fix(registry): veo 3.1 pricing tiers + roll up open registry PRs (glm-5.2, Qwen3.8-Flash, gemma-4-31b, scribe_v2, fireworks/databricks deepseek v4) + deprecation dates

[devin-ai-integration]

TLDR

Problem this solves:

• Registry missing recently released models; consolidates open registry PRs into one rolling PR
• Fireworks DeepSeek V4 Flash twins advertised one price, billed another
• Veo 3.1 billed one flat rate regardless of video resolution
• Some deprecation dates announced by providers were missing

How it solves it:

• Adds Veo 3.1 resolution-tier pricing on gemini and vertex_ai entries (original scope of this PR)
• Aligns Fireworks DeepSeek V4 Flash 0731 bare key with its accounts/ twin at the published price
• Adds zai/glm-5.2, together_ai/Qwen/Qwen3.8-Flash, cerebras/gemma-4-31b, elevenlabs/scribe_v2, Databricks DeepSeek V4 entries
• Adds six provider-announced deprecation dates (Azure claude, Bedrock nova-sonic, Vertex gemini-live)
• Every value verified against the official provider page cited in each entry's source

User Flow

Before: a team routing Fireworks DeepSeek V4 Flash through the gateway is billed double what the gateway advertises, and Veo 3.1 spend ignores video resolution

1. They send GET https://litellm-domain/v1/model/info and read fireworks_ai/deepseek-v4-flash-0731 at $0.14/1M input, $0.28/1M output
2. They send POST https://litellm-domain/v1/chat/completions with that model; the x-litellm-response-cost header comes back computed at $0.22/1M input and $0.66/1M output, about 2x the advertised rates, so their spend forecasts built on model info are off by half
3. They generate an 8 second 4k video with vertex_ai/veo-3.1-generate-001; https://litellm-domain/ui/?page=logs shows $3.20 spend while Google bills the project $4.80, and an 8 second 1080p fast video logs $1.20 while Google bills $0.96

After: advertised, billed, and provider prices all agree, and Veo 3.1 spend follows Google's per-resolution rates

1. GET https://litellm-domain/v1/model/info lists fireworks_ai/deepseek-v4-flash-0731 at $0.22/1M input, $0.66/1M output, matching the Fireworks serverless pricing page
2. The same POST https://litellm-domain/v1/chat/completions bills at exactly those rates; the x-litellm-response-cost header now matches what model info advertises
3. The same 8 second 4k Veo generation logs $4.80 and the 1080p fast one logs $0.96 at https://litellm-domain/ui/?page=logs, matching Google's pricing page

Changes by provider

Google (Veo 3.1, original PR scope)

• gemini/veo-3.1-* and vertex_ai/veo-3.1-* (preview and 001): per-second video pricing split by resolution tier. Standard: $0.40/s base (720p/1080p) + output_cost_per_second_4k $0.60. Fast: $0.10/s base + output_cost_per_second_1080p $0.12 + output_cost_per_second_4k $0.30. The vertex_ai entries carry the same tier keys as the gemini ones, which also enables resolution inference on vertex video responses. Sources: https://ai.google.dev/gemini-api/docs/pricing, https://cloud.google.com/vertex-ai/generative-ai/pricing
• New test pins completion_cost for all eight gemini/vertex veo 3.1 entries at the published tier rates in tests/test_litellm/test_video_generation.py

Z.AI (absorbed from #38866)

• New zai/glm-5.2: $1.40/1M input, $0.26/1M cached input, $4.40/1M output, 1M context / 128K output. Source: https://docs.z.ai/guides/overview/pricing

Together AI (absorbed from #38946)

• New together_ai/Qwen/Qwen3.8-Flash: $0.15/1M input, $0.47/1M output, 1M context. Source: https://docs.together.ai/docs/serverless-models

Cerebras (absorbed from #38675)

• New cerebras/gemma-4-31b: $0.99/1M input, $1.49/1M output, 131072 context / 40960 output, tool + vision + reasoning flags per the official model endpoint. Source: https://api.cerebras.ai/public/v1/models/gemma-4-31b

Cerebras deprecation (absorbed from #38675, requested in #37184)

• cerebras/zai-glm-4.7: deprecation_date 2026-08-17 per the official deprecations page ("2026-08-17 Deprecated zai-glm-4.7", raw-HTML verified). Source: https://inference-docs.cerebras.ai/support/deprecation

ElevenLabs (absorbed from #38860)

