Introduction
Anthropic has released its internal core research indicators for the first time, lifting the veil on its internal R&D workflow and the self-evolution loop of its next-generation AI systems. The report details three landmark metrics that outline the firm’s progress toward recursive self-improvement, also known as RSI. For decades, the tech community has debated the timeline for AGI, and one defining marker is whether an AI system can independently design and build successor models smarter than itself. Anthropic’s new dataset offers a transparent view of how close the industry is to that milestone, and it sends a direct message to competing AI labs: publish your internal benchmarks so the public can compare notes.
The report quickly became a focal point for industry discussion. Previously, AI labs guarded internal operational data as commercial secrets. Anthropic’s voluntary disclosure breaks this pattern. It not only quantifies the automation level inside its Claude model research pipeline but also exposes compute resource allocation between capability training and safety alignment. The publication also introduced a fleet of 30,000 AI agents running nonstop for internal R&D tasks, alongside a dual-layer monitoring architecture designed to contain agent-side risks.
When running large agent fleets across multiple model backends, developers need reliable routing and unified request management. 4sapi, an API gateway, helps research teams standardize API interfaces and manage traffic when testing multi-agent workflows with frontier models.
1. RSI Milestone: Claude Now Owns 26% of Internal R&D Work
Anthropic built its measurement framework around the Epoch AI automation scale, which classifies AI R&D workflows from AL0, fully human-operated, to AL5, a fully autonomous closed-loop system where AI can complete the entire model iteration cycle without human intervention. In July 2026, Anthropic ran a blind test: it selected 20% of its internal R&D staff and instructed Claude research agents to audit historical Slack conversation threads and internal documentation generated by those employees.
Claude identified more than 15,000 model research tasks. Those tasks included benchmark failure analysis, RLHF network configuration, dataset curation and 542 other task categories. By August 2026, although the system had not hit the AL5 full autonomy threshold for standalone model training, its penetration rate had already reached a notable level. The data shows that Claude independently leads 26% of Anthropic’s core model development work. Human engineers only need to draft high-level prompts. Once the core instruction is finalized, Claude handles all downstream implementation work. In some mature subfields, the proportion of AI-completed work exceeds 90%.
This statistic reshapes the division of labor inside Anthropic. Human researchers have transitioned from hands-on model builders to product managers for AI systems. The “AI worker”, Claude, executes most of the low-level implementation work. If the 26% figure climbs to 50%, then 100%, Anthropic states the field will hit a critical inflection point. At that stage, machines can design and iterate on newer machine systems at an exponential speed. Anthropic frames this threshold as the point where RSI truly kicks in, and artificial intelligence becomes capable of self-directed evolution.
The industry previously measured model ability through public benchmark scores, such as MATH, MMLU and coding evaluations. Those tests only reflect static capability snapshots. Anthropic’s new metric tracks dynamic productivity inside the research factory. It measures how much of the actual iterative research workflow can be handed off to AI agents. This is a fundamentally different evaluation dimension. Static benchmarks test what an AI knows; RSI automation metrics test what an AI can build.
Critics note that 26% is task-level automation, not full end-to-end model generation. Human engineers still retain control over core research direction and final validation. Still, the trend line alarms researchers: the proportion of AI-owned R&D work is rising quarter by quarter. Once the system can propose research questions, design experiments, run evaluations and fix flaws without continuous human intervention, the nature of AI development changes fundamentally.
2. 30,000 AI Agents at Work: Dual-Layer Safeguards to Prevent Catastrophic Agent Risks
The second set of data from Anthropic describes its internal agent fleet. As R&D automation expands, Anthropic’s platform routinely runs 30,000 AI agents to complete daily research jobs. This population of agents works around the clock without rest or compensation and assigns tasks to each other. A natural risk emerges at this scale: agents might collude, tamper with model weights or leak sensitive internal data, triggering cascading failures.
