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

Project	Deployment	Actions	Updated (UTC)
docs	Ready	Preview, Comment	Jun 12, 2026 3:11am

PR Summary

Medium Risk
Large metadata changes affect model selection, sliders, and cost display; the Bedrock ID fix is correctness-critical for those models but low blast radius elsewhere.

Overview
This PR re-validates the static model catalog in models.ts and fixes a Bedrock runtime bug where geo inference profile prefixes were applied to models that only accept bare in-region IDs (Mistral, Cohere, Titan).

Catalog & capabilities: Pricing, updatedAt, context windows, maxOutputTokens, deprecation flags, and defaults are corrected across OpenAI, Anthropic, Azure, Gemini/Vertex, xAI, DeepSeek, Groq, Cerebras, Mistral, and Bedrock. Notable capability tweaks include Mistral temperature 0–1.5 (new agent-block slider), xAI/Groq 0–2, Together 0–1, claude-fable-5 native structured outputs, gpt-5.5-pro effort/verbosity aligned to docs, Azure 5.4 reasoning effort trimmed, and ollama-cloud no longer advertising unsupported tool choice.

Provider plumbing: getModelsWithTempRange01/02 is replaced by getModelsWithTemperatureRange(max) and MODELS_TEMP_RANGE_0_15; tests and landing-catalog assertions are updated to match.

^{Reviewed by Cursor Bugbot for commit ddadc30. Configure here.}

Greptile Summary

This PR performs a two-pass validation of all ~160 models across 12 providers, correcting pricing, deprecation flags, context windows, temperature ranges, and capability fields against live provider documentation. It also ships a real runtime bug fix in getBedrockInferenceProfileId, which previously prefixed 11 Bedrock models (Mistral text, Cohere, Titan) with geo-inference profile IDs that those models don't support.

• Bedrock geo-profile fix: GEO_PROFILE_UNSUPPORTED_MODEL_IDS set added; affected models now return bare in-region IDs. New unit tests cover the three branches (prefix required, already prefixed, unsupported).
• Temperature range parameterization: getModelsWithTempRange01/02 replaced by getModelsWithTemperatureRange(max); a third slider entry (0–1.5) added in agent.ts for Mistral models; MODELS_TEMP_RANGE_0_15 exported from utils.ts.
• Data corrections: ~30 models flagged deprecated; pricing updated for DeepSeek V4, grok-4.20 trio, Gemini Deep Research, Bedrock Mistral Large 3, Nova, and many others; context windows, maxOutputTokens, and reasoning-effort value sets corrected across Azure, Anthropic, xAI, Groq, and Vertex.

Confidence Score: 5/5

Safe to merge — the only runtime change is the Bedrock geo-profile fix, which is well-tested and corrects real invalid model IDs that were being sent to AWS.

The Bedrock bug fix is correct and covered by new unit tests. All data changes (pricing, deprecations, context windows) are documented in the PR with source verification. The rest of the changes are refactors and test updates that don't affect production paths.

apps/sim/providers/models.ts — the static temperature-range list helpers don't inherit provider-level capabilities, leaving an inconsistency with the per-model query functions, though this has no current production effect.

Important Files Changed

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Model ID lookup] --> B{getModelCapabilities}
    B --> C["Merge: { ...provider.capabilities, ...model.capabilities }"]
    C --> D[supportsTemperature / getMaxTemperature]
    D --> E[Agent slider condition — CORRECT for all models incl. Groq]

    F[Static list builders] --> G{getModelsWithTemperatureRange / getModelsWithTemperatureSupport}
    G --> H[Only model.capabilities.temperature checked]
    H --> I[MODELS_TEMP_RANGE_0_2 / MODELS_WITH_TEMPERATURE_SUPPORT — MISSES Groq provider-level caps]

    J[Bedrock model ID] --> K{getBedrockInferenceProfileId}
    K --> L{Already prefixed?}
    L -- yes --> M[Return as-is]
    L -- no --> N{In GEO_PROFILE_UNSUPPORTED_MODEL_IDS?}
    N -- yes --> O[Return bare model ID]
    N -- no --> P[Add region prefix e.g. us. eu. ap.]

_{Reviews (5): Last reviewed commit: "fix(providers): default azure-openai to ..." | Re-trigger Greptile}

Re: the two items flagged in the Greptile summary (posted as outside-diff comments, so answering here):

1. DeepSeek cachedInput $0.0028 (50x discount) — verified, not a decimal error. Re-fetched https://api-docs.deepseek.com/quick_start/pricing directly: deepseek-v4-flash is $0.0028/1M on cache hit vs $0.14/1M on cache miss, $0.28/1M output (and v4-pro is $0.003625 hit vs $0.435 miss — an even steeper ~120x). DeepSeek moved away from the old 10x ratio with the V4 generation. The deepseek-chat/deepseek-reasoner aliases bill at v4-flash rates until their 2026-07-24 retirement.

2. Renamed IDs (gemini-3.1-flash-lite-preview → gemini-3.1-flash-lite, azure/gpt-5-chat-latest → azure/gpt-5-chat) — runtime routing does not fail for persisted workflows. getProviderFromModel falls back to provider modelPatterns (/^gemini/, /^vertex\//, /^azure\//) when an id isn't in the catalog, and the Azure deployment name is derived from the stored model string (request.model.replace('azure/', '')), so a workflow that stored an old id routes and executes exactly as before — it only loses catalog pricing/capability metadata. Aliases weren't kept because each old id is dead or dying upstream: the google preview was shut down server-side 2026-05-25 (calls fail regardless of our catalog), the vertex preview alias is discontinued by Google on 2026-07-09, and gpt-5-chat-latest never matched an Azure model name. Keeping deprecated alias entries would add permanent catalog noise for at most a few weeks of metadata continuity — judged not worth it, but happy to add them if you disagree.

@greptile

@cursor review