Yesterday’s launches moved AI cost from price per token to price per finished task. Today follows that shift into buying, security, staffing and valuation.
- Effort sets the bill. Anthropic pitches Fable 5.2 as a manager for cheaper models, and effort settings now swing cost per task several-fold.
- Count the task. Meta’s Muse Ember matches rivals’ list prices but runs a typical agent task for about a third of Haiku 5.5’s cost.
- Buy through commitments. ChatGPT Enterprise is now in AWS Marketplace, so AI seats can draw down existing cloud spend.
- Guard the weights. Fireworks AI disclosed that attackers reached partial model weights and datasets for some customers.
- Hire for judgment. PM and forward-deployed engineer openings are rising while engineering and design demand stalls.
- Discount the code. Martin Alderson argues AI is turning software R&D from a lasting asset into a running cost.
- Score from your best customers. Clay ranks accounts weekly from its top customers’ traits and keeps a person on every send.
- Know your ratio. ICONIQ benchmarks put median ARR per employee at $237K and top growth companies at $655K.
This week, ask your team what a completed task, not a token, costs you.
Developments
TechnologyNew Angle
Fable 5.2, Day Two: Anthropic Is Selling It as a Manager for Cheaper Models, and the Effort Setting Now Drives the Bill (November 9–10, 2026)
Beyond the price cut covered yesterday, Anthropic launched Fable 5.2 with two routing modes: Advisor, where Haiku, Sonnet or Opus escalate hard problems to Fable, and Orchestrator, where Fable hands simpler subtasks down to them. Its own charts show Fable at High effort roughly matching Opus 5.5’s best coding score at about half the cost per task, while Max effort costs more than three times Xhigh and scored lower; an optional fast mode doubles the price to $16 and $80 per million tokens. Early users at Uber and Every describe unattended runs lasting 10 hours to a full weekend. Anthropic also says the model suspects it is being tested more often than recent models.
So what
The cost of frontier AI now depends less on which model you license than on how your team sets effort, speed and routing. Ask your engineering lead to set default effort levels, require approval for fast mode and Max effort, and route through cheaper models first, with Fable as the escalation path. Reserve overnight runs for large, well-specified projects with a named reviewer, and budget them like a contractor engagement rather than a seat license.
Hype check
Benchmark and cost-per-task figures are Anthropic’s own, published weeks before its expected IPO, and The Neuron notes its named outside testers did not review the launch materials.
Confidence: Established fact — modes, pricing and effort results are from Anthropic’s launch page; customer run times are vendor-quoted testimonials.
Go deeper Anthropic · The Deep View · Every Vibe Check
Competition
Meta’s Muse Ember Matches Rivals’ List Price but Runs a Typical Agent Task for About $0.14, a Third of Claude Haiku 5.5 (November 9, 2026)
Meta released Muse Ember, its smallest model, built for the repetitive steps inside agent workflows such as tool calls, routing, data extraction and long-document reading. List prices match Claude Haiku 5.5 and OpenAI’s GPT-6 Luna at $0.10 per million input tokens and $0.50 output, but Artificial Analysis estimates an average task costs $0.14 on Ember, $0.15 on Luna and $0.41 on Haiku, because Ember uses fewer tokens. Meta pitches it as the volume worker beneath its larger Muse Wonder, and The Information reports Meta is building an enterprise sales push through the big consulting firms.
So what
Per-token price is the wrong comparison: at identical list prices, cost per completed task here differs almost threefold. Ask for a cost-per-task view of your three highest-volume agent workflows and test a budget model on one of them this quarter. Before moving production work to Meta, check its enterprise data terms, support and uptime commitments, which are newer than its rivals’.
Hype check
Benchmark wins come from Meta’s announcement and launch-day Artificial Analysis runs; there are no independent production results yet.
Confidence: Established fact — launch and pricing confirmed by TechCrunch; cost-per-task and benchmark figures are early and should be treated as Watch this.
Go deeper Meta · TechCrunch · The Deep View
Go-to-Market
ChatGPT Enterprise Is Now Sold Through AWS Marketplace, So Seats Can Count Toward Existing AWS Commitments (November 2026)
OpenAI listed ChatGPT Enterprise in AWS Marketplace, letting organizations buy and manage subscriptions through existing AWS billing and apply eligible spending toward their AWS commitments, according to OpenAI’s announcement as summarized by The Neuron and TLDR AI. It extends the AWS relationship that began when OpenAI’s models arrived on Amazon Bedrock earlier this year. Labs are increasingly selling through the channels enterprises already use: The Information reports Meta is building a consulting-firm sales push, and Cisco said it will add Anthropic’s Claude Mythos to its AI Defense product.
