fix(cost-map): add perplexity, openrouter, voyage and nebius models and fix registry metadata

[devin-ai-integration]

TLDR

Problem this solves:

• Seven Perplexity Agent API models and two OpenRouter models have no registry entry
• Three Voyage models listed on the Voyage pricing page have no entry
• Perplexity openai/gpt-5.6-sol is priced above the published Perplexity rate
• Several Gemini rows miss documented reasoning, caching or URL context flags
• Nebius Qwen3.8-27B is missing and three Nebius rows have wrong context limits
• Retired DeepSeek and Gemini models and some Azure rows carry wrong deprecation_date

How it solves it:

• Adds 12 entries from official Perplexity, OpenRouter and Voyage sources
• Corrects six Perplexity GPT-5.6 Sol price fields to the official table
• Adds Gemini, Nebius capability flags and fixes the Lyria 3.5 input limit
• Keeps the earlier Nebius limit and deprecation date fixes from this rolling PR

User Flow

Before: a proxy user calling the new Perplexity GPT-6.1 Sol model gets no cost, and GPT-5.6 Sol is overcharged

1. They call perplexity/openai/gpt-6.1-sol through POST https://litellm-domain/v1/responses
2. Cost lookup fails with "This model isn't mapped yet", so spend is logged as $0
3. A perplexity/openai/gpt-5.6-sol call is logged at $5/M input and $30/M output tokens

After: both calls are priced at the rates Perplexity publishes

1. They call perplexity/openai/gpt-6.1-sol through POST https://litellm-domain/v1/responses
2. Spend is logged at $2/M input and $10/M output tokens, $4/$15 above 272k
3. The perplexity/openai/gpt-5.6-sol call is logged at $4/M input and $20/M output tokens

Registry changes by provider

Perplexity

Source for every value: the raw HTML table at https://docs.perplexity.ai/docs/agent-api/models (Input, Output and Cache read columns, read row by row). Cache rows given as a percentage discount were converted from the input price

Added, all litellm_provider perplexity, mode responses:

Key	Input /M	Output /M	Cache read /M	Above tier
perplexity/anthropic/claude-fable-5-1	$10	$50	$0.25	none
perplexity/anthropic/claude-opus-5-5	$4	$20	$0.20	none
perplexity/openai/gpt-6.1-sol	$2	$10	$0.10 (95% off)	>272k: $4 in, $15 out, $0.20 cache
perplexity/openai/gpt-6-sol	$2	$10	$0.20 (90% off)	>272k: $4 in, $15 out, $0.40 cache
perplexity/openai/gpt-6-luna	$0.10	$0.50	$0.01 (90% off)	>272k: $0.20 in, $0.75 out, $0.02 cache
perplexity/google/gemini-3.8-flash	$0.75	$3.75	$0.075	none
perplexity/xai/grok-4.7	$2	$6	$0.50	>200k: $4 in, $12 out, $1 cache

The table notes that not every third-party model supports every feature, so no supports_function_calling or supports_web_search flags were set on the new rows

Changed: perplexity/openai/gpt-5.6-sol row says $4/$8 input and $20/$30 output at 272k with a 90% cache discount. input_cost_per_token 5e-06 -> 4e-06, output_cost_per_token 3e-05 -> 2e-05, cache_read_input_token_cost 5e-07 -> 4e-07, input_cost_per_token_above_272k_tokens 1e-05 -> 8e-06, output_cost_per_token_above_272k_tokens 4.5e-05 -> 3e-05, cache_read_input_token_cost_above_272k_tokens 1e-06 -> 8e-07. Every other existing Perplexity row was compared against the same table and matches

OpenRouter

Source: the official OpenRouter models API https://openrouter.ai/api/v1/models

openrouter/apodex/apodex-1.1-mini:free added: prompt and completion price 0, context_length 262144 -> max_input_tokens, top_provider.max_completion_tokens 235929 -> max_output_tokens/max_tokens, text only input, tools, tool_choice, structured_outputs and reasoning in supported_parameters

openrouter/unbiased/pareto-26.10-preview added: prompt 0.0000008, completion 0.0000032, input_cache_read 0.00000003, context_length 1048576, max completion 131072, input modalities text and image, tools and tool_choice in supported_parameters

Voyage

Sources: https://docs.voyageai.com/docs/pricing (price rows) and https://docs.voyageai.com/docs/reranker, https://docs.voyageai.com/docs/embeddings (context rows)

voyage/voyage-large-2-instruct added: $0.12/M, 16000 token context, 1024 dimensions. voyage/rerank-1 added: $0.05/M, 8000 token context. voyage/rerank-lite-1 added: $0.02/M, 4000 token context

