feat(302ai): add deepseek-flash (DeepSeek-V4.1-Flash)

[JI4JUN]

Summary

Add DeepSeek-V4.1-Flash to the 302.AI catalog. 302.AI serves this model under the routing ID deepseek-flash (catalog SKU created 2026-09-08), so the provider file is providers/302ai/models/deepseek-flash.toml, inheriting lab metadata via base_model = "deepseek/deepseek-v4.1-flash".

Sources & Documentation

• Catalog & Pricing API: POST https://api.302.ai/v1/models?include_custom_models=1&llm=1 (retrieved 2026-09-14)
  • SKU deepseek-flash: context 1,000,000; USD per MTok input = $0.150, output = $0.600, cache_read = $0.003 — matches DeepSeek first-party pricing for V4.1-Flash (providers/deepseek/models/deepseek-flash.toml, off-peak rate per https://api-docs.deepseek.com/quick_start/pricing).
• Live behavior probes (POST https://api.302.ai/v1/chat/completions, 2026-09-14, SKU deepseek-flash):
  • Plain completion returns reasoning_content in the message plus completion_tokens_details.reasoning_tokens — side channel confirmed.
  • thinking.type = "disabled" suppresses reasoning entirely (no reasoning_content, reasoning_tokens = 0) — toggle verified.
  • Invalid reasoning_effort values are rejected (-10003) — the effort field is parsed upstream, not silently ignored.

Reasoning Controls (Relay Classification)

Per AGENTS.md / audit-reasoning-options: 302.AI is a multi-model relay; controls mirror the underlying lab surface (DeepSeek first-party maps thinking.type toggle + reasoning_effort = low|high|max) and are additionally live-verified on this host (see probes above):

# Toggle: thinking.type = enabled|disabled
# Effort: reasoning_effort = low|high|max
reasoning_options = [
  { type = "toggle" },
  { type = "effort", values = ["low", "high", "max"] },
]

[interleaved]
field = "reasoning_content"

Notes

• The 302.AI catalog marks this SKU is_featured = false; it is added here explicitly per host request, and its reasoning metadata flag is false while the model provably streams reasoning (probes above), consistent with the catalog's unreliable reasoning metadata on other SKUs (e.g. gpt-5.6-sol).
• Legacy host SKU deepseek-v4.1-flash-expires-on-0910 also serves this model and shows identical behavior; not added since it is a dated/expiring alias.

Verification

• bun validate: Passed with 0 errors.

[github-actions]

No actionable findings.