Product & Strategy
The Product Desk
Gemini 4 Argon's per-task cost doubles to about $4 after Nov 30, passing GPT-6 Astra.
The $1.99 a task that puts it level on the Intelligence Index is calculated at $2/$10 per million tokens. After the cutoff that rate becomes $4/$20, and at the same token mix the cheaper option ends up costing more per task than the rival at roughly $3.30. A model-swap case built this month rests on a price that lasts about six weeks, so the comparison you take to finance should be run at both rates.
In Play
Argon's Promotional Price Window
Artificial Analysis rates Gemini 4 Argon, generally available since Oct 14, level with GPT-6 Astra on its Intelligence Index at $1.99 per task. That figure uses $2/$10 per million tokens, a rate that doubles to $4/$20 after Nov 30. Any model-swap case your team builds this month uses prices that expire in about six weeks. Reviewers also disagree on quality. Cursor puts Argon on par with Opus 5.5, while one tester logged a false-claim rate above 50% on codebase analysis.
Ask ClarityDecision Models Split Off the Frontier
Microsoft launched Decision-1 on Oct 13 for routing, classification and verification. Computerworld reports AWS, OpenAI and Cloudflare are adopting the same bounded-decision layer. These models return a choice plus a probability, which makes them cheap gates for agent actions. One test agent built by Claude Code shows the risk: all 46 of its tests passed, yet a fallback misattributed confidence in a way that could authorize a tool call. Microsoft has published no benchmarks for its efficiency claim.
Ask ClarityAI ROI Still Measured in Hours Saved
A Wharton and GBK Collective survey of 1,400 managers finds only about 27% of firms track revenue from customer-facing AI, while 40% measure hours saved. The share expecting AI budgets to rise substantially next year fell to 10% from 19% in 2025. Your 2027 funding will follow features that can show growth. Raw usage dashboards won't make that case, because Visier finds 48% of workers exaggerate their AI use.
Ask ClarityAgents Meet Locked Doors and Echo Attacks
The Information reports JPMorgan said on its Oct 13 earnings call that it will not let 'just any outside agent' into client accounts. CSO Online separately reports researchers bypassed six agent frameworks with a 'Repeat-After-Me' prompt and verified 253 bugs across 118 open-source MCP servers. If your agents act inside customers' systems, they now need a permission model a risk committee will accept. Sublime also ties part of a roughly 33,000% rise in calendar-invite phishing to agents that auto-accept invites.
Ask ClarityAI Vendors Look Less Durable
Character.AI filed for Chapter 11 on Oct 12 with Disney as first bidder at $20M. Its owner warns the platform shuts down if no buyer keeps it running by the sale's late November or early December close. The Pragmatic Engineer's excerpt of Sam Newman's new book offers Cursor as the warning case: when Anthropic changed its pricing in 2025, Cursor passed a significant cost increase on to its customers. Fallback models and degraded modes now belong in your plan. Capacity is tight too, with The Information counting $531B of Anthropic compute deals in 11 months.
Ask Clarity
Deep Dives
- ●
Argon's Lead Is Real, but Its Price Expires on Nov 30
On paper Argon is the cheapest frontier-class option, but that may not hold at December prices. Its weakest skill is also the one most AI features rely on: stating facts.
What the price change does to the ranking The per-task numbers move once the price increase is applied. Artificial Analysis measures $1.99 per task at the $2/$10 rate. If the token mix holds, $4/$20 puts Argon near $4 per task…
3 action items
- ●
Cheap Decision Models Hand You a Probability, and Your Code Turns It Into Permission
Moving routing and classification off the frontier model saves money, but authorization then sits in fallback code that tests rarely exercise.
How a missing action got a 0.90 confidence An incident agent pages the team on a number that belonged to another action. Daily Dose of Data Science had Claude Code build the agent from one prompt: 30 files and 3,383…
3 action items
- ●
Workday Will Bill for Agent Actions Before Buyers Can Measure What They Return
Metered agent pricing makes every action a line item for finance teams. Those teams still judge AI by hours saved and by usage numbers that workers admit they pad.
The bill arrives before the proof Picture a finance analyst at a Workday customer opening an invoice with a new line for agent actions. That line traces to Workday's Oct 13 announcement, the clearest sign of where enterprise AI pricing…
3 action items
The edition continues
Take the signal into the room.
Sign up or log in to read all 3 deep dives in full, plus the final take.
Read the full editionContinue with LinkedIn