You Can't Compete on Cheap Models Anymore

YouTube Summary · 06 Jul 2026

VIDEO

⚡ TL;DR

EXECUTIVE SUMMARY

AI made execution cheap, so value moved upstream to imagination — the human ability to pose questions no backlog contains. Mitchell Hashimoto's $40 frontier-model job outperformed what cheap models and even he could do, because he imagined a task nobody had thought to ask. Cheap execution is table stakes; the 10x multiplier comes from frontier imagination grounded in deep model familiarity and real context. You need both layers — cheap execution as the engine, frontier imagination as the steering.

Source: YouTube — You Can't Compete on Cheap Models Anymore · Duration: 15:35 · Creator: AI News & Strategy Daily (Nate B Jones)

🎯 Key Highlights

Hashimoto's experiment — Fable 5 ($9, 40min) tied cheap models ($1) on ordinary tasks, but on a gnarly systems-code optimization the cheap models couldn't touch, it hit a performance level Hashimoto himself couldn't reach — for $40 in 2 hours.
The task nobody assigned — The $40 job wasn't on any backlog, sprint, or PM priority. It existed only because an expert with thousands of hours inside models suspected something new was possible and spent money to find out.
Value moves upstream — When execution commoditizes, value doesn't disappear; it moves to whoever decides what's worth executing. The cheaper execution gets, the more valuable each frontier question becomes.
Blackberry vs Apple — Comparable execution muscle, but Apple imagined a different answer to 'what a phone even is.' Imagination set the 100x multiplier, not execution quality.
The imagination test — Has your task list changed in 12/6/3 months? If you're doing your old list faster and cheaper, you have an imagination shortage, not a tooling problem.
Imagination ≠ artistic gift — It's fingertip awareness of where capability lines have moved, built from hundreds of hours inside models. You can't imagine with capabilities you haven't touched.
Factory electrification analogy — Bolting electric motors onto steam-era factory layouts yielded barely any gain. The payoff came when managers redesigned the building around what cheap distributed motors made possible. AI is the same.
Stripe's 50M-line migration — One day to migrate 50M lines, but the real story is years of building review systems and task coverage that could verify changes at that speed. They redesigned the building first, then harvested frontier value.
You can't hire imagination — An external 'AI visionary' has imagination but no context. Imagination only fires next to context. The job is manufacturing it: give context-holders tools, permission, and budget to make bets.
Fable 5 blackout survival — The model disappeared for 72 hours, but the imagined questions and redesigned workflows survived. When it came back, the people with imagination just continued. The asset was never the model.

⏱ Topics by Timeline

00:00 The sameness problem — better tools, dropping prices, but everything looks the same. Value didn't disappear; it moved.

01:45 Hashimoto's model test — Fable 5 ($9) vs GLM 5.2 ($1) vs GPT 5.5 ($1.50) on ordinary work: all tied. Cheap looks like a rip-off of expensive.

02:30 The $40 frontier job — optimizing gnarly systems code. Cheap models couldn't touch it. Fable 5 reached a level Hashimoto himself couldn't hit.

03:30 Who assigned that task? — Not a backlog, sprint, or PM. It came from an expert who suspected something new was possible. AI can only do work someone has imagined.

04:25 Two-layer stack — cheap open execution as the engine, frontier imagination as the steering. They're not in competition; the cheaper execution gets, the more valuable frontier questions become.

05:20 Blackberry vs Apple — same industry, comparable execution. Imagination set the 100x multiplier. The $40 job had no competition because nobody else had thought of it.

06:15 The imagination test — has your task list changed in 12/6/3 months? Old list + faster = imagination shortage, not AI transformation.

07:00 Imagination = fingertip awareness — not artistic gift. Hundreds/thousands of hours inside models. You can't imagine with capabilities you haven't touched.

07:45 Fable 5 porch marketing example — mapping unshaded porches via Google Maps, 3D modeling, custom mailers. Frontier model for prototyping, cheaper models for production pipeline.

10:20 Personal practice — two-layer: cheap models for daily execution + dedicated scouting hours with frontier models. Are you taking scouting seriously?

11:00 Factory electrification — bolting motors onto steam layouts = no gain. Redesigning the building = payoff. The unit of change was the building, not the motor. AI is the same.

11:50 Stripe 50M-line migration in 1 day — the impressive number isn't the day; it's the years of infrastructure (review systems, task coverage, model-driving expertise) built beforehand.

13:00 Can't hire imagination — external visionary has imagination but no context. Imagination fires next to context. Manufacture it: give context-holders permission, tools, and budget to bet.

