Model Price Shock vs Vendor Lock-In: Two Defenses Compared

Switch vendors against a sustained price spike and absorb a temporary one: a $15,000-$30,000 migration pays back in 2-3 months against a sustained 200% spike (which pushes a $28k MRR product's AI cost from $6,200 to $18,600/month), but loses against a brief one. The two halves of the problem rarely get analyzed together: the price-shock stress test is standard, the switching-cost side is the rarer analysis, and the eval-suite cost is the hidden variable that flips the math. This article runs both halves on the same product and pulls the decision rule out.

1. The $28k MRR setup at 51% margin

The scenario: $28,000 MRR AI SaaS at 51% gross margin, $6,200 monthly AI cost (Claude Sonnet across user-facing flows). Gross profit today: $14,280. Non-AI cost: $7,520. The 51% margin is below SaaS norms (70-75%) because AI cost is high relative to revenue at this scale. The stress-test engine confirms the today line and reports three price-drop scenarios:

The engine models drops. For the lock-in analysis, the relevant scenarios are price spikes, which the same arithmetic runs in the inverse direction (not produced by the engine, which only models drops): a 50% spike makes AI cost $9,300; a 100% spike $12,400; a 200% spike $18,600. The 200% case is the conservative planning shock, not the expected one.

2. The price-shock side: 200% spike scenario

Pricing the 200% spike:

Today:
  Monthly AI cost           $6,200
  Total monthly cost        $13,720
  Gross profit              $14,280
  Gross margin              51%

After 200% AI spike:
  New monthly AI cost       $18,600
  Total monthly cost        $26,120
  Gross profit              $1,880
  Gross margin              ~7%
  Cumulative annual hit     $148,800

After full pass-through to customers (22% price cut to defend share):
  Revenue                   $21,840
  AI cost                   $18,600
  Gross profit              -$4,280   (negative)

Stanford HAI's AI Index documents that token prices have fallen 50-60% across major models in 2023-2024^[4], but the absolute spend trend has been upward as products grow MAU. A 200% sustained spike is hypothetical but not impossible — vendor capacity crunches, regulatory pressure, or compute-cost shifts could produce it. Planning against the spike is the right defensive posture.

3. The lock-in side: switching cost components

Switching vendors costs three things:

• Eval-suite rebuild ($8,000-$15,000 of engineering time). The existing model's behavior on the product's specific prompts has been tuned over months. A new model behaves differently. The eval suite (a test set of 100-500 representative inputs with expected outputs) needs to be re-run, recalibrated, and revalidated. Building from scratch takes 2-4 weeks of focused engineering.
• Prompt re-engineering across all routes ($4,000-$10,000). Prompts that work on Claude don't always work on GPT-5.x or Gemini. Each user-facing route needs its prompt rewritten and tested. Time: 1-3 weeks depending on route count.
• Migration debt and quality risk (hard to price, 60-90 days exposure). While switching, the product runs two pipelines or operates on a less-tested vendor. Customer-facing quality is uncertain. Bug reports rise temporarily. Some customers churn during the window — the implied cost is the churn the product would not have suffered if it stayed on the original vendor.

Run a representative configuration through the lock-in engine — current $6,200/month spend, moderate prompt complexity, a 120-input eval suite, 80 hours of retraining engineering at $150/hour, 2 days of downtime, and a 30% new-vendor discount:

The engine returns a $25,013.33 one-time switching cost (164 engineering hours at $150 plus downtime opportunity cost), against $1,860/month of savings at the 30% discount — a payback of roughly 13.5 months if the spike persists. That sits inside the realistic $15,000-$30,000 band, plus the 60-90 day quality risk that the engine does not price. Together AI's pricing data shows alternative inference hosting starts around 20-50% cheaper than first-party Anthropic or OpenAI^[3].

4. Comparing absorb vs switch

The break-even math:

If shock lasts 1 month:
  Absorb cost                 $12,400  (extra AI spend for 1 month)
  Switch cost                 $25,013  (engine one-time total)
  Winner                      Absorb

If shock lasts 3 months:
  Absorb cost                 $37,200
  Switch cost                 $25,013 + maybe 2-month migration cost
  Winner                      Switch

If shock lasts 6 months:
  Absorb cost                 $74,400
  Switch cost                 $25,013 + 2-month migration
  Winner                      Switch (decisively)

If shock is permanent:
  Absorb cost                 $148,800/year
  Switch cost                 $25,013 one-time
  Winner                      Switch (overwhelmingly)

The break-even is roughly 2 months ($25,013 switch cost / $12,400 monthly absorb cost). For shorter shocks, absorb. For longer shocks, switch. The challenge is that solo founders rarely know how long a shock will last when it begins — vendors are unpredictable about both pricing and stability.

5. The eval suite: the hidden variable

The single largest variable in switching cost is the maturity of the eval suite. Products with a mature, automated eval suite (100+ test cases, automatic regression detection, CSAT correlation) can swap vendors in 1-2 weeks. Products without one are starting from scratch — 4-8 weeks of building and validating tests before the actual migration begins.

