Remove deprecated deepseek-chat and deepseek-reasoner TOML files

[sumansahoo1]

Problem

deepseek-chat and deepseek-reasoner are deprecated aliases for deepseek-v4-flash, fully removed by DeepSeek on 2026-07-24 (22 days from now). But both TOML files still live on dev, so consumers like OpenCode show 4 DeepSeek models instead of the correct 2 (v4-flash + v4-pro).

Only deepseek-v4-flash and deepseek-v4-pro should be listed.

Source

https://api-docs.deepseek.com/quick_start/pricing

deepseek-chat and deepseek-reasoner will be discontinued at 2026/07/24 23:59, please migrate to deepseek-v4-flash as soon as possible.

Affected files (should be removed)

• providers/deepseek/models/deepseek-chat.toml
• providers/deepseek/models/deepseek-reasoner.toml

Related

• Add retirement_date field and backfill it across all first-party labs #1923 — Previously tried adding retirement_date to these files, closed without merging
• fix(models.dev): correct deepseek-chat/reasoner cache_read pricing #2037 — Fixed pricing on both, PR body acknowledged they're deprecated aliases

[jkemp814]

Bumping this: the DeepSeek discontinuation date (2026-07-24) has now passed, and DeepSeek's official Models & Pricing page (https://api-docs.deepseek.com/quick_start/pricing) still lists only deepseek-v4-flash and deepseek-v4-pro — deepseek-chat / deepseek-reasoner are fully gone from their API.

The stale TOML files are still present on dev (confirmed as of today), so OpenCode's /models list still shows 4 DeepSeek models instead of 2:

• providers/deepseek/models/deepseek-chat.toml
• providers/deepseek/models/deepseek-reasoner.toml

Note: the base model metadata also still exists and should be removed alongside the provider entries:

• models/deepseek/deepseek-chat.toml
• models/deepseek/deepseek-reasoner.toml

Alternative to outright deletion: set status = "deprecated" on the provider entries so they stop surfacing in /models while keeping history. Either way, the goal is the same as the original report — OpenCode should list only deepseek-v4-flash and deepseek-v4-pro.

[jkemp814]

Cross-referencing for visibility: raised at anomalyco/opencode#40577 since opencode's /models list surfaces these stale entries for every DeepSeek user.

[jkemp814]

Additional official evidence from DeepSeek's changelog (https://api-docs.deepseek.com/updates):

The two legacy API model names, deepseek-chat and deepseek-reasoner, will be discontinued in three months (2026-07-24). During the current period, these two model names point to the non-thinking mode and thinking mode of deepseek-v4-flash, respectively.

Two takeaways: these names were never separate models — just aliases for deepseek-v4-flash modes — and the 2026-07-24 discontinuation date has now passed, so they are fully retired. This supports removing the stale catalog entries (or marking them deprecated) so /models stops listing dead aliases alongside the real models.

[jkemp814]

Heads-up: a fix for this already exists as an open, mergeable PR — #3371 (fix(deepseek): deprecate legacy model aliases).

It marks deepseek-chat and deepseek-reasoner as status = "deprecated" (the alternative I flagged above as acceptable) and cites the same official DeepSeek changelog as evidence. It's been sitting unmerged since 2026-07-22.

Could a maintainer review/merge it? One gap to consider: it only touches the provider entries — the base metadata under models/deepseek/*.toml would still need the same treatment to fully stop the aliases from surfacing.