xAI launches Grok 4.3 with API and Bedrock access

xAI launched Grok 4.3, and its documentation lists the model name as “grok-4.3” and the alias as “grok-4.3-latest.”[1][2] VentureBeat reported availability through the xAI API and partner OpenRouter.[3] xAI says the model is generally available on Amazon Bedrock, while SpaceXSI says Bedrock access is available to developers in supported AWS regions.[4] The documentation lists availability in the us-east-1 region and support for the Batch API.[2] Teams choosing a route should verify that the relevant region and interface match their deployment requirements.

Grok 4.3 accepts text and image inputs, can be connected to external tools and systems, and supports structured outputs in specified formats.[2] Its documentation lists reasoning efforts of none, low, medium, high and xhigh, with low as the default.[2] Those options give developers several configuration points to test when comparing response quality, format reliability and tool use in their own tasks. The documentation describes the model as fast and reliable and highlights tool-calling and instruction-following capabilities; those descriptions are xAI’s characterization, not a guarantee of results for a particular workflow.[2]

The listed rates are $1.25, $2.50 and $0.20 per 1 million tokens.[2] The evidence available here does not identify which rate corresponds to which token category, so teams should confirm the rate breakdown in the pricing documentation before building a cost estimate. xAI says it charges higher rates for requests exceeding the 200,000-token context threshold.[2] That threshold makes prompt length a relevant input to budgeting; test expected workloads, including longer requests, rather than relying on a single per-token figure.

Artificial Analysis reported a score of 53 for Grok 4.3 on its Intelligence Index, four points above the latest Grok 4.20.[1] On GDPval-AA, it reported an Elo of 1,500, 321 points above Grok 4.20 0309 v2.[1] Artificial Analysis also reported that Grok 4.3 maintained an 81% IFBench score from that comparison model.[1] Its analysis found an eight-point gain in AA-Omniscience Accuracy alongside an eight-point drop in Non-Hallucination Rate.[1] These results offer comparison points, but the paired change is a reason to test factual reliability on the tasks a team actually handles.