Current State and Baseline We estimate that Anthropic currently holds a modest lead over Google DeepMind in AI R&D uplift in value, placing the present ratio roughly between 1.1 and 1.3. Anthropic has aggressively integrated agentic workflows, with Claude now "leading" (AL4) 26% of its measured AI R&D tasks https://www.anthropic.com/institute/measuring-pace-of-ai-development https://www.anthropic.com/institute/measuring-pace-of-ai-development. A preliminary, third-party assessment by METR estimated Anthropic's overall acceleration at approximately 1.5x, though the underlying evidence was not shared. In contrast, Google DeepMind's disclosures suggest a lower baseline or more conservative measurement: Gemini 4 Argon remains below Google's Critical Capability Levels for ML R&D https://deepmind.google/models/model-cards/gemini-4-argon/ https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-4-Argon-Model-Card.pdf https://deepmind.google/models/model-cards/gemini-4-argon/, and safety documentation notes "no significant end-to-end acceleration" https://storage.googleapis.com/deepmind-media/gemini/gemini_4_safety_report.pdf.

The Case for Anthropic Divergence The primary driver for the right-skewed tail—where the ratio stretches toward 1.4 and approaches 1.8—is the compounding nature of recursive self-improvement (RSI). Anthropic has demonstrated a strong early focus on dogfooding its own agents. If Anthropic enters a compounding RSI loop even a few months ahead of Google, a modest chronological lead could translate into a massive absolute gap in R&D multipliers by late 2028. Furthermore, disputed reports of DeepMind staff utilizing Claude Code suggest that Anthropic's tooling advantage may be tangible enough to actively influence competitor workflows https://www.businessinsider.com/google-deepmind-ai-tool-divide-internal-tensions-2026-4 https://www.businessinsider.com/steve-yegge-google-ai-adoption-curve-demis-hassabis-2026-4.

Compute Bottlenecks and Google's Structural Edge However, several structural constraints strongly bound Anthropic's upside and could pull the ratio toward or below parity. Crucially, research uplift in value is not equivalent to coding throughput; it requires running experiments and validating hypotheses, which makes compute a strict bottleneck. Google's massive in-house TPU infrastructure, including its Ironwood clusters, gives it a distinct advantage in converting automated coding into actual research milestones and acts as a stronger constraint against an Anthropic breakaway October 8, 2026 - Nathan Benaich AIR STREET CAPITAL.. Anthropic relies on cloud partners and explicitly names compute, alongside human review and research judgment, as a binding limit https://www.anthropic.com/institute/recursive-self-improvement.

Convergence and the Final Distribution Ultimately, the extreme upper tails are constrained by both Google's talent density and the broader diffusion of capabilities between frontier labs. Furthermore, if leadership at both labs pursue proposed pacing policies, artificially capped R&D speeds would force the ratio closer to 1.0. Our median projection of 1.15 reflects Anthropic's persistent but bounded edge in agent deployment. The 10th and 25th percentiles are 0.85 and 1.0, respectively, capturing a meaningful chance that Google's overwhelming compute resources and deep research bench allow it to match or overtake Anthropic. The 75th percentile is 1.4, while the 90th percentile reaches 1.8 to capture scenarios where Anthropic achieves a rapid, unmitigated self-improvement breakout, with the upper tail slightly compressed because structural compute constraints weigh heavily against massive unmitigated divergence.