Intelligence Track

Design

9 July 2026

The Brief

The result everyone will repeat from today's model benchmark — that OpenAI's newest model edges out its rival — is 70% one person's taste and its winner flips depending on whether you're writing a PRD, prototyping, or debugging, which makes the ranking the least useful thing in it; the portable lesson is that you now build your own eval and route models per task. That reframe rhymes with the rest of the day. An economist's case that frontier models drift toward commodity margins, with value captured up the stack the way mobile carriers built a trillion-dollar business and flat stock, says the model is the cheap, interchangeable part. An Instagram reorg into four-to-six-person generalist pods and a new 'product staff' role treats judgment, taste, and curation as the one thing AI can't absorb. And two travel-money moves — a card-first fintech competing on the monthly statement instead of the fare, and a leisure platform giving away corporate booking to own the SME's future leisure trips — run the same play on distribution: own the relationship, not the transaction. The commodity keeps climbing a layer — from room to ride to model to execution to content — and value follows whoever owns the layer that directs it. For anyone sitting between commoditized supply and a commoditizing model, the question is which layer neither side can absorb, and how quickly the window to own it is closing.

Lenny's Newsletter · 9 Jul 2026

GPT-5.6 Sol vs. Claude Fable: Why OpenAI’s new model crushes my benchmark

A personal 'How I AI' evaluation — weighted 70% the author's taste, 30% Terminal Bench 2.1 — ran GPT-5.6 (Sol, Terra, Luna), Claude Fable 5, and Sonnet 5 across PRDs, prototypes, wireframes, debugging, and agentic voice; Sol took the weighted index overall, but the per-task winner flipped, with Terra preferred for PRDs and Sonnet 5 for debugging and agentic voice. Which models were even accessible shaped the practical verdict as much as the scores did.

Industry lens

Once the top-scoring model reaches general availability rather than limited preview, does the per-task winner ordering hold up under independent evaluation, or do access and cost — not benchmark rank — decide which model teams standardize on?

Reading as

Adam Mosseri: AI is a tailwind for authenticity

Behind a headline about AI content and authenticity sits the sharper disclosure: the canonical product team is collapsing from roughly thirteen-person specialist groups into lean four-to-six-person generalist pods, with a new 'product staff' role fusing PM, design, data science, and research into one operator, and taste, curation, and strategy named as the inputs AI does not cover. The authenticity claim is the flip side — as synthetic content floods feeds, audiences are expected to seek verified humans over volume.

Why it matters

If the canonical team consolidates into generalist operators and judgment and curation become the protected core, hiring rubrics and leveling should select for cross-functional range and taste, and narrow-specialist headcount becomes the role most exposed.

Industry lens

Do other large product organizations that consolidate into generalist 'product staff' pods retain decision quality over the next few cycles, or does the specialist depth they shed resurface as degraded product judgment?

Lenny's Newsletter·9 Jul 2026

Ways to think about token pricing

Token prices sit in a transient supply crunch, and the case made is that every visible dynamic points toward foundation models becoming low-margin commodity infrastructure with value captured up the stack — the mobile-data parallel being a trillion-dollar industry whose carrier stocks went nowhere. A quieter point does the heavier work: the current crunch is driven almost entirely by one use case, software development, and today's infrastructure could not serve a consumer use case with hundreds of millions of daily users at any price.

The binding constraint on scaling a consumer-facing agentic feature may be capacity rather than model quality or even price, so a product betting on always-on AI agents at consumer scale should stress-test availability and cost against a world where a single consumer product-market-fit event exhausts inference supply.

Benedict Evans·9 Jul 2026