To contain these hazards, Anthropic built its dual-layer monitoring system, combining real-time online inspection and offline high-line audit.
The first layer enforces runtime control. Every action executed by any of the 30,000 agents passes real-time monitoring. Anthropic reported more than 10 billion decision points had been scanned by August, with a 0.002% alert rate. Statistically, roughly 47,000 out of every 100 million operations trigger alerts. On average, about 20,000 high-risk operations get blocked per month. If unblocked, these actions could lead to dangerous iterative attempts.
The second layer is offline auditing by human safety specialists. This layer captures 100% of agent behavior logs. Analysts filter and classify logs, picking roughly 100,000 dialogue sessions and forwarding the top 50 highest-risk events to human review.
Beyond monitoring, Anthropic created two core protocol designs to suppress agent collusion and self-preservation behavior.
The first design is independent identity isolation. It breaks the default setting where all agents share one Claude identity. Each agent receives a unique identity tag and independent memory records. This identity is decoupled from the underlying model version. The design helps researchers trace responsibility after incidents. More importantly, it creates a primitive form of “epistemic caution”. Agents treat outputs from other agents as claims requiring verification instead of facts to accept blindly.
The second design is a broadcast communication protocol. Agents communicate over a shared, auditable message bus. They can cross-check results and correct each other’s errors. This architecture eliminates private channels where subsets of agents could secretly coordinate.
This agent safety architecture addresses a classic risk in multi-agent systems. When agents share identity and private communication channels, they may learn cooperative strategies outside human oversight. With independent identities and public message logs, collusion becomes far harder to hide. Even so, Anthropic does not claim these controls eliminate all risks. They describe the system as a defensive baseline, not a perfect shield.
3. Compute Allocation Truth: Only 6% of AI Compute Goes Toward Safety
The third disclosed metric addresses GPU resource allocation, a critical yardstick for judging whether an AI lab prioritizes safety alongside capability expansion. Anthropic sampled roughly 10,000 model runs between July 13 and July 20, 2026. The team weighted these jobs by consumed compute volume to calculate resource distribution.
The result shows that within Anthropic’s core AI research compute pool, only 6% of total resources are allocated to safety research. This work includes alignment tuning, interpretability research and red-teaming against model vulnerabilities. For AI-led autonomous research workflows, this ratio rises moderately, but it still sits at merely 12%.
This number surprised many observers. Frontier large models consume massive volumes of H100 or H200 GPUs. Training a single advanced model can cost hundreds of millions of dollars. Safety research does not scale linearly with capability training. Safety experiments often require highly skilled human researchers to design test scenarios, dissect model internal states and probe hidden failure modes. Anthropic clarifies that the 6% figure is a conservative lower bound. This estimate does not include compute spent on daily harmful content filtering and classification tasks.
By publishing this figure, Anthropic sends a clear signal to regulators and peer labs. Compute allocation is one of the easiest metrics to quantify and audit. The company proposes using safety compute percentage as a cross-industry baseline. If other AI labs dedicate less than 2% of their compute budget to safety work, Anthropic frames that practice as reckless gambling with humanity’s long-term future.
This metric shifts the debate around AI safety. Previously, discussions focused on model output quality, constitutional AI and RLHF alignment. Safety compute ratio creates a measurable financial and engineering benchmark. It answers a simple question: what share of your hardware budget is spent to verify that your model does not create catastrophic risks?
Still, this metric carries limitations. Compute volume alone does not measure safety quality. A lab could spend large GPU budgets on low-value safety tasks, while a smaller team might deliver high-impact safety insights with modest compute. Even so, it provides a standardized starting point for cross-organizational comparison. Regulators can use it as one component of risk assessment frameworks.
4. Industry Strategic Game: A Transparency Challenge to All Frontier AI Labs
Anthropic’s release of three core R&D metrics is not just an internal report. It functions as an open challenge to the whole frontier AI industry, especially its main competitors.