Gemini API

Source for each flag: the raw HTML of the model page under https://ai.google.dev/gemini-api/docs/models/ (Capabilities table, "Supported" rows)

supports_reasoning true added to gemini/gemini-2.5-flash-native-audio-preview-12-2025, gemini/gemini-3.1-flash-live-preview, gemini/gemini-3.8-live and gemini/gemini-3.8-live-extended-thinking (Thinking: Supported)

supports_prompt_caching true added to gemini/gemini-3.8-flash-tts and gemini/gemini-3.8-flash-lite-tts (Caching: Supported, and both rows already carry a cache read price from https://ai.google.dev/gemini-api/docs/pricing)

supports_url_context true added to gemini/gemini-2.5-pro (URL context: Supported)

gemini/lyria-3.5 max_input_tokens 1048576 -> 131072 (Input token limit 131,072 on https://ai.google.dev/gemini-api/docs/models/lyria-3.5)

Nebius

Sources: https://tokenfactory.nebius.com/model-catalog.md and the public catalog API https://tokenfactory.nebius.com/api/public/models_info (both official Nebius)

nebius/zai-org/GLM-5.3-Flash gains supports_vision true: the catalog lists it as image2text with image in its use cases

nebius/Qwen/Qwen3.8-27B is new (earlier commit): input_cost_per_token 4.5e-07, output_cost_per_token 3e-06, max_input_tokens 262144 (catalog max_model_len), supports_function_calling and supports_reasoning from the catalog use cases, source https://tokenfactory.nebius.com/models/catalog/text2text/Qwen%2FQwen3.8-27B. It is text2text, so no vision flag

nebius/deepseek-ai/DeepSeek-V4-Pro-0813 gains max_input_tokens 979000. nebius/zai-org/GLM-5.3 goes from 1048576 to 1024000 for max_input_tokens and max_tokens. nebius/deepseek-ai/DeepSeek-V4.1-Flash goes from 1048576 to 1048000 for max_input_tokens. All match the catalog max_model_len

Greptile's note that Nebius wildcard discovery misses these rows is real but is a code issue: the discovery list only collects rows whose provider is nebius-chat-models, while all 60 existing Nebius chat rows use nebius. Changing that is out of scope for a data-only PR

Deprecation dates (earlier commit)

DeepSeek, source https://api-docs.deepseek.com/updates: deprecation_date 2026-07-24 added to deepseek-chat, deepseek-reasoner, deepseek/deepseek-chat and deepseek/deepseek-reasoner

Gemini API, source https://ai.google.dev/gemini-api/docs/deprecations: deprecation_date 2026-07-28 added to gemini/gemini-2.5-computer-use-preview-10-2025

Azure OpenAI, source https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/model-retirement-schedule:

Rows	Old	New
azure/gpt-4.1-nano, azure/gpt-4.1-nano-2025-04-14, azure/us/gpt-4.1-nano, azure/us/gpt-4.1-nano-2025-04-14, azure/eu/gpt-4.1-nano	2027-04-14	2026-10-14
azure/gpt-4o-transcribe	2026-12-31	2026-10-15
azure/gpt-realtime-2.1, azure/gpt-realtime-2.1-mini	2027-07-31	2027-06-25

Other open registry PRs

This is the rolling registry PR. Open PRs touching the registry were diffed field by field against their own merge base and re-checked against current main and official sources. None had a change that could be verified and was not already on main, so nothing was absorbed and no PR was closed

Dropped as unverifiable or already covered (data-only PRs):

• feat(models): add kimi-k2.5 to model cost map and backup #43879 (kimi-k2.5): moonshot/kimi-k2.5 already on main, 8192 output cap not in Moonshot docs
• feat(model_prices): add deepseek-v3 to model price maps (#43660) #43753 (deepseek-v3): https://api-docs.deepseek.com/quick_start/pricing has no deepseek-v3 row
• Add Tinker (Thinking Machines) models to model cost map #43859 (Tinker): tinker is not a LiteLLM provider
• feat(models): sync together_ai model registry #41775 (Together): deprecation-only change, out of scope here
• fix(prices): fix xai/grok-4 capabilities #41376 (xAI Grok 4 flags), fix: correct cache pricing for Claude 3 Haiku CRIS variants on Bedrock #40368 (Bedrock Claude 3 Haiku cache): target rows are no longer on main
• fix(pricing): restore Azure gpt-5.6 terra/luna rates to Azure's published prices #36274 (Azure GPT-5.6 tiers): values conflict with the Azure Retail Prices API https://prices.azure.com/api/retail/prices
• feat(fireworks): add GLM 5.2 Fast router and fix glm-5p2 context window #31460 (Fireworks GLM-5.2): new rows already on main, cache and context edits conflict with https://api.fireworks.ai/v1/serverless/models
• feat(pricing): add xAI Grok 4.3 on AWS Bedrock #30775 (Bedrock Grok 4.3): no matching model in Bedrock list_foundation_models
• fix(bedrock_mantle): correct grok-4.3 and gemma-4 token limits #36385, feat(bedrock): add bedrock_mantle google.gemma-4-e4b #33930 (Bedrock Mantle): no official per-token row found
• fix: correct embed-multilingual-light-v3.0 pricing (1000x too high) #32117 (Cohere embed light): https://cohere.com/pricing lists Model Vault hourly prices, not per-token prices
• fix(cost): price cached GLM 5.2 tokens for OpenAI-compatible routes #32842, feat: add qwen3.7-max model (Alibaba/Qwen via Anthropic protocol) #29920, fix(anthropic): set supports_web_search on current Claude models priced for web search #33920, feat(github_copilot): add gpt-5.6 luna, terra, sol #33966, feat(github_copilot): add missing models to registry and unit tests #29679, fix(scaleway): remove support for hcompany/holo2-30b-a3b #36084: duplicates, deletions, or no official per-model price to verify

