🏆 THE AGENT REPORT — WEEKLY DIGEST

Hi AI builder,

Here are the top 5 AI Agent stories from this week:

1. Anthropic's $16.8 Billion Quarter: The Arithmetic Behind a $110 Billion Run Rate Anthropic told investors Q3 revenue hit $16.8bn, up 45% quarter over quarter, and that it is now cash-flow positive — with $1.70 of revenue per $1 of compute, against $0.63 in 2025. Analysts at the briefing extrapolated a $110bn run rate by year-end, which quietly demands 64% sequential growth in Q4. A November listing near $2tn is now expected, sitting on top of at least $518bn in decade-long compute commitments.

2. Anthropic: AI Worms on $55 Smart TV Boxes Are Months Away Anthropic's Frontier Red Team concluded that models at 30B parameters and under can already execute each individual step of autonomous replication, and that a small model should clear the full capability floor before the end of 2026. Their proof of concept sharded a 26B model across 15 cut-price Android TV boxes bought for about $55 each. Scaled to a botnet like Kimwolf — over 2 million infected boxes — the fleet would serve roughly 80,000 tokens per second with no central point to switch off.

3. The $20 Billion Hole in OpenAI's Revenue Projections OpenAI told investors its annualized revenue was approaching $50bn at the end of September, roughly $20bn below the $68–70bn figure circulated a week earlier. Most of the gap is a measurement difference — grossed-up partner sales versus OpenAI's own net share — not vanished business, but the market read it as a demand signal anyway: Nvidia fell 3%, Oracle nearly 6% and CoreWeave nearly 8%. OpenAI still targets $70bn annualized by year-end, with its IPO pushed to 2027 at the earliest.

4. 816B, 560B, 120B in One Day: The Open-Weight Wave Now Hits a Memory Wall Three large open-weight models landed inside 24 hours: Meta's Muse Spark Open (816B total, 42B active, 1M-token context), InclusionAI's Ling 3.1 Flash (560B/25B, MIT) and NVIDIA's Nemotron 3.5 Super VL plus Bolt 74B-A9B. All are sparse mixture-of-experts running 4.5% to 12% of their parameters per token, so compute is now mid-sized — memory is not. Muse Spark Open wants a 1.5 TB BF16 checkpoint across four 8×H200 nodes, Ling wants 8 H200s at full context, and Bolt fits in 24GB, which tells you exactly where self-hosting actually lives now.

5. 741% More Code, 20% More Software: The Review Bottleneck The viral "741% more code, 20% more shipped software" chart blends two separate 2026 studies: the 741%/20% pair comes from an NBER paper on developers, while the 700-firm review-bottleneck evidence is a different Harvard study. The 741% counts lines of code, the weakest productivity proxy there is — the firm-level study measured a far smaller +30% and found no statistically significant gain in completed work. Both converge on the same mechanism: human code review absorbs the gains, lengthening the merge queue by about 49%, and AI review tools barely dent it — 80% of firms used them by March 2026, yet humans still wrote 77% of review comments.

📖 Read the full coverage: https://the-agent-report.com/latest/ 💌 Subscribe: https://buttondown.com/theagentreport See you next week! — The Agent Report