• New elevenlabs/scribe_v2: $0.22/hour = $6.11e-5/second batch transcription. Scribe v2 Realtime ($0.39/hour) is a separate tier, not covered. Source: https://elevenlabs.io/pricing/api

Fireworks AI (absorbed from #38797)

• fireworks_ai/deepseek-v4-flash-0731: input $0.22/1M, output $0.66/1M, cache read $0.007/1M, brings the bare key in line with the already-correct accounts/fireworks/models/ twin. Source: https://docs.fireworks.ai/serverless/pricing
• Includes the source PR's pinned-pricing + twin-consistency tests in tests/test_litellm/test_fireworks_serverless_model_costs.py

Databricks (absorbed from #38933, field-level)

• New databricks/databricks-deepseek-v4-flash-0731 (2.0 DBU/M in, 4.0 DBU/M out, 0.4 DBU/M cache read) and databricks/databricks-deepseek-v4-pro-0813 (18.857 / 56.571 / 1.886 DBU/M), converted at $0.070/DBU. Source: https://www.databricks.com/product/pricing/foundation-model-serving
• Context/max output use the DeepSeek-published model limits (1M context, 384K max output): https://api-docs.deepseek.com/quick_start/pricing
• The source PR's other entries already exist on the base branch. Databricks publishes no cache-write rate for these models, so cache_creation_input_token_cost is set to the input rate (no cache-write premium), matching the existing convention for databricks gpt/gemini/kimi/glm entries and the databricks cost-calculator tests

Deprecation dates (provider-announced, raw-HTML verified)

• azure_ai/claude-sonnet-5 2027-06-30, azure_ai/claude-opus-5 2027-07-08, azure_ai/claude-opus-4-8 2027-09-01, azure_ai/claude-fable-5 2027-12-05. Source: https://learn.microsoft.com/en-us/azure/ai-foundry/concepts/model-lifecycle-retirement
• amazon.nova-sonic-v1:0 2026-09-14. Source: https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html
• gemini-live-2.5-flash-native-audio 2026-12-13. Source: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions
• All other OpenAI / Anthropic / Gemini / Vertex / Bedrock / Azure deprecation dates in the registry were audited against the official lifecycle pages and are already correct

Already covered by the base branch (source PRs closed as superseded)

• Add moonshot/kimi-k3 pricing #35592 moonshot kimi-k3: base already has richer entry matching https://platform.kimi.ai/docs/pricing/chat-k3
• fix(xai): correct grok-4.5 cached-input pricing #36413 xai grok-4.5 cached-input pricing: base matches https://docs.x.ai/docs/models
• fix(voyage): add voyage-4 embedding models to model cost map #32987 voyage-4 family: base matches https://docs.voyageai.com/docs/pricing
• fix(models): update deepseek-v4-flash and pro max tokens to 384k #32107 deepseek 384K max output: base already has 393216 (=384K) per https://api-docs.deepseek.com/quick_start/pricing

Dropped as unverifiable (left open, not absorbed)

• fix: set supports_vision to false for novita/openai/gpt-oss-120b #37585 (novita gpt-oss-120b supports_vision: false): Novita's official models API lists image in input_modalities, contradicting the PR
• docs: added prices for GLM-5.2 and deepseek-v4-flash-0731 on scaleway #38502 (Scaleway glm-5.2 / deepseek-v4-flash): pricing verified in EUR, but the exact API model IDs could not be confirmed from official Scaleway pages
• fix: correct embed-multilingual-light-v3.0 pricing (1000x too high) #32117 (cohere embed-multilingual-light-v3.0 price): Cohere no longer publishes per-token pricing for this model
• fix(anthropic): set supports_web_search on current Claude models priced for web search #33920 (anthropic supports_web_search on 32 entries): Anthropic's web-search doc does not enumerate supported models in its static content

Left alone (code/behavior or too large to re-verify field-by-field)

#38976, #38789, #38516, #38384, #38165, #37972, #37433, #38881, #36416, #36422, #36279, #36274, #35720, #33185, #33930, #31907, #31884, #31018, #29204, #32197, #26016, #32059

Relevant issues

Deprecation dates support #26900 (proactive model deprecation alerts)

Linear ticket

Resolves LIT-6408
Resolves LIT-6565

Pre-Submission checklist

• I have added meaningful tests: pinned twin/bare Fireworks prices, twin-consistency sweep, and Veo 3.1 tier-rate pins
• LITELLM_LOCAL_MODEL_COST_MAP=True uv run pytest -q tests/test_litellm/test_model_prices_schema.py tests/test_litellm/test_model_cost_aliases.py tests/test_litellm/test_fireworks_serverless_model_costs.py tests/test_litellm/test_video_generation.py passes locally
• python3 ci_cd/check_files_match.py passes (root and backup JSON in sync)
• My PR passes all required CI/CD checks
• My PR's scope is as isolated as possible: registry data + registry tests, no runtime code touched
• Greptile Confidence Score of at least 4/5
• 06d4521 passes /live-pr-risk

