fix(subconscious): use Anthropic Messages API and correct model metadata

[github-actions]

The file is now corrected and consistent with the approved GLM-5.2 entry for the same provider. Let me do a final review of the changes against the authoritative sources.

The corrections are complete and verified:

1. interleaved: { field = "reasoning_content" } → true. The Thinking Mode docs show reasoning streams inline in content (between think tags) for the OpenAI-compatible endpoint, with streaming reading delta.content. The reasoning_content field form is for providers that split reasoning into a separate field, which Subconscious does not do.

2. cost: { input = 0, output = 0 } → { input = 0.3, output = 3.0, cache_read = 0.15 }. The pricing page lists TIM-Qwen3.6 27B at cached $0.15 / input $0.30 / output $3.00 per 1M tokens.

3. reasoning_options: added the API-syntax comment to the existing toggle and added a budget_tokens option. The Thinking Mode docs (using subconscious/tim-qwen3.6-27b as the example) document both the OpenAI chat_template_kwargs.enable_thinking toggle and the Anthropic thinking.budget_tokens control.

4. Leading source-comment block added per AGENTS.md convention (Subconscious has no sync module, so comments are preserved).

I did not run bun validate or any shell/API commands.

---

Summary

Issue #3061 ("wrong code") pointed at the pre-existing providers/subconscious/models/subconscious/tim-qwen3.6-27b.toml. The discrepancy was flagged in the review notes of PR #3060 (which added GLM-5.2 to Subconscious): the TIM-Qwen3.6 27B entry had a zero cost that contradicted Subconscious's pricing page, and an interleaved value that contradicted Subconscious's Thinking Mode documentation. This corrects both factual errors and completes the reasoning_options to match the provider's documented API surface, bringing the entry in line with the approved GLM-5.2 entry for the same provider.

Changes

In providers/subconscious/models/subconscious/tim-qwen3.6-27b.toml:

• interleaved: { field = "reasoning_content" } → true
  • Reasoning streams inline in content (between think tags) on the OpenAI-compatible endpoint, not in a separate reasoning_content field.
• [cost]: input = 0, output = 0 → input = 0.3, output = 3.0, cache_read = 0.15 (USD per 1M tokens)
  • Added the missing cache_read (cached-token rate).
• [[reasoning_options]]: added the API-syntax comment to the existing toggle and added a second budget_tokens option
  • toggle → # API: {"chat_template_kwargs": {"enable_thinking": true | false}}
  • new budget_tokens → # API: {"thinking": {"type": "enabled", "budget_tokens": <n>}} (bounds omitted; not documented)
• Added a leading source-comment block citing the pricing and thinking-mode docs (Subconscious has no sync module, so the block is preserved on re-serialization).

No other fields changed. release_date, last_updated, limit, modalities, and capability booleans were left as-is — they were not flagged and could not be independently verified against provider docs.

Evidence

• Subconscious Pricing — lists subconscious/tim-qwen3.6-27b with per-1M-token rates: cached $0.15, input $0.30, output $3.00. Establishes the corrected cost values.
• Subconscious Thinking Mode docs — uses subconscious/tim-qwen3.6-27b as the example model and establishes three facts: (1) the OpenAI-format response carries reasoning inline in content wrapped in think tags, with streaming reading delta.content (justifies interleaved = true); (2) the OpenAI format toggles thinking via chat_template_kwargs.enable_thinking (justifies the toggle option and its API comment); (3) the Anthropic Messages format uses thinking.type = "enabled" with thinking.budget_tokens to cap reasoning (justifies the budget_tokens option). The docs show only an example budget_tokens value of 2000 and no documented bounds, so min/max are omitted.
• PR #3060 review notes (context only) — the GLM-5.2 entry's author explicitly flagged the interleaved and cost discrepancies in this file as out-of-scope for that PR; the reviewer then opened issue wrong code. reference https://github.com/anomalyco/models.dev/blob/dev/providers/subconscious/models/subconscious/tim-qwen3.6-27b.toml #3061 referencing this file. The approved GLM-5.2 entry uses interleaved = true plus the same toggle/budget_tokens reasoning-option pair, confirming the target structure.

Validation

• Verified the corrected cost values against the live pricing page (cached $0.15, input $0.30, output $3.00 per 1M tokens).
• Verified interleaved = true against the Thinking Mode docs: reasoning is inline in content for the OpenAI-compatible endpoint (the reasoning_content separate-field form does not apply to Subconscious).
• Verified both reasoning-option controls against the Thinking Mode docs, which use subconscious/tim-qwen3.6-27b as the example for both the OpenAI toggle and the Anthropic budget_tokens control.
• Confirmed the Cost schema accepts cache_read as an optional field and that budget_tokens permits omitted min/max.
• Did not run bun validate or any shell/API commands.

Review notes

• budget_tokens bounds are intentionally omitted: the docs show only an example value (2000) and publish no min/max, consistent with the audit guidance to omit unverified bounds.
• release_date/last_updated (2026-05-11) and the limit block (context/input 8192, output 5000) were not changed; the reviewed docs do not publish these values, so I could not verify them and the issue did not flag them.
• The added budget_tokens option goes slightly beyond the two discrepancies explicitly named in the PR feat(subconscious): add GLM-5.2 #3060 notes (interleaved and cost), but is required for internal consistency: AGENTS.md treats complete reasoning_options as a blocker for reasoning models, the Thinking Mode docs document the control for this exact model, and it matches the approved GLM-5.2 entry.

Closes #3061

Automated by the issue fixer: https://github.com/anomalyco/models.dev/actions/runs/28795107369

[kraplegit]

Editing the wrong file

[rekram1-node]

Follow-up in d62f00cd3 changes the provider to @ai-sdk/anthropic and removes interleaved. Subconscious currently has one cataloged model, subconscious/tim-qwen3.6-27b, and its OpenAPI explicitly supports that model on POST /v1/messages, including native thinking and tool content blocks. The Chat Completions API returns reasoning inline in content and does not document reasoning_content; inline output alone does not establish interleaved thinking between tool calls. bun validate and git diff --check pass.