Audit Together AI models and reasoning controls

[rekram1-node]

Summary

• audit every Together AI serverless chat model against official provider catalog, exact model pages, changelog, reasoning docs, and deprecation schedule
• add 9 missing active models, yielding coverage of all 22 currently listed serverless chat models
• retain 7 historical files as deprecated and preserve capabilities independently of lifecycle
• encode reasoning controls only where Together documents exact provider behavior

Evidence

• Current catalog: https://docs.together.ai/docs/serverless/models
• Serverless reasoning table and controls: https://docs.together.ai/docs/inference/chat/reasoning
• Dated pricing/releases: https://docs.together.ai/docs/changelog
• Deprecations: https://docs.together.ai/docs/deprecations
• Google Gemma 4 exact page: https://www.together.ai/models/gemma-4-31b
• Pearl Gemma 4 exact page: https://www.together.ai/models/gemma-4-31b-it-pearl
• MiniMax M2.5 exact page: https://www.together.ai/models/minimax-m2-5
• Qwen3 235B Instruct exact page: https://www.together.ai/models/qwen3-235b-a22b-instruct-2507-fp8
• DeepSeek V4 Pro controls: https://docs.together.ai/docs/deepseek-v4-quickstart

Reasoning semantics

• MiniMax M2.7 and deprecated DeepSeek R1: fixed reasoning (reasoning_options = []) based on positive fixed-mode evidence.
• Documented hybrid models use toggle; GPT-OSS uses exact low, medium, high; DeepSeek V4 Pro uses toggle plus exact high, max.
• Nemotron retains its verified hybrid toggle. Its medium/high depth switch is intentionally not encoded as generic effort because Together exposes chat_template_kwargs={"medium_effort": true}, not reasoning_effort, and the generic schema cannot represent that transport-specific boolean accurately.
• Google and Pearl Gemma 4 exact endpoint pages explicitly describe configurable thinking, so both retain reasoning = true. Their omission from Together’s serverless reasoning table conflicts with those exact pages, and no current first-party wire parameter or accepted values were found; reasoning_options is intentionally omitted.
• Deprecated MiniMax M2.5 retains reasoning = true because its exact endpoint page classifies it as reasoning. No positive Together evidence establishes fixed/toggle/effort/budget semantics, so options remain unresolved rather than fabricated.
• Qwen3 235B A22B Instruct 2507 remains reasoning = false: Together recommends it for reasoning workloads and markets reasoning aptitude, but does not document reasoning-token output or a fixed/toggle/effort/budget mode.
• All reasoning = false models omit reasoning_options.

Metadata corrections

• DeepSeek V4 Pro: $1.74 input / $3.48 output / $0.20 cached input, effective June 9.
• Google Gemma 4: $0.39 input / $0.97 output, effective May 21; structured outputs enabled. The current catalog table still shows the pre-May-21 price, while the dated changelog and exact page show the effective values used here.

Matrix

• 29 files: 22 active, 7 deprecated.
• Active: 14 reasoning, 8 non-reasoning. Controls: 1 fixed, 8 toggle, 2 effort, 1 toggle+effort, 2 unresolved.
• Deprecated: 4 reasoning, 3 non-reasoning. Controls: 1 fixed, 2 toggle, 1 unresolved.
• Explicit unresolved allowlist: google/gemma-4-31B-it, pearl-ai/gemma-4-31b-it, MiniMaxAI/MiniMax-M2.5.

Verification

• bun validate
• bun test packages/core/test/sync-runner.test.ts (9 passed)
• full 29-file matrix check with explicit unresolved allowlist
• generated Together output assertions for corrected reasoning, status, structured output, and prices
• git diff --check
• no Together API credential was available, so no live inference tests were run

Unresolved gaps

• Gemma 4: exact pages establish configurable thinking, but current reasoning docs omit the endpoints and no exact Together wire control is documented.
• MiniMax M2.5: historical reasoning capability is established, but exact control semantics are not.
• Nemotron medium/high depth is documented but is not representable by current generic reasoning option types without mislabeling it as reasoning_effort.