Building an eval suite is one of those high-impact solo-founder investments that pays off in the worst-case scenario rather than the average case. In normal operation, the eval suite catches model regressions (Anthropic ships a quiet update, GPT-4o behavior shifts on edge cases). In a price-shock scenario, it cuts switching cost by 40-70%. Total investment: 3-6 weeks of focused engineering. Total returns over 3 years: typically 4-8 weeks of saved migration time plus 30-60% lower regression-induced churn.

The eval suite is also the single most useful artifact for negotiating with vendors. A founder who shows a vendor "our eval suite shows your model degraded 8% on our use case this month" gets a different response than a founder who reports "things feel different." Vendors honor measurable regressions when they're documented; they wave off subjective complaints.

6. The hybrid defence: partial multi-vendor

The most economically efficient defense is neither pure absorb nor pure switch — it's structural multi-vendor capability:

• Build the AI layer with a routing abstraction. One config flip swaps Anthropic/OpenAI/Together/Bedrock. Architectural cost: 1-2 weeks. Insurance value: substantial.
• Run 80% of production on the primary vendor, 20% on the secondary. Real production traffic on both means the secondary is always warm — no surprises during a switch.
• Re-evaluate the mix quarterly. Vendor pricing diverges constantly. The optimal 80/20 split today is rarely the same as in 90 days.

The hybrid approach turns a 60-90 day migration into a 60-90 minute config change. Total cost: 1-2 weeks of upfront architecture work plus 15-20% higher operational complexity from running two providers. Compared to the $30,000 switching cost in a forced migration, the hybrid is dramatically cheaper insurance.

Stanford HAI's pricing-trend data shows that the gap between cheapest and most expensive frontier model has been 3-5x at any given point^[4]. A hybrid that captures the cheapest 20-40% of traffic on the cheapest provider saves money continuously, not just during a shock.

7. The 90-day defence playbook

Six moves to build resilience against price shocks:

• Week 1-4: Build the eval suite. 100+ representative test cases. Automated runner. Regression alerts. The most important defensive investment.
• Week 4-6: Build the routing abstraction. One config flag swaps providers. Test it on 5-10% of traffic to confirm correctness.
• Week 6-8: Establish the secondary provider in production. Run 10-20% of real traffic through the alternative. Track CSAT and quality metrics.
• Week 8-10: Negotiate annual contract with primary. Lock in current rates against future shocks. Volume commits typically save 10-20% on rates.
• Week 10-12: Build cost dashboard. Daily AI spend by route and provider. Alert on +30% week-over-week. Catch shocks within 24 hours, not on the monthly invoice.
• Ongoing: Re-run both engines quarterly. Track the cost ratio between providers and the cost-to-MRR ratio. When the ratio shifts materially, act.

The combined cost of this playbook (3 months of part-time engineering, roughly $15,000-$25,000 of opportunity cost) is less than a single switching event. Founders who treat the playbook as overhead pay 3-5x more during the inevitable next shock. The runway-with-AI-cost-shock calculator handles the downstream cash impact; the vendor lock-in cost engine sizes the switching cost specifically. Anthropic's^[1] and OpenAI's^[2] published prices have been stable through 2025-2026, which is the calm period to build the defence — not after the spike arrives.

A useful sanity-check pattern: chart the AI cost as a percent of revenue monthly. At under 15%, shocks are absorbable. At 15-25%, shocks require defensive action but don't threaten the business. Above 25%, the product has a structural problem (priced too low or token-intensive product surface) that no defense can fully solve. The $28k MRR scenario sits near 22% (6,200 / 28,000), so it's in the defensive-action band — the playbook above is exactly the right response, neither overreaction nor underreaction.

One additional consideration solo founders skip: the contract-side defense. Most vendors will negotiate fixed-rate annual commits at 10-25% off published pricing in exchange for a minimum spend commitment. Anthropic's enterprise contracts, OpenAI's volume tiers, and Bedrock's reserved capacity all offer this. For a $6,200/month product, committing to $50,000/year in spend ($4,167/month average) at a 15% discount produces $7,500-$9,000 of annual savings during normal operation and protects against price changes during the contract term. The downside is reduced flexibility — but for a product with predictable usage growth, the trade is usually worth it.

One last pattern. The instinct to "wait until I have to" on multi-vendor is wrong because the worst moment to migrate is during a price shock — quality risk is highest, engineering attention is competing with customer issues, and the urgency degrades the migration's quality. Build the multi-vendor capability in the quiet period; flip the config when needed. See the methodology for the full derivation^[5].

Frequently asked questions

What does a 200% price spike do at $28k MRR?

AI cost rises from $6,200 to $18,600/month. Total cost climbs from $13,720 to $26,120, leaving $1,880 of gross profit instead of $14,280. Gross margin drops from 51% to roughly 7%. Without intervention, the product is one bad month from net-negative.

What does switching vendors cost?

Three parts: eval-suite rebuild ($8,000-$15,000 of engineering time), prompt re-engineering across all routes ($4,000-$10,000), and migration debt (customer-facing quality during the switch, hard to price but real). Realistic total: $15,000-$30,000 of one-time cost plus a 60-90 day quality risk period.

When does switching beat absorbing?

When the price shock is sustained for 6+ months. A 200% spike that costs $12,400/month extra burn pays back even a $30,000 migration in 2.4 months. The switch beats absorb decisively at long shocks; absorb beats switch only at short, recoverable spikes.