Anthropic's AI R&D Just Doubled Using Internal Models

Let's unpack the METR assessment of Anthropic, because this is one of those stories that sounds like inside-baseball but actually tells us something enormous about where the entire AI industry is heading. **Technical Deep Dive** METR, the nonprofit evaluator that's become something like an independent auditor for frontier AI labs, spent eleven weeks with direct access to Anthropic's internal data and staff interviews. Their conclusion: Anthropic's AI research and development capability — specifically, how fast the company improves model performance on well-defined tasks — is running somewhere between 1.

3 and 3 times faster than its 2024-2025 pace, with a central estimate around 2x. Crucially, METR attributes roughly 80 percent of that acceleration to AI itself doing the work — not more headcount, not more compute alone, but Anthropic's own models increasingly designing experiments, writing training code, and iterating on their successors. The report also revealed the existence of at least two internal models ahead of Anthropic's publicly available Claude Opus 5.

5, including one internally called "Model 4" that cleared 50 percent on all fifteen benchmarks METR tested it against. That's a meaningful detail: the public-facing Claude lineup may already be one or two generations behind what's running in Anthropic's own labs. METR's own researcher, Ryan Greenblatt, publicly said he was "seriously wrong" in his earlier, much slower projections for this kind of speedup — which is a notable admission from someone whose entire job is calibrating these estimates.

Anthropic, for its part, redacted much of the underlying evidence, with co-founder Sam McCandlish saying the goal was avoiding "handing out a roadmap for others to go faster." That tension — wanting independent verification without accelerating competitors — is itself a sign of how seriously labs now take this recursive self-improvement question. **Financial Analysis** This news lands at a financially delicate moment for Anthropic.

As we just covered, major IPO investors are already pressuring the company to cut its expected valuation from $1.5 trillion toward something closer to $1 trillion, with the total raise potentially shrinking by $20 billion. On one hand, evidence of a genuine AI-driven R&D flywheel is exactly the kind of story that justifies a premium valuation — it suggests Anthropic's cost structure could improve dramatically as AI increasingly does the engineering work humans used to be paid for.

Remember, Anthropic's S-1 showed $33.24 billion in revenue against a $104 billion net loss; anything that improves the ratio of output to spending is financially significant. On the other hand, this is a double-edged financial signal.

Anthropic's own IPO filing explicitly defined "dramatically accelerated progress attributable to AI" as a danger threshold worth flagging to regulators and investors — and Monday's update suggested that threshold may already be close. That's not typically language that makes IPO investors comfortable in an environment where they're already nervous about data center buildouts following the midterm election backlash. Investors have to weigh whether faster internal R&D translates into faster monetizable products, or whether it mainly front-loads safety and governance costs that eat into margins before any revenue shows up.

**Market Disruption** If Anthropic really has a 2x internal R&D multiplier running on AI-driven iteration, every other frontier lab has to assume their rivals have something similar, or will shortly. This changes competitive dynamics in a few concrete ways. First, it raises the bar for what counts as a "model generation" — if Anthropic already has a Model 4 outperforming its public Opus line, the public release cadence investors and customers see may increasingly lag well behind the real frontier, making it harder for competitors like OpenAI and Google DeepMind to calibrate what they're actually racing against.

Second, it puts pressure on smaller labs and open-weight projects. We already heard from Joanna that DeepSeek's open-weight model is closing the capability gap with closed frontier systems at a fraction of the cost — if closed labs like Anthropic are simultaneously accelerating their own internal pace using proprietary AI tooling, the gap between frontier-closed and frontier-open may actually widen again just as open models were catching up. Third, this reinforces compute as the real moat.

An R&D flywheel driven by AI models doing AI research still needs enormous compute to run those experiments, which ties back directly into yesterday's story about data center expansion facing political headwinds after the midterms. If growth in research velocity depends on growth in compute capacity, and compute capacity growth is now politically contested, that's a genuine structural bottleneck worth watching. **Cultural & Social Impact** For the public, this story lands differently than a typical product launch, because it's not about a new chatbot feature — it's about the rate of change itself.

When an independent, credentialed evaluator says a company's AI is already meaningfully accelerating its own development with limited bottlenecks, it validates a concern that's been circulating in AI safety circles for years: that progress could become self-reinforcing before society, regulators, or even the companies themselves are ready. That's part of why an Arizona congresswoman, fresh off the midterms, is already calling for a pause on advanced AI and a ban on superintelligence as a top legislative priority. Whether or not that specific policy goes anywhere, it signals that "AI improving AI" is moving from a theoretical safety scenario into something politicians feel compelled to respond to.

For everyday users, the more AI capability that gets built without full visibility — because Anthropic redacted its strongest evidence — the wider the gap grows between what's publicly available and what's actually possible behind closed doors. That erodes a certain kind of public trust: people are being asked to evaluate AI's trajectory based on products that may be one or two steps behind the real frontier. **Executive Action Plan** So what should business leaders actually do with this?

First, treat internal AI R&D acceleration as a competitive signal, not just a safety headline — if labs are genuinely compounding capability gains faster than public releases suggest, your AI vendor roadmap conversations should include direct questions about what's coming in the next two quarters, not just what's shipped today. Second, build in governance review cycles now rather than waiting for regulation to force your hand; given that even Anthropic is flagging its own danger thresholds internally, enterprises relying heavily on frontier models should have an internal audit process for capability jumps, similar to how finance teams audit for material changes. Third, diversify your technical bets across both closed frontier providers and the fast-improving open-weight ecosystem Joanna highlighted — with cost-per-task falling as low as two dollars for some open models, locking into a single closed vendor at today's pricing is a riskier long-term bet than it was even six months ago.