feat(prices): price DeepSeek, at its peak rate

[yuanhao]

What

The bundled price snapshot gains DeepSeek: deepseek-flash, deepseek-v4-pro, and the legacy deepseek-v4-flash (served by V4.1 Flash, billed as Flash). Pricing stays opt-in (prices::enable_bundled()).

Rates per million tokens, from DeepSeek's pricing page, read from its raw markup, not a summary:

Model	Input (cache miss)	Cache hit	Output
deepseek-flash, deepseek-v4-flash	$0.30	$0.006	$1.20
deepseek-v4-pro	$1.32	$0.044	$3.96

DeepSeek has no cache-write charge, so cache_write = input, under the existing cache_write_at_input allowance.

Why the peak rate

DeepSeek bills half of these off-peak, which is every hour except 01:00–04:00 and 06:00–10:00 UTC on weekdays (about 79% of the week). A CostConfig has no time of day, so the snapshot carries the peak rate as an upper bound:

• costs are over-stated off-peak and never under-stated;
• a Budget stops early rather than late.

docs/concepts/pricing.md shows how to install the off-peak rates as an override for anyone who'd rather count those.

Keeping it honest

• tests/price_audit.rs now records DeepSeek's differences from models.dev instead of skipping them, and asserts them both ways:
  • DeepSeek's input, output and cache_read must be exactly twice models.dev's, which lists the off-peak rates. If models.dev starts listing the peak rate, or the ratio changes, that is drift.
  • models.dev's separate reasoning rate must equal its output rate, recorded as FLAT_RATE_GAPS with the figures pinned.
  • Positive controls: an entry at the off-peak rate is drift, and a missing reasoning gap is drift.
• The checked-in models.dev slice is refreshed: V4 Pro's old figures there were wrong, and deepseek-v4-flash is added.
• A live models.dev fetch (enable_live) skips DeepSeek's models as priced by time of day, so its off-peak rates never replace the bundled peak ones. The skip is reported in skipped, without the warning meant for unmappable data, and the live audit asserts it stays skipped.

Tests

• Tests that used DeepSeek as "an unpriced model" (price_override_test, two in model.rs, the CostConfig doc example) now use Mistral.
• price_fetch_test counts are updated for the refreshed slice.
• Locally:
  • cargo test --all-features passes;
  • the live price audit passes, with DeepSeek's 12 fields checked as peak = 2 × off-peak and 0 drift;
  • clippy -D warnings, rustdoc and fmt are clean;
  • the live DeepSeek smoke now runs its cost check, 6/6 (it was "n/a" before).

🤖 Generated with Claude Code

https://claude.ai/code/session_01T7iq5hpndSiHQcnAsywKuG