Grok 4’s July launch put tool use, subscriber access and an API on the same release record
xAI announced Grok 4 on July 9, 2025, offering it to SuperGrok and Premium+ subscribers and through its API, with Grok 4 Heavy as a separate tier.
Historical reporting: xAI announced Grok 4 on July 9, 2025. Its post said Grok 4 was available to SuperGrok and Premium+ subscribers and through the xAI API, and introduced a SuperGrok Heavy tier for Grok 4 Heavy. xAI described reinforcement-learning training, native tools, a 256,000-token API context window and release-era benchmark scores. Those are xAI’s claims and published conditions, not independent evidence of performance, safety or enterprise deployment. An August 2025 xAI model card is later documentation and is used only to distinguish the later Web/API evaluation record from the July event.
Key takeaways
- The July announcement combined consumer subscription access, API access and a separate Heavy offering without making those routes interchangeable.
- Native tool use described a model capability and service workflow, not a guarantee that web or code results were accurate.
- xAI’s benchmark and training-scale statements need their stated tests and do not predict results for another application.
The launch named distinct ways to reach Grok 4[source 1]
On July 9, 2025, xAI announced Grok 4 and said it was available to SuperGrok and Premium+ subscribers and through the xAI API. The same announcement introduced a SuperGrok Heavy tier with access to Grok 4 Heavy. These are related but distinct product facts. A subscriber-facing model picker, an API service and a higher-capacity tier can have different limits, pricing, interfaces and operational controls. The announcement establishes that xAI placed all three in the release record; it does not prove that every subscriber had identical access, that an API customer used the Heavy variant, or that a model was deployed through every partner xAI mentioned as a future route.
Tool-use training added a decision layer beside language generation[source 1]
xAI said Grok 4 was trained with reinforcement learning to use tools such as code interpretation and web browsing, and described real-time search across X, the web and news sources in the API. This is a meaningful architecture and workflow claim: a model can select queries or invoke a tool before producing an answer. It does not establish that every retrieval source is current, complete or authoritative, or that a tool trace turns an answer into verified reporting. Tool access can also be limited by product configuration, user authorization, network conditions and provider policies. The July article supports xAI’s stated native-tool direction while keeping factual verification, data access and human review outside the model claim.
Release-era scores and training-scale figures needed their published conditions[source 1]
The Grok 4 post connected the release to a 200,000-GPU Colossus cluster, reinforcement-learning work and results on benchmarks including Humanity’s Last Exam and ARC-AGI-2. These are vendor-published accounts of a training program and selected evaluations. They are informative when read with their stated tool settings, scoring protocol, date and comparison set. They are not independent measurements of a customer’s accuracy, cost or reliability, and they do not prove that a model will safely perform a consequential workflow. A benchmark can isolate a task while a deployment must also manage prompts, retrieved material, tool permissions, output review and user impact. The event is a release record, not a universal performance certificate.
The August model card documented a later evaluation record[source 1][source 2]
An xAI model card last updated August 20, 2025 describes evaluations for Grok 4 Web and Grok 4 API and the company’s stated approach to model-behavior testing. That is useful later documentation, but it does not change the July 9 event date, launch access routes or exact claims made at announcement. A model card can explain the provider’s evaluation method and deployment framing without certifying an individual application. It also cannot be used to infer a legal, security or safety outcome for a user’s configuration. The historical conclusion remains bounded: xAI announced Grok 4 with its named subscriber and API paths, tool-use claims and vendor-reported evaluation material; later documentation and later service changes belong on their own dates.
Sources
- xAI: Grok 4 (July 9, 2025)
- xAI: Grok 4 Model Card
Continue reading
Related reading
News analysisGrok 4 Fast combined quick and reasoning responses while xAI split API access into two models3 min read
News analysisxAI introduced Grok 3 as an early preview, with Think variants in beta while training continued3 min read
News analysisxAI released Grok Code Fast 1 for coding agents, with partner trials and an API route kept separate3 min read