14:00 Fable 5 blackout — model gone for 72hrs, back into a price war. But imagined questions and redesigned workflows survived. The asset was never the model; it was the imagination.

15:00 Conclusion — spend frontier where it multiplies: real context, real bets, real questions that exercise technical imagination. No substitute.

🧠 Hermes Integration

💡 Formalize a Scouting Mode

Hermes already runs cheap models (GLM-5.2 cloud) for daily execution. The skill: create a 'scouting' mode — dedicated sessions with frontier models (Fable 5 / GPT-5.5-class) where the goal isn't execution but exploration: 'What can this model do that I've never been able to ask before?' Track these sessions separately from task execution.

✅ Actionable: Add a scouting prompt template + flag to Hermes that routes to frontier models with imagination-first instructions

💡 Imagination Audit via GBrain

The video's test: 'Has your task list changed in 12/6/3 months?' Hermes can automate this. Track the types of tasks/prompts M~ sends over time in GBrain as facts or timeline entries. Periodically surface: 'Your task patterns have shifted X% vs last quarter' or 'You're running the same 5 task types — imagination shortage alert.'

✅ Actionable: Log task-type metadata to GBrain facts; build a quarterly imagination-audit cron job

💡 Context-Injected Subagent Prompts

delegate_task subagents get context but the video's key insight is imagination fires next to context. Ensure subagent prompts for frontier tasks include: (1) the full project context, (2) explicit permission to explore beyond the stated task, (3) a 'what if' framing that encourages novel questions rather than just executing the brief.

✅ Actionable: Add an optional 'explore_mode' flag to delegate_task prompts that shifts from execution to imagination framing

💡 Capability Line Tracking

Maintain a GBrain page tracking where the model capability line has moved — what's newly possible with each model upgrade. This builds the 'fingertip awareness' the video describes. Each time a new model lands (Fable 5, GPT-6, etc.), log what it can do that previous models couldn't. This becomes the imagination fuel.

✅ Actionable: Create a GBrain 'capability-frontier' page updated on every model upgrade; query it before scouting sessions

💡 The $40 Question Permission Protocol

The video asks: 'Who on your team is allowed to pose a $400 question without asking anyone?' For Hermes as M~'s system, this means: should Zeus be empowered to spend frontier-model budget on exploration without explicit per-task approval? Define a scouting budget + auto-approval threshold for frontier model calls that aren't tied to a specific task.

⚠️ Worth discussing: Set a monthly scouting budget cap + auto-approve frontier calls under $X for exploration

😈 Critical Thinking & Devil's Advocate

Survivorship bias in the $40 story

We hear about Hashimoto's $40 job that worked. We don't hear about the 99 $40 frontier-model jobs that produced nothing useful. The video uses one anecdote to argue for a strategy. How many failed frontier experiments does it take before the 'imagination premium' doesn't pay off? The expected value calculation is missing.

'Imagination' is unfalsifiable

If a frontier job produces great results, you 'imagined well.' If it fails, you 'didn't imagine enough.' This is a convenient framing that can't be tested. The video never defines a measurable threshold for when frontier spending beats cheap routing. It's rhetoric dressed as strategy.

The frontier advantage window is closing

The entire argument rests on frontier models doing things cheap models can't. But cheap models are improving faster than frontier models are pulling ahead. The porch-marketing example? A $1 model will do that in 6 months. The two-layer stack may be a transient phenomenon, not a durable strategy. The video treats a moving target as a fixed insight.

Blackberry analogy is oversimplified

Apple didn't win on imagination alone — they won on execution + ecosystem (App Store) + timing + supply chain + brand. Reducing it to 'imagination set the multiplier' erases the dozens of other factors. This is the classic single-cause fallacy. Nokia also 'imagined' a different phone; it didn't save them.

Stripe's success is engineering, not imagination

The Stripe story actually contradicts the thesis. Their 50M-line migration worked because of years of infrastructure investment — review systems, test coverage, team expertise. That's execution discipline, not imagination. The video rebrands good engineering as 'technical imagination' to fit the narrative.

'Manufacturing imagination' is harder than it sounds

Giving context-holders 'permission and budget to make bets' sounds liberating, but real organizations have risk aversion, compliance requirements, budget cycles, and politics. The video hand-waves the structural barriers. Most companies can't just 'let people spend $400 on frontier questions' — there are approvals, audits, and accountability frameworks for good reasons.

Maybe the bottleneck is model limits, not human creativity

The video assumes humans have unlimited imagination and models are the constraint. But perhaps the real bottleneck is that models still can't do reliable multi-step reasoning, maintain context over long tasks, or handle ambiguous requirements. Framing it as a human imagination problem lets model makers off the hook for capability gaps.