Problem
The #510 vendored models.dev refresh deleted retired first-party model entries while reseller providers kept their copies. flatten (crates/relayburn-sdk/src/analyze/pricing.rs:183-222) only protects a model when the primary vendor still publishes it (PRIMARY_PRICING_PROVIDERS, pricing.rs:127-131), so last-wins silently hands such models to a reseller tariff instead of flagging them unpriced.
36 models switched from first-party to reseller pricing on main. Worst cases:
claude-sonnet-4-20250514: anthropic 3/15 (cache_read 0.30, cache_write 3.75) → jiekou 2.7/13.5 with NO cache fields. cache_read.unwrap_or(0.0) (pricing.rs:205) bills cache reads at $0.00 and cache_write.unwrap_or(input) (pricing.rs:209) drops cache creation 3.75→2.7 — cache_read is usually the dominant bucket for Claude Code sessions, so cost is drastically under-reported.claude-opus-4-20250514: 15/75 → 13.5/67.5, cache pricing gone.gemini-2.5-flash-preview-05-20: output 0.6 → 3.15 (5.25× over-report).gemini-2.5-pro-preview-05-06: input 1.25 → 2.5 (2× over).gpt-5-codex: 1.25/10 → 1.125/9 — this is why BURN_GOLDEN=1 golden cases summary/summary-json fail on main tip (snapshot $0.0065 vs binary $0.0059).
Attribution: all three inputs to the failing golden cases are identical at main tip 5b531d0 — the defect is on main, not in any open branch.
Fix
When a primary vendor previously published a model and the refreshed snapshot no longer carries it, treat the model as unpriced (route through the unpriced-usage path #510 added) rather than falling through to a reseller. Do NOT regenerate the golden snapshots — that would bake reseller prices into the expectations.
Related
#509 (override loading), #503/#510 (precedence groundwork), #522 (scheduled refresh workflow — its test gate will trip on this class of change until fixed).
Found during retroactive adversarial review of PR #511's merge delta.
🤖 Generated with Claude Code
Problem
The #510 vendored models.dev refresh deleted retired first-party model entries while reseller providers kept their copies.
flatten(crates/relayburn-sdk/src/analyze/pricing.rs:183-222) only protects a model when the primary vendor still publishes it (PRIMARY_PRICING_PROVIDERS, pricing.rs:127-131), so last-wins silently hands such models to a reseller tariff instead of flagging them unpriced.36 models switched from first-party to reseller pricing on main. Worst cases:
claude-sonnet-4-20250514: anthropic 3/15 (cache_read 0.30, cache_write 3.75) → jiekou 2.7/13.5 with NO cache fields.cache_read.unwrap_or(0.0)(pricing.rs:205) bills cache reads at $0.00 andcache_write.unwrap_or(input)(pricing.rs:209) drops cache creation 3.75→2.7 — cache_read is usually the dominant bucket for Claude Code sessions, so cost is drastically under-reported.claude-opus-4-20250514: 15/75 → 13.5/67.5, cache pricing gone.gemini-2.5-flash-preview-05-20: output 0.6 → 3.15 (5.25× over-report).gemini-2.5-pro-preview-05-06: input 1.25 → 2.5 (2× over).gpt-5-codex: 1.25/10 → 1.125/9 — this is whyBURN_GOLDEN=1golden casessummary/summary-jsonfail on main tip (snapshot $0.0065 vs binary $0.0059).Attribution: all three inputs to the failing golden cases are identical at main tip 5b531d0 — the defect is on main, not in any open branch.
Fix
When a primary vendor previously published a model and the refreshed snapshot no longer carries it, treat the model as unpriced (route through the
unpriced-usagepath #510 added) rather than falling through to a reseller. Do NOT regenerate the golden snapshots — that would bake reseller prices into the expectations.Related
#509 (override loading), #503/#510 (precedence groundwork), #522 (scheduled refresh workflow — its test gate will trip on this class of change until fixed).
Found during retroactive adversarial review of PR #511's merge delta.
🤖 Generated with Claude Code