fix: deepseek-v4-flash reasoning_options omit low, which the API accepts and honors

[github-actions]

Summary

deepseek-v4-flash documents a distinct reasoning_effort=low level (maps to low on Flash), but several catalog entries only listed high/max. That hid the only effort value that differs meaningfully from the default on Flash. Effort values are now ["low", "high", "max"] on the first-party DeepSeek entry and the opencode / opencode-go Flash hosts.

Changes

File	Field	Old	New
providers/deepseek/models/deepseek-v4-flash.toml	reasoning_options effort	["high", "max"]	["low", "high", "max"]
providers/opencode-go/models/deepseek-v4-flash.toml	reasoning_options effort	["high", "max"]	["low", "high", "max"]
providers/opencode/models/deepseek-v4-flash.toml	reasoning_options effort	["high", "max"]	["low", "high", "max"]
providers/opencode/models/deepseek-v4-flash-free.toml	reasoning_options effort	["high", "max"]	["low", "high", "max"]
providers/deepseek/provider.toml	leading comment	said low/medium→high	documents Flash low→low vs Pro low→high

Left unchanged: deepseek-v4-pro remains high/max (docs map Pro low→high). Did not add xhigh (Flash maps it to high) or medium (not a native OpenAI-format effort for this model).

Evidence

• DeepSeek Thinking Mode guide — OpenAI-format effort is low/high/max; default effort is high. Mapping table: Flash low→low, high→high, xhigh→high, max→max; Pro low→high.
• Issue deepseek-v4-flash reasoning_options omit low, which the API accepts and honors #3962 — opencode-go deserializer accepts low; measured reasoning-token means show low ≈ 879 vs default/high ≈ 2540–2580 through the go gateway.

Validation

• Checked official Thinking Mode docs against the mapping table for Flash vs Pro.
• Diff-reviewed the four model TOMLs and provider comment for leading source comments and effort lists only.
• Did not run bun validate or live API calls in this environment.

Review notes

• Other relays (Fireworks, Alibaba, Ollama Cloud, etc.) still list Flash as high/max only; left alone without per-host verification.
• opencode / opencode-go shapes kept as-is (toggle only where already present); only the effort list changed.
• AGENTS.md still summarizes DeepSeek V4 as high/max in places; that text was not edited (catalog-only change).

Closes #3962

Automated by the issue fixer: https://github.com/anomalyco/models.dev/actions/runs/30759490163

[sigco3111]

Hi! Quick question on the trigger for this fix — was the low option explicitly rejected by the DeepSeek v4-flash API response, or was this caught during upstream SDK behavior validation? I'm asking because the cost-optimization guidance differs depending on which case it is (silent accept vs explicit reject behaves very differently for retries).

I'm maintaining a Korean dev community cost-optimization guide for DeepSeek models, so getting the low option behavior right matters a lot for our users.

Thanks for the great work on models.dev! 🙏