LLM Vendor Lock-In Cost: Claude to Open-Source, Priced

Vendor lock-in is the most over-discussed and under-quantified risk in AI infrastructure. Founders worry about it endlessly and almost never compute the actual switching cost. This article runs the math on a realistic Claude-to-open-source migration, breaks down where the cost lives, and names the threshold below which switching is dead money.

1. The $4k/mo Claude scenario, priced literally

The inputs to the engine: current monthly Claude spend $4,000, prompt complexity 7 (out of 10, meaning prompts have multi-step instructions and few-shot examples but no exotic chain-of-thought wrappers), eval suite size 120 test cases, retraining engineering estimate 60 hours, expected downtime during migration 1.5 days, engineering hourly cost $150 (loaded rate for a senior solo founder), new vendor discount 60% (a hosted Llama 3.1 70B on Together AI vs Claude Sonnet 3.5).

The engine returns:

# llm-vendor-lock-in-cost (computed live from /engines/llm-vendor-lock-in-cost.js)
Engine input
  current_monthly_spend_usd= 4000
  prompt_complexity     = 7
  eval_suite_size       = 120
  retraining_engineering_hours= 60
  downtime_days         = 1.5
  hourly_engineering_cost= 150
  new_vendor_discount_percent= 60

Engine output
  promptRewriteHours    = 28
  evalRewriteHours      = 60
  totalEngineeringHours = 148
  engineeringDollarCost = 22200
  downtimeOpportunityCost= 200
  totalSwitchingCost    = 22400
  monthsOfSpendEquivalent= 5.6
  monthlySavingsAtDiscount= 2400
  paybackMonths         = 9.33

The plan pays back in 9.3 months and saves $2,400/month thereafter, or $28,800/year. Over a 24-month horizon, total savings net of switching cost are about $35,200. Not life-changing for a solo founder, but real money. The relevant question is whether the assumptions hold.

2. Prompt rewrite and eval rewrite are the bulk of the cost

Of the 148 engineering hours, 88 are prompt and eval rewrite work. The reason: open-weight models do not respond to the same prompt structure as Claude. Few-shot examples that work on Sonnet 3.5 may need restructuring on Llama 3.1 to hit similar quality. Eval suites built to grade Claude outputs need recalibration to grade the new model fairly (the rubric is the same, but the failure modes are different).

This is the cost that founders consistently under-estimate. The default mental model is "swap the API endpoint, done." The reality at complexity 7 prompts is that maybe 60% of prompts work unchanged, 30% need minor reformatting, and 10% need substantial restructuring or example replacement. Each non-trivial prompt is 1 to 3 hours of work to rewrite and validate.

The token cost optimization playbook covers the related prompt-engineering work that happens during migration. Founders running this migration usually find that the new prompts are also better for the original vendor — the migration forces a rewrite that should have happened anyway.

3. Downtime opportunity cost is smaller than it feels

The calculator returns $200 of downtime cost on 1.5 days. This looks suspiciously low until you realize what downtime actually means in this context: most LLM migrations run shadow-mode for weeks before the cutover, so real production downtime during the switch is hours, not days. The $200 is the cost of a brief flag flip and rollback safety.

The implication: founders who refuse to migrate because they fear downtime are mispricing the risk by orders of magnitude relative to the engineering cost they are also implicitly avoiding. Downtime fear is rarely a defensible reason not to switch when the engineering cost is the actual barrier. The build vs buy 2026 article covers the broader pattern.

That said, the calculator under-prices catastrophic-quality-regression risk. If the new model gets a critical feature 10% wrong instead of 1% wrong, the cost is customer churn, not downtime. Run a 4-week shadow comparison on production traffic before the cutover. The engineering cost of the shadow comparison is roughly 10 to 20 additional hours, well within the budget.

4. What discount does open-source actually deliver?

The 60% discount input is realistic for the Llama 3.1 70B vs Claude Sonnet 3.5 comparison as of May 2026. Together AI lists Llama 3.1 70B at $0.88 per million combined input+output tokens^[3]. Claude Sonnet 3.5 list pricing is $3 input + $15 output per million^[1], blended at typical 1:1 ratio to $9 per million. That is a 90% discount on per-token rates, but real workloads use more output than input, dropping the realized discount toward 60% to 75% for most products.

