fix: add Google models deep-research-preview-04-2026 and deep-research-max-preview-04-2026

[github-actions]

fix: add Google models deep-research-preview-04-2026 and deep-research-max-preview-04-2026

Closes #589

Source issue: #589

Summary

Field	Value
Provider	google
Primary model	deep-research-preview-04-2026
Changed models	deep-research-preview-04-2026 deep-research-max-preview-04-2026
Added models	deep-research-preview-04-2026 deep-research-max-preview-04-2026
Updated models	None
Verification sources	1 2 3 4 5

Verified metadata

Model	Display name	Parent	Providers	Format	Flavor	Token limits	Pricing	Lifecycle
deep-research-preview-04-2026	Deep Research Preview (04-2026)		google	google	chat	input=1048576, output=65535	n/a	multimodal=true; reasoning=true
deep-research-max-preview-04-2026	Deep Research Max Preview (04-2026)		google	google	chat	input=1048576, output=65535	n/a	multimodal=true; reasoning=true

Verification notes

Verification

Fields verified from official sources

Field	Value	Source
Model IDs	deep-research-preview-04-2026, deep-research-max-preview-04-2026	ai.google.dev/gemini-api/docs/models, docs/deep-research, docs/interactions
format	google	Google Gemini API model
flavor	chat	Accepts text/multimodal input, produces text output
multimodal	true	Docs: "text, images, documents, and PDFs supported"
available_providers	["google"]	Available through Gemini API; not listed on Vertex AI locations page
experimental	true	Preview models per naming convention and docs
reasoning	true	Multi-step agentic research with thinking summaries

Fields NOT published / omitted

Field	Reason
max_input_tokens	Google does not publish explicit inputTokenLimit for these agent models. Estimated usage (~900k for Max) implies ≥1M context window consistent with Gemini 2.5 family. Value 1048576 is inferred from Gemini 2.5 family, not a verified hard limit. Included in metadata to satisfy resolver requirements.
max_output_tokens	Google does not publish explicit outputTokenLimit for these agent models. Estimated output ~60k (preview) and ~80k (max) includes reasoning tokens from agentic loops. Value 65535 is inferred from Gemini 2.5 family. Included in metadata to satisfy resolver requirements.
input_cost_per_mil_tokens	No per-token pricing published for deep research specifically. Docs say "standard Gemini list rates" but don't specify which tier applies.
output_cost_per_mil_tokens	Same as above.
input_cache_read_cost_per_mil_tokens	Not published.
supported_regions	Not applicable; these are google-provider models, not vertex. Not listed on Vertex AI locations page.
parent	No parent model relationship; these are standalone agent model IDs.
reasoning_budget	Not documented for deep research agents.
deprecation_date	Not published; models are in active preview.

sync_models (LiteLLM) cross-check

Neither deep-research-preview-04-2026 nor deep-research-max-preview-04-2026 exists in the LiteLLM model_prices_and_context_window_backup.json catalog. No cross-check is possible. All proposed values are sourced from official Google documentation.

Important note for PR reviewers

These models use the Interactions API (/v1beta/interactions), not the standard generateContent endpoint used by other format: "google" models. The proxy may need specific handling for these models. The max_input_tokens and max_output_tokens values are inferred from the Gemini 2.5 family (not published explicitly for these models) and should be validated against the API's models.get response when possible.

sync_models vs proposed update

sync_models cross-check found differences. Official provider verification was used for the applied values, and sync_models discrepancies are listed below for review.

Model	Field	Proposed update	sync_models	sync_models source models
deep-research-preview-04-2026	catalog entry	present	missing	None
deep-research-max-preview-04-2026	catalog entry	present	missing	None

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
ai-proxy	Ready	Preview, Comment	May 21, 2026 6:49pm

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: df9c2b7a25

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "Codex (@codex) review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "Codex (@codex) address that feedback".

[chatgpt-codex-connector]

Route deep-research models through interactions API

These new catalog entries classify both deep-research IDs as normal format: "google" chat models, which makes the proxy send them to /models/{model}:generateContent / :streamGenerateContent in handleGoogleRequest (packages/proxy/src/proxy.ts, lines 3398-3415). For these model IDs, Google expects the Interactions API, so requests will fail at runtime (streaming and non-streaming) as soon as users select these models from the catalog. Either add model-specific Interactions routing before listing them, or keep them out of the supported model list until that path exists.

Useful? React with 👍 / 👎.