropic will likely hold a slight, persistent lead over DeepMind due to focused execution, though structural compute constraints will prevent the gap from widening dramatically. Over a 2.4-year horizon, the rapid diffusion of model capabilities, agent tooling, and talent strongly favors this convergence. The left tail (encompassing a solid chance that GDM outperforms Anthropic, pushing the ratio down to 0.94 at the 25th percentile and 0.76 at the 10th percentile) reflects scenarios where Google’s compute abundance and specialized automated discovery systems allow it to scale overall value uplift faster. Conversely, the fat right tail (reaching 1.35 at the 75th percentile and 1.75 at the 90th percentile) covers scenarios where Anthropic's focused approach compounds into a genuinely distinct research regime, while GDM's R&D value remains diluted across a wider array of product, legacy, and infrastructure demands.