Model Registry and Provider API Audit — 2026-10-07

Status: implemented; authenticated vendor availability remains account-gated

This audit covers the current first-party model catalogs and materially new API surfaces for OpenAI, Anthropic, Google, and xAI. The wider registry still contains DeepSeek, Groq, Mistral, and Ollama records from the previous full catalog audit. Official model pages, API guides, pricing pages, release notes, and the installed first-party SDK types were treated as the sources of truth.

Implemented registry changes

OpenAI

Added GPT-6.1 Sol, GPT-6 Sol, and GPT-6 Luna alongside GPT-6 Astra, with verified reasoning efforts, endpoints, context/output limits, cache-write and long-context pricing, processing tiers, and Responses extension flags.
Added access-controlled GPT-5.6 Cyber (gpt-daybreak-red-latest) and GPT-Rosalind Research. Unknown published limits remain null; the registry does not invent them.
Added the gpt-daybreak-blue-latest alias for GPT-5.6 Sol.
Added GPT Image 2.5 Sunburst and Flare.
GPT-6 context guardrails record the documented 1,050,000-token total context window. The 128,000-token output cap remains separate.

Anthropic

Added Claude Opus 5.5 and Sonnet 5.5 as preferred public models with 1M input context, 128K output, current pricing, and model-specific effort metadata.
Marked Opus 5 and Sonnet 5 as legacy with 5.5 replacements. They remain callable; legacy is not treated as retired.
Retained Fable 5.1 and Mythos 5.1 lifecycle/access metadata from the September refresh.

Google

Added Gemini 3.8 Flash, Gemini 3.8 Live, and Gemini 3.8 Live Extended Thinking, including Interactions/Live endpoints, multimodal capabilities, processing tiers, and pricing.
Added Gemini Nano Banana 2.1 and Gemini 3.8 Flash/Flash-Lite TTS models.
Updated replaced Live entries to point to Gemini 3.8 Live.

xAI

Added Grok 4.7 with its 500K context window, supported reasoning efforts, Responses endpoint, cache and long-context pricing, and current hosted-tool capabilities.
Added Grok Voice Transcribe 2.0 and retained current xAI image, video, TTS, STT, and Voice Agent records.

Implemented provider/API changes

OpenAI

Upgraded to openai 7.30.
Managed Agent execution now preserves Responses reasoning and encrypted compaction items across tool turns. Direct and streaming continuations expose provider state without flattening it into visible text, and streaming replay retains the original cross-type output order.
Native tools cover web search, Code Interpreter, file search, remote MCP, computer use, hosted shell, apply patch, tool search, and image generation on capability-declared models. Direct responses retain native call payloads, and typed computer/shell/patch/tool-search output items let trusted hosts continue client-executed tool loops without raw SDK calls.
Added connector-first OpenAIDecisions, OpenAILiveSession, and OpenAIVoices wrappers.
Existing Astra support remains: async function/custom tools, in-band configuration_update, explicit compaction, Responses WebSocket steering, and misalignment alert retrieval. EU Fast/priority validation covers all four current GPT-6 models, and EU Ultrafast requests are rejected before inference.
Sampling validation follows the effective GPT-6 reasoning effort. Sol and Luna permit sampling only with none; Astra and GPT-6.1 Sol reject none.

Anthropic

Upgraded to @anthropic-ai/sdk 0.131.
Added the current web search (web_search_20260318), web fetch (web_fetch_20260318), and code execution (code_execution_20260521) wire versions.
Added deferred tool definitions, programmatic code-execution callers, regex/BM25 tool search, and server-side compaction request/response handling.
Added model-specific validation: Opus 5.5 is always adaptive; Sonnet 5.5 uses adaptive or between_tools, with between_tools limited to high or lower.
Claude Opus 5 disabled mode is likewise limited to high or lower, and contradictory provider-neutral thinking configurations fail locally.
Unknown Anthropic provider blocks are stored as opaque provider state and replayed verbatim. This is required for correct tool-search and server-tool continuation, including streams.

Google

Upgraded to @google/genai 2.27.
Gemini 3.5+ continues to use Interactions by default. Interactions now map continuation tokens, Gemini 3.8 file search, computer use, and file-search usage telemetry.
Added connector-first GoogleLiveSession, including background interaction status, and GoogleVoices lifecycle access.
Google TTS distinguishes built-in names from custom voice_…, voicekey_…, and voices/… identifiers. Its local voice list remains deterministic while GoogleVoices owns the remote custom catalog.
Computer use intentionally remains host-executed. OneRingAI returns normal tool calls; the host owns confirmation, action execution, and screenshots.

xAI

Added a dedicated GrokTextProvider while retaining xAI's OpenAI-compatible Responses transport.
The adapter advertises and maps xAI web search, X search, and hosted code execution only for registry-declared models; preserves reasoning state through tool turns; and converts native tool counts and provider-reported cost ticks.
xAI-specific capabilities are not inferred from OpenAI model names.

Compatibility and safety decisions

Connector remains the only credential source. New high-level APIs accept a connector name or instance and never accept raw keys in request options.
Runtime native-tool validation is model- and adapter-gated. Unsupported tools fail before a provider call rather than silently falling back.
Provider-owned continuation blocks use ContentType.PROVIDER_STATE. They are retained for replay but excluded from user-visible text and compaction summaries.
Access-controlled models are represented in the registry even when a caller cannot invoke them. availability communicates that distinction.
Unknown limits and prices are null/omitted. The registry does not turn absent vendor data into synthetic zeroes or guessed values.
Google computer use and all consequential local actions remain host-owned; provider tool events are telemetry, not authorization.

API migration notes

Check agent.getAdvancedCapabilities(model) before exposing a native tool or Responses extension in a product UI.
Use response.continuation_token and pass it back as continuationToken for Gemini Interactions continuation.
Use runDirect()/streamDirect() when the host owns OpenAI async-tool or Gemini computer-use continuations.
Claude 5.5 applications should set explicit thinking.mode only when they need between_tools; ordinary adaptive mode is the safe default.
xAI callers can request { capability: 'x_search' } only through a Grok connector/model. The OpenAI adapter rejects it.
Registry aliases must be globally unambiguous within each registry; a canonical model ID can no longer also be claimed as another model's alias.
Media callers should continue using ImageGeneration, TextToSpeech, and SpeechToText; new media records are selected through their registries.

Validation

The implementation is covered by registry completeness/capability tests, provider request/response and streaming tests, connector-first wrapper tests, runtime reasoning-map tests, TypeScript, ESLint, examples compilation, package build/export verification, and generated API documentation. Live API access is organization-, region-, and credential-dependent and is intentionally separate from deterministic unit validation.

Official sources

OpenAI: latest model, models, pricing, changelog, async tool calling, steering, and reasoning.
Anthropic: model overview, Opus 5.5, Sonnet 5.5, tool use, and pricing.
Google: models, Interactions, Live, computer use, file search, and pricing.
xAI: models, Grok 4.7, tools, pricing, and release notes.