Screenshots / Proof of Fix

Live A/B on two local proxies, 2 uvicorn workers each, LITELLM_LOCAL_MODEL_COST_MAP=True, same config: fireworks_ai/deepseek-v4-flash-0731 (real Fireworks key), vertex_ai/veo-3.1-generate-001, vertex_ai/veo-3.1-fast-generate-001, gemini/veo-3.1-fast-generate-preview. Before leg on port 51185 at the merge base, after leg on port 20034 at the tip. The chat completion is a real paid Fireworks call.

Before (ec3f818)

Fireworks DeepSeek V4 Flash: advertised vs billed price

1. curl http://127.0.0.1:51185/v1/model/info -H "Authorization: Bearer $KEY" shows input_cost_per_token 1.4e-07, output_cost_per_token 2.8e-07, cache_read 2.8e-08
2. curl -D - http://127.0.0.1:51185/v1/chat/completions -d '{"model":"deepseek-flash","messages":[...]}' returns usage 18 prompt / 16 completion (0 cached) and header x-litellm-response-cost: 1.452e-05
3. 1.452e-05 = 18 x 2.2e-07 + 16 x 6.6e-07: the call billed at the accounts/ twin's rates, roughly 2x the 7.0e-06 the advertised rates predict

Veo 3.1 resolution tiers on /v1/model/info

1. Same curl /v1/model/info: veo-3.1-fast (vertex and gemini) shows output_cost_per_second 0.15 and no output_cost_per_second_1080p / output_cost_per_second_4k keys; veo-3.1-generate-001 shows flat 0.4 and no 4k key
2. Every resolution of an 8s standard generation would be billed $3.20 (Google bills $4.80 for 4k); an 8s fast 1080p would be billed $1.20 (Google bills $0.96)

After (06d4521)

Fireworks DeepSeek V4 Flash: advertised vs billed price

1. curl http://127.0.0.1:20034/v1/model/info -H "Authorization: Bearer $KEY" shows input_cost_per_token 2.2e-07, output_cost_per_token 6.6e-07, cache_read 7e-09, matching https://docs.fireworks.ai/serverless/pricing ($0.22 / $0.007 cached / $0.66 per 1M)
2. Same curl /v1/chat/completions returns usage 17 prompt / 16 completion (0 cached) and header x-litellm-response-cost: 1.43e-05
3. 1.43e-05 = 17 x 2.2e-07 + 16 x 6.6e-07: billed exactly at the advertised rates; advertised, billed, and provider prices now agree

Veo 3.1 resolution tiers on /v1/model/info

1. Same curl /v1/model/info: veo-3.1-fast (vertex and gemini) shows output_cost_per_second 0.1, output_cost_per_second_1080p 0.12, output_cost_per_second_4k 0.3; veo-3.1-generate-001 shows 0.4 base plus output_cost_per_second_4k 0.6, matching https://cloud.google.com/vertex-ai/generative-ai/pricing
2. LITELLM_LOCAL_MODEL_COST_MAP=True uv run pytest -q tests/test_litellm/test_video_generation.py passes, pinning 8s costs at $3.20 (standard 720p/1080p), $4.80 (standard 4k), $0.80 (fast 720p), $0.96 (fast 1080p), $2.40 (fast 4k) for all eight gemini/vertex veo 3.1 entries

Type

🆕 New Feature
🐛 Bug Fix

Caveats (if any)

Low

• Veo 3.1 logged spend shifts by design (the fix itself): fast tiers $0.10-0.30/s, standard 4k $0.60/s
• Fireworks DeepSeek V4 Flash advertised/estimated cost doubles to match what calls already billed (live A/B shows the before leg billing at the accounts/ twin's rate); dashboards built on model info shift, billed chat spend does not
• No live Veo generation in the proof (several dollars per video, data-only diff, tier mechanism pre-existing and tested); tier math pinned by tests over the real registry instead
• elevenlabs/scribe_v2 per-second rate rounds $0.22/hr to 6.11e-05 (exact 6.1111e-05, 0.18% under), repo decimal convention
• osv-scan fails repo-wide on a new browserslist advisory in the dashboard lockfile, unrelated to this diff and not a required check

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

Link to Devin session: https://app.devin.ai/sessions/a6934bd774524c00a458d340f440e9ac
Open in Devin Desktop: https://app.devin.ai/desktop/session/a6934bd774524c00a458d340f440e9ac?variant=devin

---

Note

Medium Risk
Registry-only, but it changes reported costs for Veo 3.1, Fireworks DeepSeek V4 Flash, and newly added models—downstream spend dashboards and alerts will shift without code deploys beyond the JSON sync.