So what
For buyers with committed cloud spend, AI seats can now skip a new vendor approval and draw down money already budgeted, which shortens sales cycles for whoever is listed. If you sell AI software to enterprises, confirm your marketplace listing counts toward customers’ cloud commitments, since that is becoming a deal requirement. If you buy, ask finance whether unused commitment can fund your 2027 AI pilots before approving new budget.
Confidence: Established fact — reported by two newsletters citing OpenAI’s announcement; the OpenAI page could not be opened directly.
Go deeper OpenAI · The Neuron
Legal & Regulatory
Fireworks AI Says Attackers Reached Billing Data and, for Some Customers, Partial Model Weights and Datasets (Updated November 9, 2026)
The AI hosting company updated its disclosure of an October incident in which attackers used stolen internal credentials. They accessed resource metadata, billing addresses, invoices and credit usage, plus some secrets and, for certain customers, incomplete model weights and datasets; Fireworks says card numbers were not stored and its zero-data-retention inference products were unaffected. It ties the breach to an “LLMjacking” group that resells stolen model access, and the Rundown, citing Forbes, notes the update came days before a planned raise at a $30 billion-plus valuation.
So what
Fine-tuned models and training data are now assets attackers go after, and they often sit with smaller infrastructure vendors. Ask each AI hosting vendor where your weights and datasets live, whether zero data retention is available, and how fast they will notify you of credential misuse. Rotate any API keys or secrets stored in third-party evaluation or fine-tuning jobs, and add these questions to vendor security reviews.
Confidence: Established fact — from Fireworks’ own security bulletin; the number of affected customers is not disclosed.
Go deeper Fireworks security bulletin · The Neuron
Perspectives
People & Organization
Lenny Rachitsky’s Job Market Data: PM Openings at a Four-Year High, Engineering Openings Flat, and Forward-Deployed Engineer Roles Up 4.6x (November 10, 2026)
Using TrueUp openings at more than 9,000 tech companies and Live Data Technologies hiring records, Rachitsky finds open PM roles up 75% from the March 2023 low, while engineering openings grew just 2.5% in the past year. U.S. software engineer hires fell to about 113,000 over 12 months, under half the 2022 peak, and design openings hit a record low at 3.8 open PM roles per design role. Forward-deployed engineer openings rose 4.6x to about 1,500, partly from relabeled sales engineer jobs, and “LLM” is now the most-mentioned skill in PM job descriptions.
So what
Demand is moving toward people who decide what to build and people who make AI work at the customer’s site. Review 2027 headcount plans through that lens: fewer narrow specialists, more product judgment and customer-facing technical talent. If you sell complex AI products, start hiring or developing forward-deployed engineers now, because that talent pool is tightening quickly.
Confidence: Informed perspective — proprietary job-posting and hiring datasets with the author’s interpretation; openings are not the same as hires.
Go deeper Lenny’s Newsletter · Lenny’s Newsletter home
Business
Martin Alderson: AI Is Turning Software R&D From a Lasting Asset Into a Running Cost (November 8, 2026)
Alderson argues software’s high margins came mostly from a scarcity of people able to build it, and agents are eroding that scarcity. Building is now about ten times faster, he says, while sales and marketing are as slow as ever; copying an app from screenshots is getting easier, and agents can pull customer records out of an old app even without a clean export. His conclusion: past engineering spend looks less like durable intellectual property, and acquisition premiums justified by “two years and 30 engineers to rebuild” should shrink.
So what
For boards and investors, this challenges any valuation that leans on the code itself as the moat. Pressure-test your own story: if a well-funded rival could rebuild your product in a quarter, what keeps customers (data, workflow depth, distribution, trust)? Put that answer at the center of your 2027 strategy and your next fundraising narrative, and weigh it in any build-versus-buy acquisition.
Confidence: Informed perspective — a practitioner’s argument with illustrative examples, not market data.
Go deeper Martin Alderson · TLDR Founders
