Anthropic opens the lab’s R&D pace metrics.
Anthropic opens measurements for the pace of AI development inside frontier labs
As of August 2026, on Epoch AI’s automation ladder: Claude is not fully autonomous for any measured AI R&D subset; Claude “leads” 26% of Anthropic’s AI R&D; the share at or above “AI collaborates” is above 90%. Roughly 30,000 agents run research/engineering work at once on the most-used internal platform — 100% of those actions pass online monitors before execution and offline monitors after; over a billion August decisions, 0.002% (~1 in 47,000) were blocked online.
Compute snapshot July 13–20: about 6% of AI-R&D compute and about 12% of AI-driven AI-R&D compute classified as safety (conservative — dual-purpose work counted as R&D). Anthropic plans embedded third-party evaluators with internal-comparable access to verify practices and metrics.
Why it matters: operators and policymakers get named, publishable gauges for recursive R&D automation and agent oversight — treat every figure as vendor self-report until third parties re-run the methods.
Sources: measurements · newsroom
| Model / package | Org | Notes |
|---|---|---|
| Ollama v0.34.3-rc1 ~8:02 PM ET Sep 18 |
Ollama | RC thinking /api/show · Nemotron H MLX |
| llama.cpp b11064 Sep 20 ~11:19 AM ET |
ggml | Nightly Metal dsv4_hc_pre arbitrary hc |
| llama.cpp tip b11070 Sep 21 ~6:49 AM ET |
ggml | Tip Hexagon DMA64 |
| SGLang v0.5.20 ~6:41 PM ET Sep 18 |
LMSYS | GA new models · sampling masks |
| mlx / mlx-lm / mlx-vlm Standing |
MLX | Quiet 0.32.2 / 0.31.3 / 0.7.1 |
| Item | Beat | Status |
|---|---|---|
| Anthropic R&D pace metrics Sep 17 |
26% leads · 30k agents · 6% safety compute | Policy |
| Accenture embedded eval Sep 18 |
≥$1B / 5y each · inside-lab access | Watch |
| AI slowdown antitrust suit Sep 19–20 press |
NDCA filing — secondary coverage | Other |
| Speakeasy OpenAPI OSS Sep 17 |
AGPLv3 generators · MCP docs servers | Active |
Critical · Pace metrics + RC pins
Treat Anthropic’s 26% / 30k / 6% as lab-reported until embedded evaluators publish. Do not ship production on Ollama v0.34.3-rc1 without an explicit RC accept — wait for GA badge + assets.
Watch · Embedded eval / lawsuit
Accenture-inside-Anthropic is the first named embedded evaluator; standards and reporting norms still unsettled. Slowdown antitrust suit is Other news until a docket or company response changes operator obligations.
Opportunity · Metal K3 / SGLang 0.5.20
llama.cpp ≥b11064 if you hit Metal CPU fallback on Kimi-K3 dsv4_hc_pre. SGLang ≥0.5.20 for Flash-Next / Hy4 / sampling-mask RL rollouts. Optional: Speakeasy generators for typed agent SDKs (AGPL on the tool).
Probe Ollama thinking via /api/show (RC)
On a throwaway install of ≥v0.34.3-rc1, curl /api/show for a thinking model and confirm thinking.values / default before promoting to a pinned GA.
Metal Kimi-K3 path: llama.cpp ≥ b11064
Pull a nightly ≥b11064 so Metal dsv4_hc_pre accepts non-4 hc instead of falling back to CPU on Kimi-K3-style residual stacks.
Issue 21 · Monday, September 21, 2026 · Not investment advice · Sources cited inline · Other news: slowdown antitrust suit (secondary) · Issue 20 OpenAI misalignment / LSVP / Ollama 0.34.2 / Agent Anomaly / Chrome Compliance / b11029 not re-led