Overview
Updates the shipped model cost registry (model_prices_and_context_window.json and backup) so billing and model metadata match current provider pricing and lifecycles.

Veo 3.1 (Gemini + Vertex) fast variants drop the default per-second rate (e.g. $0.15 → $0.10) and gain 1080p / 4k tier fields; standard variants add 4k tiers. Pricing source URLs move to the official pricing docs. A new test asserts completion_cost for create_video matches those tier rates.

New registry entries: Databricks DeepSeek V4 Flash/Pro (DBU-derived token costs), zai/glm-5.2, together_ai/Qwen/Qwen3.8-Flash, cerebras/gemma-4-31b, and elevenlabs/scribe_v2 (audio transcription).

Price corrections: fireworks_ai/deepseek-v4-flash-0731 input/output/cache-read rates are raised to match the published serverless page and the account-prefixed twin; tests pin twins and cross-check account vs bare keys.

Deprecation metadata: deprecation_date is set on Bedrock amazon.nova-sonic-v1:0, several Azure AI Claude SKUs, cerebras/zai-glm-4.7, and gemini-live-2.5-flash-native-audio.

Databricks cost-calculator tests include the two new DeepSeek models in the published DBU table.

^{Reviewed by Cursor Bugbot for commit 06d4521. Bugbot is set up for automated code reviews on this repo. Configure here.}

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• 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 all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ mateo-berri
❌ devin-ai-integration[bot]
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[devin-ai-integration]

PR #38990 has no labels (no enterprise), so it's out of scope — no GitHub or Linear changes made.

[greptile-apps]

Greptile Summary

Adds and updates provider registry metadata for Veo, Databricks, Fireworks, Z.AI, Together AI, Cerebras, and ElevenLabs, along with provider-announced deprecation dates.

• Adds resolution-tier pricing for Gemini and Vertex AI Veo 3.1 models.
• Adds new model entries and corrects Fireworks DeepSeek V4 pricing.
• Adds focused Databricks, Fireworks, and video-cost coverage while keeping the primary and backup registries synchronized.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Updates the primary model registry with new models, corrected pricing tiers, capability metadata, and deprecation dates.
litellm/model_prices_and_context_window_backup.json	Mirrors the primary registry changes and remains synchronized with it.
tests/test_litellm/llms/databricks/test_databricks_cost_calculator.py	Extends published-DBU pricing coverage to the two new Databricks DeepSeek models.
tests/test_litellm/test_fireworks_serverless_model_costs.py	Adds pinned pricing and twin-entry consistency checks for Fireworks DeepSeek V4.
tests/test_litellm/test_video_generation.py	Adds local cost-calculation coverage for Gemini and Vertex AI Veo 3.1 resolution tiers.

_{Reviews (5): Last reviewed commit: "fix(registry): add vertex veo 3.1 resolu..." | Re-trigger Greptile}

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing devin_ai_1788201394-veo31-pricing-tiers (06d4521) with litellm_internal_staging (6661e91)}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[ryanl-cerebras]

Thanks for rolling #38675 into this registry PR. I verified both registry files on the current head.

The cerebras/gemma-4-31b entry is present, but the other change from #38675 is still missing: cerebras/zai-glm-4.7 needs "deprecation_date": "2026-08-17" in both model_prices_and_context_window.json and litellm/model_prices_and_context_window_backup.json.

That field is absent from both the current base and this PR's diff. Could you include it before merging? The deprecation is tracked in #37184.

[devin-ai-integration]

@greptileai

[devin-ai-integration]

bugbot run

[devin-ai-integration]

Added deprecation_date: 2026-08-17 for cerebras/zai-glm-4.7 in both registry files, verified against https://inference-docs.cerebras.ai/support/deprecation.

[devin-ai-integration]

@greptileai

[devin-ai-integration]

bugbot run

[devin-ai-integration]

@greptileai

[mateo-berri]

@greptileai

[mateo-berri]

bugbot run

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 06d4521. Configure here.}

[mateo-berri]

LGTM