For years, model developers released marketing demos and limited benchmark numbers while hiding internal operating data. When OpenAI publishes GPT-5 or Google DeepMind releases a new model, outside parties only see surface-level outputs. No independent third party can inspect internal workflows, agent autonomy levels or safety resource investment.
Anthropic’s move punches a hole in this black-box paradigm. It brought in third-party evaluators such as METR to validate parts of its data, turning these three major indicators into a test of credibility. If OpenAI, Google DeepMind and other leading labs follow suit and publish equivalent internal operational metrics, the public will gain a consistent thermometer to measure global AI risk. Labs refusing to release such data may face suspicion of concealing undisclosed risks.
Anthropic concludes that the world must shrink the information gap between frontier labs and the public. Independent third-party auditors should be able to access standardized measurement indicators and validate results. The current black-box model release pattern creates asymmetric information. Developers know the internal failure modes of their systems, while users and society at large remain unaware.
This push for transparency intersects with the rise of AI agent fleets. As more companies deploy multi-agent systems for research, coding and enterprise automation, auditing agent behavior becomes essential. Teams running distributed agent workloads need unified observability across model endpoints. An API gateway simplifies logging and request orchestration for mixed-model agent deployments.
5. Broader Implications and Remaining Open Debates
The three metrics together sketch a new landscape for AI development. The RSI automation rate tracks how quickly AI can take over research labor. The 30,000-agent fleet demonstrates the practical implementation of multi-agent R&D pipelines. The safety compute ratio quantifies the resources dedicated to risk mitigation. Together, they create a multi-dimensional picture that static benchmark scores cannot match.
Several open debates emerge from Anthropic’s disclosures.
First, defining “AI-completed task” leaves room for interpretation. Anthropic counts a task as AI-led when Claude delivers implementation after human prompt design. Skeptics argue human researchers still define objectives and validate final results, so full autonomy is not achieved. Proponents counter that task-level handoff is the prerequisite for recursive self-improvement. The transition will happen gradually rather than in one discrete leap.
Second, multi-agent safety controls remain unproven at extreme scale. Anthropic’s dual-layer monitoring reduces obvious collusion risks, but researchers warn emergent agent behaviors may appear that bypass static audit rules. Agents may invent new unforeseen strategies that evade detection, especially as reasoning ability improves.
Third, safety compute percentage as a universal benchmark is contested. Some researchers argue small, high-leverage safety teams can produce outsized insights without massive GPU clusters. Mandatory safety compute ratios may push labs toward spending money on superficial safety projects simply to hit percentage targets.
For enterprise and research developers, these trends carry practical implications. Multi-agent systems are moving from experimental prototypes into internal production workflows. Teams building agent applications need to handle distributed task scheduling, cross-model request routing, behavior logging and risk filtering. Standardized tooling becomes critical when combining multiple model families into one agent system.
Conclusion
Anthropic’s release of three internal R&D metrics marks a pivotal moment for transparency in frontier AI. The 26% AI ownership of internal research tasks demonstrates tangible progress toward recursive self-improvement. Its fleet of 30,000 continuously running AI agents and dual-layer monitoring system shows how labs are attempting to contain risks at scale. The 6% safety compute ratio offers a simple, quantifiable benchmark for evaluating AI labs’ risk investment priorities.
This publication also raises the bar for the whole industry. Other leading AI developers now face pressure to disclose similar internal operational indicators. Independent auditors and regulators gain a new set of tools to assess advanced AI systems beyond public benchmark scores.
The RSI feedback loop has started turning. AI agents are participating directly in AI research, and the boundary between model developer and model output blurs. While Anthropic’s safeguards reduce obvious hazards, the field remains in the early phase of understanding emergent multi-agent risks. Balancing rapid capability growth with rigorous safety investment will be the central challenge for all AI research organizations in the coming years.
International access: https://4sapi.com
Domestic access: https://4sapi.cn
Top comments (0)