A house newsletter · daily · researched and written by the Kurzia AI
Saturday 7 November 2026 · Issue 6 · 3 min read
AI bills stall as usage grows; Vercel's 10-to-1 lead agent; Murati ships the next open-weight frontrunner
The main story
TechCrunch · yesterday
Ramp data: AI bills flatlined while usage grew
Takeaway: $600 million off revenue growth
Ramp's data, reported by TechCrunch on November 6, shows AI token spending grew just 1.5% across August and September, against 42% growth in July — usage rose while bills didn't. Price cuts did it: OpenAI halved API prices for two GPT-6 models in September, and Anthropic said in a filing its Fable 5.1 and Opus 5.5 cuts took about $600 million off revenue growth. Almost all AI spend comes from venture-backed tech companies, per Ramp's lead economist, and the heaviest spenders stalled first. October rebounded — up 8.9%, per Ramp — and Bloomberg reported Anthropic's expected valuation has slipped to $1.2–1.3 trillion. Cheap tokens help every company buying them; the hurt lands on AI's own investors. Tell any founder you back who builds on these APIs: lock today's prices into a contract, and don't build a plan on ever-rising AI spend.
New rounds
Physicl — synthetic 3D environments and training data for robot and world-model builders; seed; $10 million; led by Inovia Capital; reported by TechCrunch.
What the AI labs did
Claude Fable 5.1 — Anthropic's new Fable 5.1 costs an estimated 25% less than Fable 5 for typical workloads — up to about 45% for agent-heavy use — with cache reads at $0.25 per million tokens ($10 in, $50 out). Its new Enterprise Frontier Safeguards give enterprise customers zero-data-retention privacy with data stored in their own cloud, rolling out in phases this fall. That trims costs for startups building Claude-based agents and undercuts vendors selling privacy wrappers. Founders on Claude should re-run their unit economics this month. Anthropic
Inkling 2 Small — Mira Murati's Thinking Machines Lab released Inkling 2 Small on November 5: 594B parameters with 17B active, scoring 31 on the Artificial Analysis index. Weights are due under Apache 2.0 next week after external safety testing; per Artificial Analysis, via Trending Topics, it would then pass Mistral Large 4 as the strongest open-weight model outside China — squeezing startups selling open-weight hosting. Open-model founders should benchmark against it when the weights land. Thinking Machines Lab
For your portfolio
SaaStr · yesterday
Vercel's qualification team went from ten people to one
Takeaway: 93% of all cases
On November 6 SaaStr published Vercel COO Jeanne DeWitt Grosser's account of running go-to-market on internal agents: an agent built with 20% of one engineer's time now does the work of a ten-person inbound lead-qualification team, and the in-house support agent handles 93% of all cases.
The lead agent costs under $5,000 a year in tokens and infrastructure; the support agent cost about $150K a year all-in and took three engineers two months. When a customer's own team can prototype your product over a weekend, per-seat pricing for SDR or support automation stops holding. Ask every B2B founder you back: if a customer's engineer can rebuild what you sell in a weekend, what do they still pay you for — and does your product have the APIs and webhooks to survive Vercel's rule that tools their agents can't call get dropped?
“Our best AI SDR reviewed every single output to improve the flywheel. And then finally, she couldn't improve it anymore.”
The AI News Report · yesterday
Ask HN thread on coding-agent setups draws 663 comments from 530 developers
Takeaway: Claude Code leads with 180 users
An Ask HN thread from November 4 drew 663 comments; a November 6 analysis counted the tools named by 424 of the commenters: Claude Code leads with 180 users, Codex has 103, and the open-source agents OpenCode and Pi each have nearly twice Gemini's users.
148 of the 424 run open-weight or local models such as DeepSeek or Qwen, and 120 use both an Anthropic and an OpenAI product — multi-vendor is the norm, not the exception, which caps what any single model or per-seat tool can charge developers. Ask the engineering founders you back how their teams actually run agents: if the answer is one assistant per developer rather than parallel agents in separate git worktrees with a second model reviewing every change, they are behind what the thread's best setups already do.