fix(deepseek): correct max_output_tokens for deepseek-v4 models

[New1Direction]

The deepseek-v4-flash and deepseek-v4-pro entries had max_output_tokens and max_tokens set to 8192, which is carried over from the older deepseek-chat models. DeepSeek's pricing page lists 384K max output for both v4 models (the 1M context window was already correct).

This corrects all four affected entries — deepseek/deepseek-v4-flash, deepseek/deepseek-v4-pro, and the non-prefixed deepseek-v4-flash / deepseek-v4-pro aliases — in both the root model_prices_and_context_window.json and the bundled litellm/model_prices_and_context_window_backup.json.

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

[CLAassistant]

All committers have signed the CLA.

[codspeed]

Congrats! CodSpeed is installed 🎉

🆕 16 new benchmarks were detected.

You will start to see performance impacts in the reports once the benchmarks are run from your default branch.

Detected benchmarks

• test_cost_per_token_openai: 629.4 µs
• test_get_model_info_with_provider: 79.9 µs
• test_token_counter_long_content: 1.7 ms
• test_get_llm_provider_anthropic: 147 µs
• test_get_model_info_openai: 84.2 µs
• test_cost_per_token_anthropic: 625.7 µs
• test_token_counter_with_tools: 417.4 µs
• test_token_counter_multi_turn: 563.3 µs
• test_get_llm_provider_azure: 147.9 µs
• test_token_counter_simple_message: 236 µs
• test_token_counter_raw_text: 183 µs
• test_get_llm_provider_openai: 143.4 µs
• test_get_model_cost_key_exact_match: 79.7 µs
• test_get_llm_provider_with_prefix: 140.3 µs
• test_get_model_info_anthropic: 83.3 µs
• test_get_model_cost_key_case_insensitive: 82.1 µs

---

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[greptile-apps]

Greptile Summary

Corrects stale max_output_tokens and max_tokens values for deepseek-v4-flash and deepseek-v4-pro (both prefixed and un-prefixed aliases) in the root and backup model pricing JSON files — the previous value of 8192 was inherited from older deepseek-chat entries, while DeepSeek's official pricing page specifies 384K for both V4 models.

• Updates four entries in model_prices_and_context_window.json and the matching four in litellm/model_prices_and_context_window_backup.json from 8192 → 384000.
• Adds targeted assertions to the two existing test_deepseek_v4_models_in_*_cost_map test functions, verifying both max_output_tokens and max_tokens equal 384,000 for all four model keys; no network calls are made.

Confidence Score: 5/5

Safe to merge — changes are purely data corrections to two JSON pricing files, verified against DeepSeek's official docs, with new test assertions to lock in the values.

All changes are confined to JSON metadata fields and additive test assertions. The corrected 384K value is confirmed by DeepSeek's pricing page (the same URL already cited in the JSON source field). Both JSON files are updated consistently, and the new tests exercise all four affected model keys in both cost maps without making any network calls.

No files require special attention.

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Updates max_output_tokens and max_tokens from 8192 → 384000 for all four deepseek-v4 model entries; values confirmed against official DeepSeek pricing docs
litellm/model_prices_and_context_window_backup.json	Mirror of the root JSON change — same four entries corrected identically to keep the backup in sync
tests/test_litellm/test_utils.py	Additive assertions appended to two existing test functions to verify the corrected 384K token limit for all four model keys in both the root and backup cost maps; no network calls made

_{Reviews (3): Last reviewed commit: "test(deepseek): assert max_output_tokens..." | Re-trigger Greptile}

[Sameerlite]

Thanks for the contribution! A couple of things to help get this ready:

• Could you add proof that the change works as expected? Screenshots, a sample request/response, test output, or a before/after comparison really help speed up review.

Also kicking off a fresh Greptile review on this.

@greptileai

[New1Direction]

Thanks! Here is a before/after using litellm.get_model_info (LITELLM_LOCAL_MODEL_COST_MAP=True):

# before (current data)
deepseek/deepseek-v4-flash → max_output_tokens = 8192
deepseek/deepseek-v4-pro   → max_output_tokens = 8192

# after (this PR)
deepseek/deepseek-v4-flash → max_output_tokens = 384000
deepseek/deepseek-v4-pro   → max_output_tokens = 384000

(max_input_tokens stays 1000000 in both cases.)

I also pushed test coverage. The existing test_deepseek_v4_models_in_cost_map / test_deepseek_v4_models_in_backup_cost_map checked the prices and max_input_tokens, but not max_output_tokens — the field this PR corrects. I extended both to assert max_output_tokens == 384000 for all four entries; the new assertions fail on the old 8192 data and pass on this change, so the value stays guarded going forward.

Source: DeepSeek lists 384K max output for both v4 models — https://api-docs.deepseek.com/quick_start/pricing

[Sameerlite]

Thanks for your patience on this one! Your changes look solid — the Greptile review is just stale because of commits since the last review. Kicking off a fresh pass now.

@greptileai

[Sameerlite]

Great work on this fix, @New1Direction — the data correction is well-documented and the new test assertions are a nice touch to lock in the values going forward! Just one thing blocking merge:\n\n- Merge conflicts — the branch has conflicts with the base branch. Could you rebase or merge the latest from litellm_internal_staging to resolve them?\n\nOnce that's done, this should be ready to go!