July 13, 2026. Lost in the noise of GPT-5.6 going public was another frontier release the day before it: Grok 4.5, launched July 8 as xAI's new model for coding, agentic tasks, and knowledge work. The launch got one news cycle and then the general availability wave buried it, which is itself a sign of where 2026 has landed: a major lab shipping its flagship family is no longer guaranteed a week of attention. For operators the release matters for a plainer reason. The cheap end of the frontier just got another serious option, and your routing math is out of date again.
What shipped
- The model and price. Per xAI's release notes, Grok 4.5 is available on the xAI API at $2 per million input tokens and $6 per million output tokens, with configurable reasoning effort (low, medium, or high, defaulting to high). The official announcement positions it for coding, agentic tasks, and knowledge work. Note the branding: after the SpaceX merger, the docs now describe it as SpaceXAI's model.
- Reported specs. Independent coverage and gateway listings report a 500,000 token context window and cached input at roughly $0.50 per million tokens, a 75 percent discount. xAI has said the model is built on its new V9 foundation and trained on real coding-agent data, a claim to read alongside SpaceX's acquisition of Cursor in June.
- Where it runs. Beyond the API, Grok 4.5 is the default model in Grok Build (xAI's coding agent), is available inside Cursor, is the default in xAI's Microsoft Office add-ins, and is listed on gateways including OpenRouter.
- Consumer access. Full access starts with the $300 per month SuperGrok Heavy tier, with a staged rollout to the $30 per month SuperGrok plan following the launch.
What it means for operators
The sub-$10 agentic tier is now genuinely crowded: GPT-5.6 Terra at $2.50 and $15, Luna at $1 and $6, Claude Sonnet 5 at its $2 and $10 introductory rate, and now Grok 4.5 at $2 and $6. Two moves follow. First, re-run cost per completed task on your two or three highest-volume automations, the same exercise the GPT-5.6 GA playbook called a tuning pass, and treat every vendor's benchmark claims as vendor-reported until you test on your own work. Second, notice the packaging: a frontier model arriving as the default inside an IDE, an Office add-in, and a coding agent on day one is the clearest sign yet that model choice is becoming a property of the tool, not a decision the user makes. If your business runs on AI automation, keeping the stack model-agnostic is what lets weeks like this one lower your costs instead of forcing a migration. Our AI engineers spend a lot of July doing exactly these swap-and-measure passes.
Frequently Asked Questions
On the xAI API it is $2 per million input tokens and $6 per million output tokens, with reasoning effort configurable at low, medium, or high (high is the default). Independent listings also report cached input at about $0.50 per million tokens. Consumer access is via SuperGrok Heavy at $300 per month, with a staged rollout to the $30 SuperGrok plan.
xAI positions it for coding, agentic tasks, and knowledge work. It ships as the default model in Grok Build, xAI's coding agent, and is available inside Cursor and as the default in xAI's Microsoft Office add-ins, which signals a focus on hands-on-keyboard work rather than chat.
On list price it sits at the cheap end of the frontier alongside GPT-5.6 Luna ($1/$6), Terra ($2.50/$15) and Sonnet 5's introductory $2/$10. Capability claims from all vendors are self-reported, so the honest comparison is cost per completed task on your own workloads, measured before you migrate anything.
Grok 4.5 launched the day before GPT-5.6 went public, and the GA wave dominated the week's coverage. That density of frontier releases is part of the story: pricing and defaults now change faster than most teams re-evaluate their stack, which is why a scheduled routing review beats reacting to headlines.