Left alone because they change code or CI behavior and only touch the JSON as part of that: #35617, #43581, #43390, #43363 and the remaining non data-only PRs

ci_cd/cost_map_pins.json is not on main (it is added by #43390). This PR does not touch the Fireworks MiniMax M3 rows that file pins

Checked and not changed

Groq allam-2-7b: https://console.groq.com/docs/model/allam-2-7b lists pricing as pending, so it was not added. Together Kimi K3 and media models, Cohere new models (Embed v5, Command A variants, North): no per-token price on https://www.together.ai/pricing or https://cohere.com/pricing for the missing rows. Bedrock luma.ray-v2:0 and new regional inference profiles: no per-unit row matched with confidence on https://aws.amazon.com/bedrock/pricing/. Google Antigravity previews and aqa: no per-token price on https://ai.google.dev/gemini-api/docs/pricing. OpenRouter openrouter/auto-beta, openrouter/fusion, openrouter/pareto-code are routers with variable pricing, and the Lyria previews are music models, so they were not added

Anthropic models and prices (https://api.anthropic.com/v1/models, https://platform.claude.com/docs/en/about-claude/pricing), OpenAI (https://developers.openai.com/api/docs/pricing), DeepSeek (https://api-docs.deepseek.com/quick_start/pricing), Mistral (https://api.mistral.ai/v1/models) and xAI model lists all matched the registry on the rows checked. Mistral codestral-2508 keeps 128k since the model card wins over the API's 256k. Google image model limits where the API and model page disagree were left alone

Fireworks qwen3p8 rows: pricing was not published in https://api.fireworks.ai/inference/v1/models, the raw or r.jina.ai copy of https://fireworks.ai/models, or the rendered page, so nothing was added

Pre-Submission checklist

• The handful of test files covering my change pass locally
• My PR's scope is as isolated as possible; it only solves 1 specific problem

Screenshots / Proof of Fix

Probe: LITELLM_LOCAL_MODEL_COST_MAP=True python -c "litellm.get_model_info(<model>)"

Before (fcf8797)

1. perplexity/openai/gpt-6.1-sol -> no registry entry, so lookup fails
2. openrouter/unbiased/pareto-26.10-preview -> no registry entry, so lookup fails
3. voyage/rerank-1 -> no registry entry, so lookup fails
4. gemini/lyria-3.5 -> max_input_tokens 1048576
5. perplexity/openai/gpt-5.6-sol -> input 5e-06, output 3e-05

After (8ef6abd)

1. perplexity/openai/gpt-6.1-sol -> input 2e-06, output 1e-05, above 272k 4e-06 and 1.5e-05
2. openrouter/unbiased/pareto-26.10-preview -> max_input_tokens 1048576, input 8e-07, output 3.2e-06
3. voyage/rerank-1 -> max_input_tokens 8000, input 5e-08
4. gemini/lyria-3.5 -> max_input_tokens 131072
5. perplexity/openai/gpt-5.6-sol -> input 4e-06, output 2e-05

python3 ci_cd/check_files_match.py passes, the schema generator leaves the schema unchanged, tests/unit/test_model_prices_schema.py, tests/unit/test_model_cost_aliases.py and tests/unit/test_cost_map_guard.py pass (98 passed), and the cost map guard passes against main

Type

🐛 Bug Fix

Link to Devin session: https://app.devin.ai/sessions/d50c257f78224426bbe6e9a4db7c8264
Open in Devin Desktop: https://app.devin.ai/desktop/session/d50c257f78224426bbe6e9a4db7c8264?variant=devin

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring

[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]

Confidence Score: 4/5

[Medium risk] Updates model pricing and context window metadata.

The PR should not merge until the new Qwen model is available through Nebius wildcard discovery and provider-based listings.

Findings

1.  Nebius discovery misses Qwen ▶

Summary

The PR adds pricing and context metadata for a Nebius Qwen model and updates three Nebius context limits in both registry copies.

• The new model is not included in Nebius provider-based discovery or wildcard expansion.

_{Reviews (1) · Last reviewed commit: "fix(cost-map): add nebius qwen3.8-27b an..."}

[greptile-apps]

Nebius discovery misses Qwen When a proxy uses Nebius wildcard model discovery, this new model is omitted. The registry adds models to the Nebius discovery list only when their provider is nebius-chat-models, but this entry uses nebius. Direct lookup can price the model, yet nebius/* expansion and provider-based model listings cannot expose it.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_model_registry_audit_1790796018 (8ef6abd) with main (fcf8797)}