One platform. Every frontier model.

Claude Opus 5.5 is Anthropic's current flagship Opus model for complex agentic coding, deep analytical reasoning, and enterprise workloads — delivering significant improvements in test-time compute scaling, instruction precision, and tool-call reliability over Opus 5. 1M context window, 128K max output. Adaptive thinking is ON by default (omitting the thinking parameter runs adaptive); disabling thinking is only allowed at effort 'high' or lower. Raw thinking tokens are never returned. Full effort ladder (low/medium/high/xhigh/max), 512-token prompt-cache minimum, elevated cybersecurity safeguards. Separate rate-limit bucket from the Opus 4.x pool.

Meta Superintelligence Labs' frontier multimodal reasoning model for long-running agentic, multi-agent and coding work, served on the Meta Model API. 1,048,576-token context and a 131,072-token output ceiling. Text, image, PDF and video in; text out. Tool calling, structured output and prompt caching. Reasoning depth is set with 'reasoning_effort' (minimal/low/medium/high/xhigh, plus 'max' on this standard tier only); 'none' is NOT supported and returns HTTP 400, so the platform never sends it. Meta reports roughly 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 on the same work. The cheaper contributor tier is deliberately NOT registered: its discount is consideration for Meta's data-contribution terms, so selecting it would make customer prompts training-eligible.

Claude Sonnet 5.5 is Anthropic's current Sonnet model — the best combination of speed and intelligence for everyday coding, agent, and enterprise work. Successor to Sonnet 5 at the same input/output price, with cheaper cache reads and the same tokenizer. 1M context window, 128K max output. Adaptive thinking is ON by default; thinking: {type: "disabled"} returns 400 (the lowest setting is {type: "between_tools"}, accepted only at effort 'high' or lower). Effort defaults to 'high' with levels recalibrated from Sonnet 5. Forced tool_choice ('any'/'tool') returns 400. Setting temperature/top_p/top_k to non-default values returns 400. 512-token prompt-cache minimum; safety classifiers can return stop_reason 'refusal' (cyber, bio, frontier_llm, reasoning_extraction, general_harms).

Anthropic's current Haiku model for high-volume, latency-sensitive tasks (classification, extraction, routing, subagents). 1M context window, 128K max output. Adaptive thinking with the effort parameter (default medium); non-default temperature/top_p/top_k return HTTP 400.

OpenAI's GPT-6 generation flagship — their most capable model, built for the hardest end-to-end work, and state of the art on computer use, browsing, software engineering, cybersecurity, science and professional work. 1.05M-token context window (922K max input), 128K max output, advanced reasoning with a low/medium/high/xhigh/max effort ladder, image input, and hosted tool use (web search, file search, code interpreter, apply_patch, computer use, MCP). First model OpenAI classifies at the Critical cybersecurity capability level under its Preparedness Framework, and ships with its most extensive safety stack to date.

OpenAI's smartest and most precise model — uses extra compute to think harder and provide consistently better answers on complex tasks. Available for Responses API requests, including through the Batch API.

Balanced GPT-5.6 model for everyday work, competitive with GPT-5.5 while being ~2x cheaper. Previewed June 26, 2026; generally available July 9, 2026.

—

xAI's flagship Grok model for coding, agentic tasks, and knowledge work. 500k context, text and image input, function calling, structured outputs, and configurable reasoning effort (low/medium/high/xhigh, default high).

Mistral's flagship 675B MoE model (41B active). Open-weight (Apache 2.0) general-purpose multimodal model with 256K context and vision input. Released by Mistral on 2025-12-02; available on Vertex AI Model Garden as self-deploy (open-weight) since 2025-12-09; not currently offered as a fully-managed Vertex partner-model API.