Bring-your-own-key (BYOK) pricing is having a moment. Mesrai offers it, Kodus offers it, several others have added it in the last year. The pitch is simple: you already pay your LLM provider, why pay your code review vendor for the same tokens?
The pitch is correct — until it isn't. This post is the math, with real numbers from a 20-engineer team over six months. Not marketing. Receipts.
The setup
Team: 20 engineers, mixed seniority. Stack: TypeScript backend, React frontend, some Python data tooling. Median PR volume: 65 PRs per engineer per month, so about 1,300 PRs per month across the team.
Review tools compared: Mesrai on BYOK (Anthropic claude-opus-4-7) vs Mesrai on BYOK (OpenAI gpt-4o-mini) vs CodeRabbit's per-seat included plan as the included-LLM reference.
The numbers
Three real cost lines from the past six months:
BYOK · Anthropic claude-opus-4-7
Tokens per PR (avg): 31,400
PRs per month: 1,300
Anthropic invoice: $312/mo
Mesrai team seat fee: $0 (free for first 25 seats on BYOK plan)
Total: $312/mo
BYOK · OpenAI gpt-4o-mini
Tokens per PR (avg): 34,200
PRs per month: 1,300
OpenAI invoice: $189/mo
Mesrai team seat fee: $0
Total: $189/mo
Included · CodeRabbit per-seat
20 seats × $57/seat/mo: $1,140/mo
LLM cost: included
Total: $1,140/mo
On pure dollars: BYOK with Anthropic is 73% cheaper than per-seat included. BYOK with gpt-4o-mini is 83% cheaper. The math is real.
The hidden costs
BYOK is cheaper on the invoice. It is not free.
Admin time. Someone needs to: set the LLM key, set spending caps with the provider, monitor for cost spikes, rotate the key if it leaks, deal with rate-limit issues. We measured this at about 2 hours per month, every month. At a $150/hr loaded cost, that is $300/mo of engineering time per team running BYOK.
Add that: BYOK-Anthropic is $612/mo all-in, BYOK-OpenAI is $489/mo all-in. Still much cheaper than $1,140/mo per-seat — but the gap closes.
The quality difference
We ran the same 100 PRs through both BYOK setups and the included plan. Findings overlap:
Mesrai + claude-opus-4-7 caught 94 findings.
Mesrai + gpt-4o-mini caught 81 findings.
CodeRabbit included caught 88 findings.
The Anthropic model caught the most. The cheap OpenAI model caught the fewest. The included plan landed in the middle. None caught everything — there were 12 findings unique to Mesrai+Anthropic and 6 unique to CodeRabbit.
The break-even
BYOK economics depend on token volume. At low PR volume, the included plan can be cheaper because the per-seat cost is fixed but BYOK token usage scales with PR count.
Our calculation for a Mesrai-style setup:
Below ~80 PRs per engineer per month, included-plan pricing wins because token usage is so low BYOK admin overhead dominates.
Between 80–150 PRs per engineer per month, BYOK with a mid-tier model is the sweet spot.
Above 150 PRs per engineer per month, BYOK is a clear win and worth paying for the senior model.
The team in this case study sat at 65 PRs/engineer — just barely on the BYOK side once admin overhead is included. Their conclusion: they kept BYOK because the data residency story matters to them more than the dollars.
The data argument
There is one factor that has nothing to do with dollars and that the BYOK pitch buries: BYOK means your code only ever touches LLM endpoints under contracts you signed. With an included plan, your code touches the review vendor's LLM contracts.
This is fine — most vendors have strong contracts. It is also a real argument that BYOK gives you a cleaner data path. For regulated industries this is often non-negotiable.
The decision tree
Honest, non-marketing version:
Under 80 PRs per engineer per month and no data residency requirements → pay for included.
Over 80 PRs and someone can own the LLM provider relationship → BYOK saves 60–80%.
Any data residency requirement → BYOK regardless of volume.
Small team (under 5 engineers) → individual plans are usually free or near-free; the question is moot.
What Mesrai recommends
Because Mesrai supports both, we are reasonably honest about which to pick:
Start on the included trial. See if the review quality matches your team's standards. Two weeks is enough.
If you decide to keep using it, do the math above. If BYOK wins, switch. If not, the included plan is straightforward and frees you from key management.
Either way, you should not be guessing at the cost six months in. The numbers should be in front of you, in the format above, on a quarterly basis.
The takeaway
BYOK saves money at scale, costs admin time always, and is required by some compliance regimes. For a team under 80 PRs/engineer/month with no compliance constraint, the per-seat included plan is the boring right answer. For everyone else, BYOK is worth the math.