The thing most readers get wrong is that model releases are still platform choices
Claire Vo, host of How I AI and founder of ChatPRD, spent hours testing Anthropic's Opus 4.8 in early access. She found it excellent for greenfield prototypes and one-shot features. But when she turned to data-heavy strategy and roadmap work—the kind that decides what gets built next quarter—she kept reaching for Opus 4.7. Not because 4.8 wasn't impressive. Because model selection had stopped being a platform decision and become a per-task, per-week calculation that no roadmap had accounted for.
On May 19, 2026, Google held I/O and launched seven distinct products in a single keynote. Gemini 3.5 Flash. Antigravity 2.0 with projects, scheduled tasks, and subagents. AI Studio with Workspace integration. The consumer Gemini redesign. Omni for video generation. Flow for cinematic editing and character consistency. Stitch for streaming UI design. Pomelli for brand identity. Each arrived with different capabilities, different availability windows, and what Vo called "the launch-to-availability gap"—the problem where keynote demos don't map to what builders can actually ship.
When a model that benchmarks better still loses on certain task types, you can't pick a vendor and move on.
The collapse creates three operational problems that PMs don't have instrumentation for. The first is model selection as a weekly decision. Vo's experience with Opus 4.7 versus 4.8 isn't an edge case—it's the new baseline. When a model that benchmarks better still loses on certain task types, you can't pick a vendor and move on. You need to instrument which model handles which work, then re-evaluate when the next release drops the following Tuesday. That's not a technical decision. It's a product operations problem that requires tracking model performance against user retention, task completion, and cost per interaction.
The second problem is capability drift mid-build. Google's I/O blitz illustrates the planning trap. Seven products launched in one keynote, but availability lagged. Antigravity 2.0's slash commands—including a /grill-me feature that Vo described as potentially a more aggressive alternative to Claude Code's clarification flow—shipped on day one. But when Vo tested other announced features live during her I/O recap, she hit access issues. AI Studio's Workspace integration, positioned to own internal productivity use cases, wasn't universally available.
Flow's avatar creation worked, but character consistency—the feature that would make it viable for production video—produced different hair lengths, inconsistent backgrounds, and emotional expressions Vo described as "100% uncanny valley." A PM who wrote a feature spec in March based on the I/O preview would be managing expectation drift by May, not shipping.
The third problem is the complete absence of published retention data. Vo's testing methodology—building prototypes, running strategy exercises, generating video avatars—surfaces what works in a single session. But product retention is a 30-day question, and the models didn't exist 30 days ago. That means every feature decision is a bet with no backward-looking data. When Gemini 3.5 Flash launched with benchmarks showing it rivals flagship models at four times the speed, the pitch was clear: frontier intelligence at Flash-series latency, often at less than half the cost of comparable models. But cost per token isn't the same as cost per retained user.
This isn't a capabilities gap. It's a planning horizon collapse. Quarterly roadmaps assume the substrate is stable. When Vo opens Opus 4.7 for strategy work instead of 4.8, she's making a substrate decision in real time. When Google launches seven products with staggered availability, it forces adoption decisions before the tools are fully usable. When no team shares retention data on model-driven features, every PM is running the same experiment in parallel with no shared learning.
The forward question, the one that determines 2027 headcount planning, is whether automation means more work or fewer humans. Benedict Evans, the independent analyst and former Andreessen Horowitz partner, framed it precisely: "The right question about your job isn't what percent can AI do, but is this a task or a job?" A task is atomic. A job is a bundle of tasks plus judgment about which tasks matter. If AI automates tasks, you still need humans to decide which tasks to automate and in what order. But if AI automates jobs—not just writes the code but decides what code to write—the planning model changes.
Dan Shipper, co-founder and CEO of Every, a media and software company of about 30 people who are AI early adopters, sees the paradox playing out in real time. His company uses AI extensively, yet he's not cutting headcount. Instead, he's hiring forward-deployed engineers—the role that sits between the AI and the team, translating business problems into agent instructions. That's the tell. If AI were purely substitutive, Every would be shrinking. Instead, it's adding a new role to manage the operational tax of working with models that change weekly.
The operational tax is hidden because it doesn't show up in benchmark scores or pricing sheets. It shows up in the time PMs spend re-evaluating model choices, the sprints lost to capability drift, and the inability to forecast retention on features that didn't exist when the roadmap was written. When Vo says she's "still reaching for Opus 4.7 on data-heavy strategy," she's describing a workflow that no PRD template accounts for. When Google launches seven products in one keynote, each with a different launch-to-availability gap, it externalizes the integration cost to every team trying to build on top. When Evans asks whether AI automates tasks or jobs, he's pointing at the data PMs need to answer the 2027 headcount question—and that data doesn't exist yet because the models are too new.
The missing framework is the operational one. Not which model is best. Not how to monetize AI features. Not whether your job is safe. The question is: how do you build a product when the substrate changes weekly, and what does it mean that no team has shared a working model-selection or retention framework yet?
The answer, right now, is that teams are operating without one. Vo's testing methodology—try the new model, compare it to the old one, decide per task—is manual and doesn't scale. Google's launch strategy—announce everything, ship availability in tranches—pushes the coordination cost downstream. Evans's framing—task versus job—identifies the strategic question but doesn't provide the instrumentation to answer it. And Shipper's hiring pattern—add forward-deployed engineers—suggests the solution is more humans, not fewer, at least until the operational layer stabilizes.
The substrate is moving. The frameworks haven't caught up. And no one has published the playbook yet.