grok-4.3

Grok 4.3 is xAI's latest agentic flagship, available on Microsoft Foundry with the catalog model id `grok-4.3` (learn.microsoft.com/azure/ai-foundry/foundry-models/concepts/models-sold-directly-by-azure). On CallMissed you call `"model": "grok-4.3"` on `/v1/chat/completions` using our OpenAI-compatible API — the same integration path as OpenAI GPT models, without exposing Azure infrastructure details in the model string.

Microsoft documents Grok 4.3 as a preview chat-completion model with tool calling, text input and output, and deployment type "Global standard" across regions. The Foundry table lists up to 200,000 input tokens and 8,192 output tokens for deployments (verify limits in your Azure portal quota view). xAI and Microsoft marketing emphasize agentic execution: improved instruction following, multi-step tool use, structured outputs, and workflows that combine reasoning with actions such as search, code execution, or retrieval — ideal for autonomous agents that must decide when to call functions.

The Azure AI catalog highlights always-on reasoning, large-context understanding, and native function calling as core capabilities (ai.azure.com/catalog/models/grok-4.3). Teams choose Grok 4.3 when they want an alternative reasoning style to OpenAI/Anthropic models while staying inside a single CallMissed API key and billing envelope. Because the model is sold directly by Azure, it draws from the same credit-covered hosting path as other Foundry-first-party listings in our stack.

Pricing on CallMissed is $3.50 per million input tokens and $15.00 per million output tokens — premium relative to GPT-4.1 but competitive with other frontier agent models. Use Grok 4.3 for research agents, coding assistants with heavy tool use, orchestrators that fan out to multiple APIs, and analytics copilots that must cite intermediate steps. English is the primary language listed in Azure docs; evaluate quality before deploying customer-facing multilingual experiences.

Safety and compliance: Microsoft notes that Grok models may require additional system safety prompting versus other Azure Direct models and recommends reviewing the Foundry model card before production (ai.azure.com/catalog/models/grok-4.3). Content filtering policies from Azure still apply at the hosting layer. Test jailbreak resistance and refusal behavior for your domain.

Implementation tips: define JSON schemas for tools clearly; Grok 4.3 is aimed at reliable function calling. Stream responses for long agent traces. Keep human review on high-stakes actions even when tool calling is enabled. If you previously used OpenRouter's `x-ai/grok-4.20`, note that `grok-4.3` on CallMissed is a distinct Foundry-hosted deployment with its own context and pricing — update model strings explicitly rather than assuming equivalence.

Agent design patterns: Grok 4.3 targets tool-centric agents — define narrow tools with strict JSON schemas (`search`, `fetch_url`, `run_sql`, `send_email_draft`) rather than one mega-function. Let the model plan calls, then validate arguments server-side before execution. Log each tool invocation for replay debugging.

Foundry deployment details: Azure lists Grok 4.3 as Global Standard with rate limits published per region in the model reference table (TPM/RPM vary — consult current Microsoft docs). CallMissed abstracts quota but you may still receive 429 under heavy load; backoff and queue accordingly.

Benchmark context: xAI markets Grok 4.x family improvements in instruction following and multi-step reliability versus earlier Grok versions. Treat published marketing benchmarks as directional; run your own agent trajectories measuring task success rate, tool error rate, and human edit distance.

English-first but multilingual experiments: primary language on Azure card is English. For Hindi or other Indian languages, compare quality against Sarvam models on CallMissed before committing — Grok may be fine for code-switching chat but Sarvam tiers win many Indic benchmarks.

Cost control: at $3.50/$15 per million tokens, long agent traces add up. Summarize tool results before re-injecting into context. Prune conversation history with rolling summaries.

Compliance reminder: Microsoft explicitly notes Grok may need additional safety system prompts versus other Azure Direct models — review Foundry responsible AI guidance and test refusal behavior for finance/health use cases.

Legacy id note: unrelated to `x-ai/grok-4.20` on OpenRouter — update model strings explicitly when moving integrations to CallMissed Foundry-hosted Grok.

OpenAI-compatible request shape recap: POST `https://api.callmissed.com/v1/chat/completions` with Bearer `cm_*` key, JSON body `{ "model": "grok-4.3", "messages": [...], "tools": [...] }`. Responses follow the familiar `choices[0].message` pattern with optional `tool_calls`. Streaming uses SSE chunks identical to OpenAI clients — set `stream: true` in the OpenAI SDK. Structured outputs consumers should validate JSON after receipt; do not execute tool arguments without server-side schema validation.

Enterprise procurement teams often ask whether Grok on CallMissed is the same as xAI's consumer Grok — the Foundry-hosted weights and safety stack follow Microsoft's Azure AI catalog, not necessarily identical to xAI.com chat snapshots. Treat them as related but distinct products and benchmark your workflows.

Observability: emit OpenTelemetry spans around each completion and tool round-trip; Grok agent traces can span dozens of LLM calls. Aggregate metrics: `agent_success_rate`, `avg_tool_calls`, `p95_latency_ms`, `cost_usd_per_task`. Alert when tool error rates spike — often indicates schema drift in your API, not model regression.

Roadmap-friendly design: isolate model id strings in configuration so swapping `grok-4.3` for another agentic model during bake-offs does not require code changes — only config and eval harness updates.