Claude Is Helping Build Its Successor: Inside AI Research

Claude is helping Anthropic build its next AI model. Explore what Claude does in model development and why human oversight matters as AI takes on more research.

By ELYMENT Insights
Claude Is Helping Build Its Successor: Inside AI Research

Anthropic reports that Claude led 26% of its measured AI research and development work in August 2026, under human supervision. The company rated more than 90% of the measured work as at least collaborative and recorded no fully autonomous work in the index. For Sydney and NSW businesses testing AI agents, the distinction is practical: delegating a task and granting authority to approve or release the result are separate decisions.

A failed data pipeline offers a concrete example of what Anthropic means when it says Claude can lead research work. In an example set out in the company’s methodology, an engineer gives Claude a failure alert. Claude investigates the problem, tests a fix and reports what changed. The engineer decides whether to deploy it. That handover is where the lab draws the line between leading a task under supervision and completing it autonomously.

Anthropic’s research on the pace of AI development, reported by Reuters in September 2026, offers three ways to look inside that process: how much research work Claude takes on, how the lab monitors agent activity, and where it puts its computing resources. Each figure has a different denominator.

What the 26% actually counts

Anthropic built its prototype R&D Automation Index from internal work records. A Claude agent helped catalogue about 15,000 granular tasks from sampled staff records in July 2026. Anthropic grouped that work into categories and weighted them using an approximation of staff time. It then rated those categories using a scale developed by Epoch AI.

Three levels of AI involvement relevant to Anthropic’s findings

Collaborates or higher

  • What it means in practice: At the collaboration threshold, Claude completes substantial parts of a task with close human direction.
  • August 2026 measurement: More than 90% of the weighted work reached this threshold or higher.

Leads

  • What it means in practice: Claude can handle most of a task from a high-level instruction while a human supervises.
  • August 2026 measurement: 26% of the weighted work, up from under 1% in February.

Fully autonomous

  • What it means in practice: Claude carries out the work without requiring a human in the loop.
  • August 2026 measurement: No measured work reached this level.

The 26% sits inside the broader figure of more than 90%. Both refer to a weighted set of work categories. Hours saved, research output and changes in staffing would each require separate measurement. The index also holds the July task categories fixed to compare them over time, so work that emerges later can fall outside that baseline. Anthropic used a separate Claude instance to rate the evidence and compared its ratings with staff assessments. The company acknowledges room for judgement at the boundary between collaboration and leadership.

Who sees what an agent does

Anthropic reported about 30,000 concurrently active research and engineering agents on its most-used internal platform as of August. On that platform, the company says every agent action passed through an automated monitor before execution and was processed by another monitor afterwards. Humans reviewed blocked actions within a week. The highest-priority cases flagged by the retrospective system were also escalated to people.

Coverage shows which actions pass through a monitor. Detection requires a separate test of how often the monitor catches a known problem. The distinction matters as agents work for longer without constant supervision. Evidence that a system could independently produce and release a successor would have to cover design, training, evaluation and release decisions across the entire process. The August index establishes the extent of task delegation within its measured scope. Elyment’s coverage of Anthropic’s arrangements for independent AI evaluation examines what an outside reviewer can verify.

The computing figures cover one week

From 13 to 20 July 2026, Anthropic estimates that about 6% of the computing capacity it used for AI R&D went to safety work. The figure was about 12% within the narrower category of AI-driven AI R&D. The company counted workloads serving safety and model capability equally outside the safety category for this calculation. It also excluded compute used by safeguard classifiers from these safety estimates.

Compute measures the processing resources used to run the work. A training run can consume considerable capacity in a short period, while designing a safety experiment can take substantial staff time and comparatively little machine time. A longer series using consistent categories would help show whether resource allocation is shifting. The July figures capture one week of use.

What a Sydney team can measure in its own pilot

Consider a Sydney property operations team piloting an AI agent to compare fit-out scopes from tender documents. The agent could produce a comparison table, link each finding to its source and flag scope gaps. The project lead would resolve ambiguous exclusions and approve what goes to suppliers. That gives the team three measures it can use together: preparation time, reviewer correction time and cost per accepted comparison. It also gives the agent a clear stopping point before anything leaves the organisation.

The National AI Centre’s adoption guidance addresses accountability, evaluation and human oversight. NSW government agencies also have their own AI Assessment Framework. For private operators, Elyment’s guide to workflow automation for Sydney operations teams and its coverage of training engineers to deploy Claude in businesses look at how these decisions carry into day-to-day work.

Review an AI workflow before expanding it

Bring one recurring task, a sample of the work and the decision it feeds into. Elyment can help Sydney and NSW teams scope a pilot, define who approves the output and agree how to measure the result.

Request an AI Workflow Review

Sources and references


Explore more ELYMENT articles