feat(models): sync together_ai model registry

[mateo-berri]

Automated daily sync of the together_ai entries in model_prices_and_context_window.json against GET https://api.together.ai/v1/models?serverless and https://docs.together.ai/docs/deprecations.md by scripts/sync_together_ai_models.py.

Added (1)

• together_ai/zai-org/GLM-5.3

Updated (1)

• together_ai/Qwen/Qwen3.8-2.4T-A95B: input_cost_per_token: 2e-06 -> 2.5e-06; output_cost_per_token: 6e-06 -> 6.25e-06; cache_read_input_token_cost: 2.5e-07 -> 5e-07

Marked deprecated (0)

• none

Returned to the catalog (0)

• none

Warnings needing a human call (8)

• together_ai/zai-org/GLM-5.3 added without a capability rule; review its tools/vision/reasoning support and add one
• together_ai/BAAI/bge-base-en-v1.5 is absent from the serverless catalog with no removal date in the docs; needs a human deprecation call
• together_ai/Qwen/Qwen2.5-7B-Instruct-Turbo is absent from the serverless catalog with no removal date in the docs; needs a human deprecation call
• together_ai/baai/bge-base-en-v1.5 is absent from the serverless catalog with no removal date in the docs; needs a human deprecation call
• together_ai/deepseek-ai/DeepSeek-V3 is absent from the serverless catalog with no removal date in the docs; needs a human deprecation call
• together_ai/moonshotai/Kimi-K2-Instruct is absent from the serverless catalog with no removal date in the docs; needs a human deprecation call
• together_ai/togethercomputer/CodeLlama-34b-Instruct is absent from the serverless catalog with no removal date in the docs; needs a human deprecation call
• together_ai/zai-org/GLM-4.6 is absent from the serverless catalog with no removal date in the docs; needs a human deprecation call

Catalog model types outside the sync's token-pricing scope, skipped: audio (5), image (29), transcribe (4), video (38)

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[greptile-apps]

Greptile Summary

The PR synchronizes Together AI model metadata in both copies of the model registry.

• Adds the GLM-5.3 catalog entry with pricing, caching, and token-limit metadata.
• Updates Qwen3.8 input, output, and cached-input pricing.
• Keeps the primary and backup registries aligned.

Confidence Score: 4/5

The GLM-5.3 output limit should be corrected before merging because the current metadata permits requests that the upstream model rejects.

The new entry conflates GLM-5.3's one-million-token context window with its 128K output ceiling, causing local validation and token budgeting to accept unsupported output sizes.

Files Needing Attention: model_prices_and_context_window.json and litellm/model_prices_and_context_window_backup.json

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Adds GLM-5.3 and updates Qwen3.8 pricing, but overstates GLM-5.3's maximum output length by using its full context window.
litellm/model_prices_and_context_window_backup.json	Mirrors the primary registry changes, including the incorrect GLM-5.3 maximum output-token value.

_{Reviews (1): Last reviewed commit: "feat(models): sync together_ai model reg..." | Re-trigger Greptile}

[greptile-apps]

GLM-5.3 output limit overstated

When a GLM-5.3 request specifies more than 131072 output tokens, this metadata treats the model’s full context window as its output ceiling, causing LiteLLM to accept and forward a request that the upstream model rejects.

Knowledge Base Used: Provider adapters and capabilities

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_together_registry_sync_2026-08-29 (994b2cf) with litellm_internal_staging (4d71441)}

[devin-ai-integration]

Superseded by the rolling registry PR #38693, which absorbs this sync's GLM-5.3 addition (with a max_output_tokens correction to 128000 and capability flags) and the Qwen3.8 repricing.