Does Claude Opus 5.5 fully automate AI research tasks?

METR's predeployment evaluation of Claude Opus 5.5 — conducted under an unpaid AI R&D assessment agreement with Anthropic, with METR drafting the initial summary, Anthropic given "the opportunity to review and edit the text," and METR signing off on the final text — assesses two separate claims: (A) whether Opus 5.5 would now dramatically accelerate AI R&D at Anthropic, and (B) whether AI already dramatically accelerated the R&D that produced Opus 5.5. On (A), using five capability tasks (Budget NanoGPT Speedrun, LMCA, Gaming Bot, Sunlight, plus trend data from METR's Frontier Risk Report), METR finds Opus 5.5 "improves upon Fable 5.1 across both verifiable tasks ... and harder-to-verify tasks," but is "an incremental improvement ... rather than a discontinuous jump," and is "unlikely to be able to fully automate AI R&D."

METR's reasoning for withholding a full-automation verdict is a skills gap, not a scores gap: full automation would require "large improvements in foresight, prediction, creating one's own feedback loops, and generally other skills that might typically be referred to as researcher 'judgement' or 'taste,'" and "the evidence we have does not suggest that Claude Opus 5.5 represents a large improvement over Fable 5.1 in these 'judgement' skills." On (B), METR is explicit that it did not independently verify the figure it cites for the model's own development: it relays, without endorsing, a "highly experimental and preliminary report from a separate METR assessment of AI R&D acceleration inside Anthropic," which estimated "~1.5X overall acceleration in capabilities due to AI ... with perhaps 30% chance of 2X acceleration." METR's own team states it uses that report's evidence "as an input to our assessment but do not argue directly in defense of its claims."

This sits in tension with Is AI development already being handed to AI systems?, whose cited speedup runs from about 3x to about 52x — far larger than, and independently sourced from, the ~1.5x (30% chance of 2x) that METR relays here secondhand. It also complements Does AI assistance erode the skills needed to oversee it?: METR's judgement-and-taste gap gives a capability-side account of why full delegation stays bounded, alongside that survey's oversight-side account of the same ceiling. And it gives a live instance of the verifiable-versus-harder-to-verify split that Do fixed-budget efficiency gains translate to real research progress? has to bridge from a benchmark score to an R&D trend, here at the point of an actual predeployment evaluation rather than a benchmark proposal.

The excerpt does not establish what rate of continued incremental gains would eventually close the judgement gap, nor whether the ~1.5x figure it relays is itself a reliable measurement — METR flags both as unresolved, noting "the data we have is insufficient for distinguishing consistent, accelerating, or decelerating rates of improvement." The implication, at the strength the evidence allows, is that the two headline risk cases this batch keeps circling — autonomous AI R&D and self-accelerating model development — currently rest more on cited internal estimates and relayed, unverified figures than on capability measurements an outside evaluator could independently confirm.