M1 · THE ECONOMIC CASE · P03

The Model-Swap Playbook

OpenAI changed the model. Your company didn’t notice.

10 SLIDES · ~10 MIN · PLAYBOOK OVERVIEW
Key Takeaways
  • If you don't own model-swap rights, you don't own your product.
  • ~6 — silent model updates per quarter at major vendors (estimated).
  • 20-40% — of enterprise prompts drift in output after a swap.
  • 14+ — days from swap to discovery in unmonitored systems.
  • Add a Model-Swap Notification Clause to every AI vendor contract — 30-day notice minimum.
MODULE 1 · PLAYBOOK 03 · ECONOMIC FRAME

The Model-Swap Playbook

OpenAI changed the model. Your company didn't notice.

BuildClub Academy
01 · WHY THIS MATTERS

Silent model updates are now your product risk.

~6
silent model updates per quarter at major vendors (estimated).
OpenAI deprecation log
20-40%
of enterprise prompts drift in output after a swap.
Anthropic release notes
14+
days from swap to discovery in unmonitored systems.
operator network interviews
02 · THE THESIS

If you don't own
model-swap rights,
you don't own your product.

The Thesis
03 · VOCABULARY

Six terms every CEO needs.

Model Version
The specific weights running behind the API. Changes when the vendor retrains or fine-tunes.
Endpoint
The API URL your code calls. Same endpoint can serve different model versions over time.
System Prompt
The instructions you wrap around user input. Brittle when the model behind the endpoint changes.
Regression Suite
Frozen test cases run nightly against your prompts to catch output drift before customers do.
Frozen Snapshot
A version-locked endpoint that won't change without explicit action — your insurance for high-stakes flows.
Model-Swap Notification Clause
Contractual obligation for the vendor to notify before model changes (30+ days advance notice).
04 · WHAT BREAKS

What changes silently when the model swaps.

1
Prompt outputs break
Prompts that worked yesterday produce wrong, off-tone, or incomplete output today. Same input, new model.
2
Output schema drifts
JSON field names rename. Structure changes. Downstream code that parses output breaks silently.
3
Cost-per-call shifts
Token efficiency changes. Your $4K monthly bill becomes $11K. No one looks until finance flags it.
05 · THE ANTHROPIC STORY

From press release to P&L.

1
March 2026 — Anthropic outage
The real cost wasn't downtime. It was prompts that worked Tuesday breaking Wednesday. Production customers absorbed the variance the press release didn't mention.
2
April 2026 — GPT-5.4 quiet shift
OpenAI launched GPT-5.4. JSON output behavior changed on three customer-facing endpoints the same week. No deprecation notice, no warning.
3
The pattern
61% of enterprises name OpenAI as primary GenAI provider — that is concentration risk, not strategy. Vendors announce capability gains. The P&L impact lands on the customer.

Vendors announce. Customers absorb. The press release celebrates; the customer rewrites the prompts.

06 · THE PLAYBOOK

Four steps to own your model risk.

1 Demandnotification clause 30 days minimum 2 Nightlyregression suite diff against snapshot 3 Version-lockhigh-stakes flows until proven safe 4 Wire alt-vendorfallback tested, not theoretical
07 · VERSION-LOCK DECISION MATRIX

Where to lock the model. Where to let it move.

High-stakes / High-volume
Lock + monitor. Legal-document generation, financial summaries, regulated workflows. Drift here is balance-sheet risk.
High-stakes / Low-volume
Lock. Board materials, executive contracts, M&A diligence. Volume is low; cost of being wrong is catastrophic.
Low-stakes / High-volume
Monitor. CX support, internal Q&A, content drafts. Run regression suite; accept drift if outputs stay in range.
Low-stakes / Low-volume
Unlocked. Experimentation, prototypes, ad-hoc analysis. The cost of locking exceeds the cost of drifting.
08 · THE CAPABILITY TRAP

Single-vendor dependency is concentration risk, not strategy.

1
61% cite OpenAI as primary
Single-provider dominance creates correlated failure modes across the enterprise.
2
Anthropic March 2026 outage
Made the risk concrete: hours of downtime translated directly to customer-facing incidents.
3
73% have no fallback
PwC 2026: most enterprises have no documented alternative vendor for their primary AI workload.
CLOSING

Five things to do Monday.

Monday Morning
  1. 1
    Add a Model-Swap Notification Clause to every AI vendor contract — 30-day notice minimum.
  2. 2
    Stand up a nightly regression suite against frozen test cases for your top 3 prompts.
  3. 3
    Identify the 5 high-stakes flows that must run on version-locked endpoints.
  4. 4
    Wire one alt vendor as fallback for each critical workload — validated quarterly.
  5. 5
    Name an owner for model-swap response — one person, one runbook, accountable for the prompt-rewrite drill.
P03 · Questions CEOs Ask

Frequently Asked Questions

What is the core idea of The Model-Swap Playbook?
If you don't own model-swap rights, you don't own your product.
What does the data say a CEO should pay attention to?
~6 silent model updates per quarter at major vendors (estimated). 20-40% of enterprise prompts drift in output after a swap.
What should a CEO do Monday morning after reading The Model-Swap Playbook?
Start here: Add a Model-Swap Notification Clause to every AI vendor contract — 30-day notice minimum; Stand up a nightly regression suite against frozen test cases for your top 3 prompts; Identify the 5 high-stakes flows that must run on version-locked endpoints.
What are the steps in The Model-Swap Playbook?
1) Demand notification clause; 2) Nightly regression suite; 3) Version-lock high-stakes flows; 4) Build alt-vendor fallback.
Where do the claims in The Model-Swap Playbook come from?
The playbook cites OpenAI deprecation log; Anthropic release notes; Anthropic status page; Gartner 2026 GenAI Provider Survey.
P03 · Sources

All Sources in This Playbook

00 / 10