fix(pricing): add the Vertex AI Gemini Live GA models and correct the preview

[marty-sullivan]

TLDR

Problem this solves:

• Both GA Vertex Live models are absent, so sessions bill $0
• The existing preview entry under-prices text input by 40%
• Image input has no rate, so camera frames bill 6x under

How it solves it:

• Add gemini-live-2.5-flash and gemini-live-2.5-flash-native-audio
• Correct the preview to the published $0.50 input rate
• Set the $3.00 non-text input rate on the Live entries

User Flow

Before: a developer running a voice session against a GA Live model is billed nothing at all, so the model looks free and no budget ever trips

1. They open a WebSocket to wss://litellm-domain/v1/realtime?model=gemini-live-2.5-flash-native-audio and speak into it
2. The model answers in audio and the session reports 16 prompt and 60 completion tokens
3. They open https://litellm-domain/ui/?page=logs and the request is logged at $0.000000
4. Their key's budget never advances, so a voice workload runs unmetered against any spend cap

After: the same session is billed at Google's published Live rates

1. They open a WebSocket to wss://litellm-domain/v1/realtime?model=gemini-live-2.5-flash-native-audio and speak into it
2. The model answers in audio and the session reports 16 prompt and 60 completion tokens
3. https://litellm-domain/ui/?page=logs now shows the request at $0.000668, with the audio share dominating
4. Their key's budget advances, so a spend cap applies to voice the way it already does to chat

Relevant issues

Fixes #31087

Linear ticket

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

Screenshots / Proof of Fix

Vertex AI serves two GA Live models and neither had a cost map entry, so every Live session billed $0. This is the missing-entry half of #31087.

Which models actually exist. Probed by opening a real BidiGenerateContent socket and sending a setup per candidate, across 22 Vertex locations, with definitely-not-a-real-model as a negative control:

gemini-live-2.5-flash                 global only        (11 regional locations refuse it)
gemini-live-2.5-flash-native-audio    11 regions, not global
gemini-live-2.5-flash-preview-native-audio-09-2025   us-central1 only
definitely-not-a-real-model           absent everywhere

No Gemini 3.x Live exists on Vertex in any location tested, which matches Google documenting 3.x Live as Developer-API only. The two endpoints are disjoint: the half-cascade model is global-only and the native-audio model is regional-only.

Before, dev gateway at 7c6ba0a990, a real voice session on vertex_ai/gemini-live-2.5-flash-native-audio produced audio and a correct transcript, and its spend row read $0.00000000 because the model resolved to no entry at all.

After, dev gateway at ec9d9abfa2, the same sessions on both GA models bill from /spend/logs/v2:

  vertex_ai/gemini-live-2.5-flash               (16 in / 30 out)  $0.000318
  vertex_ai/gemini-live-2.5-flash-native-audio  (16 in / 60 out)  $0.000668

The preview correction. The existing gemini-live-2.5-flash-preview-native-audio-09-2025 entry prices text input at 3e-07. That is gemini-2.5-flash's rate, not the Live API's. Google publishes $0.50 per 1M for Live text input, and the entry's other three columns already match the Live table ($3 input audio, $2 output text, $12 output audio), so the one odd column reads as a copy from the wrong sibling. Corrected on that entry and its gemini/ twin.

The image rate. These entries had no input_cost_per_image_token, so image tokens fell through to input_cost_per_token and billed at the $0.50/M text rate instead of $3.00/M. That matters more on Live than the name suggests, because Gemini reports a Live video stream as a sequence of frames under the IMAGE modality; VIDEO never appears in a Live session's usage, so a camera feed is billed entirely through this one field.

The rate is Google's non-text Live input price, published as "$3.00 (audio / video)" on ai.google.dev/gemini-api/docs/pricing and matching the audio rate already in these entries. An independent check lines up with it: Google documents video input as "roughly 258 tokens per second", and a single 128x128 JPEG frame measured exactly 258 tokens against Vertex, so one frame is one second of video for billing. A four-frame session reported 1032 image tokens, exactly 4 x 258.