Llama 3.1 405B (the larger open-weights model) on Together AI is roughly $5 per million tokens, a 45% discount on Claude Sonnet 3.5. Quality benchmarks are mixed — Llama 405B beats Sonnet 3.5 on some MMLU subsets, loses on most reasoning-heavy tasks. The discount is real but smaller than the headline number suggests.

The discount is also vendor-volatile. Anthropic, OpenAI, and Google all cut frontier-model pricing 30% to 60% in late 2025 and early 2026. The discount window between frontier and open-weights closes on every frontier price cut. Rerunning this calculation every six months catches the case where the migration no longer pays back.

5. The 9-month payback threshold

The 9.3-month payback in the worked scenario is at the edge of what most founders would call a "good" infrastructure investment. The right framing: if monthly spend doubles to $8,000, the payback halves to 4.7 months. If monthly spend drops to $2,000, the payback extends to 18.7 months — and at that point, the question is whether the prompts will even still be relevant in 18 months.

The 9-month rule of thumb: under 9 months payback, migrate. 9 to 18 months, depends on prompt stability and revenue growth trajectory. Over 18 months, do not migrate unless there is a non-cost reason (vendor risk, regulatory requirement, customer demand for open-weights).

The vendor lock-in math article covers the related framework for evaluating which migrations are worth running and which are aspirational. The short version: spend matters, prompt stability matters more, and the migration plan should include a kill switch if quality regresses on production traffic.

6. The hidden ongoing costs of self-hosting

The calculator does not include the ongoing operational cost of running your own inference infrastructure, because the assumption is a managed open-weights provider (Together, Fireworks, Anyscale, Replicate). If a founder reads this article and decides to self-host Llama 3.1 70B on rented GPUs, the cost picture changes.

Self-hosted Llama 3.1 70B requires roughly 2x A100-80GB GPUs ($2.50 to $4.00 per GPU-hour spot, or $1,800 to $2,880 per month per GPU at 100% utilization). At 30% utilization (realistic for solo workloads), the GPU bill is $1,000 to $1,800 per month, comparable to the API spend. Plus DevOps work (scaling, monitoring, model updates) at 8 to 20 hours per month, or $1,200 to $3,000/month at loaded rates.

Self-hosting is rarely cheaper than a managed open-weights provider for solo workloads. The U.S. Bureau of Labor Statistics^[5] reports a May 2024 mean wage of $140,910 for software developers (occupation 15-1252), or about $88/hour unloaded. At $150/hour loaded, the DevOps cost dominates any infrastructure savings under 24/7 high-utilization workloads.

7. When to switch and when to stay

Three-row decision matrix:

1. Under $1,500/mo current spend: stay. The migration cost is more than 12 months of total spend; the math never works.
2. $1,500 to $4,000/mo current spend with stable prompts and 12+ month product runway: migrate to a managed open-weights provider (Together, Fireworks). Payback at 6 to 12 months, ongoing savings real.
3. Over $4,000/mo current spend or unstable prompts: migrate aggressively, but include a 4-week shadow-mode period and a defined kill-switch threshold. The savings are large enough to fund a proper migration.

The other consideration: open-weights migration is a hedge against vendor risk (pricing changes, API deprecation, terms-of-service changes that block your use case). Even at break-even payback, the hedge has real value for solo founders dependent on a single vendor for a critical product capability. The methodology behind the engine's switching-cost model is documented at the LLM Vendor Lock-In Cost methodology page^[4].

8. FAQ

Is migrating from Claude to open-source actually cheaper? At $4,000/mo spend with a 60% discount on hosted Llama, payback is 9.3 months. Below $1,500/mo spend the math does not work.

What is the biggest cost in an LLM migration? Engineering time, by 100x over downtime. Prompt and eval rewrites are 60% to 80% of that, integration and infra work the rest.

Should a solo founder ever self-host an LLM? Almost never. Use a managed open-weights provider; DevOps cost on self-hosting dominates the API-spend savings under 24/7 high-utilization workloads.

What discount should I assume on the new vendor? 50% to 70% on hosted open-weights vs Claude Sonnet 3.5. Llama 3.1 70B is roughly 40% cheaper realized on typical input/output ratios; 405B is closer to par.