Measured on a deployment carrying these entries, reading back from /spend/logs/v2, with expected computed independently from each session's own modality split:

mode     Vertex reported promptTokensDetails    billed        expected
image    [{TEXT 7}, {IMAGE 258}]                $0.00082550   $0.00082550
frames   [{TEXT 8}, {IMAGE 1032}]               $0.00313800   $0.00313800

At the text rate those same sessions would have billed $0.00014900 and $0.00052400, so the missing field was a 5.5x to 6x under-bill on any session carrying frames.

supported_output_modalities on the native-audio entries is now audio only, because Vertex closes the socket when responseModalities asks for TEXT on those models, verified directly:

native-audio, TEXT only     -> REJECTED close=1000
native-audio, TEXT + AUDIO  -> REJECTED close=1000
native-audio, AUDIO         -> ACCEPTED (setupComplete)
half-cascade, TEXT only     -> ACCEPTED (setupComplete)

supported_regions comes from those same probes, ordered so a caller in an unsupported region falls back to us-central1 rather than an arbitrary first entry. Note it lists us-west4, which Google's own region table omits but which answers setupComplete, so the list is empirical rather than transcribed.

A live call is the proof here rather than get_model_info alone, since the failure mode was a billed-at-zero session rather than a lookup error.

Type

🐛 Bug Fix

Changes

Adds gemini-live-2.5-flash (half-cascade, global) and gemini-live-2.5-flash-native-audio (native audio, regional), each as a bare key and a vertex_ai/ key, in both cost map files.

Corrects the 09-2025 preview entry and its gemini/ twin: input_cost_per_token 3e-07 to 5e-07, and supported_output_modalities to audio only.

Sets input_cost_per_image_token to 3e-06 on the four GA Live keys, matching the audio rate, so frames are priced at Google's non-text rate rather than falling back to text. Applied to both the bare and vertex_ai/-prefixed keys, since the prefixed one is what get_model_info resolves for custom_llm_provider="vertex_ai".

Sets mode to realtime on the Gemini Live entries that said chat while listing only /v1/realtime in supported_endpoints, so health checks select the realtime probe. OpenAI's gpt-audio family is deliberately untouched: those genuinely also serve /v1/chat/completions, so chat is right for them. gemini/gemini-robotics-er-2-streaming-preview has the same mode mismatch and is left alone as out of scope.

No new test, per the cost-map convention; the existing cost suite is the verification.

Caveats (if any)

• input_cost_per_video_token stays unset; Live never reports VIDEO
• supported_regions on a bare Vertex key is inert for routing
• The AI Studio gemini/ image rate is out of scope here

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

[marty-sullivan]

@greptileai

[greptile-apps]

Greptile Summary

Adds pricing and capability metadata for the Vertex AI Gemini Live GA models and corrects existing Gemini Live metadata.

• Adds bare and provider-prefixed entries for the half-cascade and native-audio GA models.
• Adds image-token pricing for Live camera-frame usage.
• Corrects preview text pricing, output modalities, regions, and realtime model classification.

Confidence Score: 5/5

The PR appears safe to merge because no blocking failure remains in the review’s eligible findings.

No blocking failure remains.

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Adds and corrects Gemini Live pricing, capability, modality, region, and realtime metadata; no eligible follow-up finding was identified.
litellm/model_prices_and_context_window_backup.json	Mirrors the primary model-cost-map updates; no eligible follow-up finding was identified.

_{Reviews (4): Last reviewed commit: "fix(pricing): price Gemini Live image in..." | Re-trigger Greptile}

[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 cu-aaii:litellm_vertex_live_ga_pricing (9c538a0) with litellm_internal_staging (b9bff09)}

[marty-sullivan]

@greptileai

[devin-ai-integration]

Superseded by rolling registry audit #37902, which reverified and absorbed the provider-documented metadata changes at field level

[marty-sullivan]

@mateo-berri Could a maintainer reopen this? #37902 closed unmerged, so nothing was absorbed, and both GA Live entries are still absent from staging

[marty-sullivan]

Update: ab1b7bf since added gemini-live-2.5-flash-native-audio alone, without image or cache-read rates. The three vertex_ai/ entries here are still absent from main

[devin-ai-integration]

Rolling registry PR #39388 now carries the eight mode: realtime flips and the native-audio image price; the remaining entries stay here pending official pricing rows

[mateo-berri]

Reopened, sorry for closing. We've since added gemini-live-2.5-flash-native-audio to staging, so please rebase onto litellm_internal_staging keeping your rates

[mateo-berri]

Closing: the realtime mode flips and image token rate landed in #39388; the remaining entries and price cuts are not on the linked pricing pages.