The week of November 16, 2026, delivered what may be the most consequential expansion of xAI’s developer platform this quarter: Grok 4.7 arrived with a 500,000-token context window, Grok Agents introduced goal-driven tool use, and Grok Voice 3.0 broadened speech workloads. xAI also retired grok-build-0.1, reduced agent-tool prices, and continued testing models reportedly associated with Grok 4.8.
Here is our professional analysis of what defined this week.
Grok 4.7: Context, Reasoning, and Tiered Pricing
According to xAI’s developer release notes, the new grok-4.7 API model accepts text-and-image input and offers a 500,000-token context window. Developers can also configure reasoning effort, giving applications a control point for balancing model deliberation against workflow requirements.
That combination targets document-heavy and multimodal systems. It does not eliminate the need for retrieval, segmentation, or input governance—it gives engineering teams a larger operating envelope in which to design them.
For prompts below 200,000 tokens, xAI lists prices of $2 per million input tokens, $0.50 per million cached input tokens, and $6 per million output tokens. The threshold matters because the release notes describe tiered pricing rather than one uniform rate across the full context window.
Enterprise cost models therefore need separate assumptions for ordinary prompts and workloads approaching the model’s maximum context. Cached-input pricing also creates a direct incentive to stabilize reusable instructions and reference material instead of repeatedly sending fully variable prompts.
Production evaluations should test grok-4.7 at the context lengths applications will actually use. A large advertised window is a capacity specification, not proof that every task benefits from filling it.
Grok Agents: Tool Use Becomes a Platform
How to Ship It reported the launch of Grok Agents, a developer platform for goal-driven agents that can browse the web, execute code, handle files, and call custom tools. This is a broader execution surface than a conventional prompt-and-response API.
The operational consequence is substantial. Web access introduces changing external content, code execution introduces runtime risk, file handling introduces data-governance requirements, and custom tools can connect model decisions to business systems. Each capability needs explicit permissions, logging, validation, and failure handling.
API Directory’s release tracking also says agent-tool prices fell by up to 50%, with charges capped at $5 per 1,000 successful calls. The same reporting says these tools support Grok 4.1 Fast models.
Pricing by successful call makes tool completion an observable budget unit, but it does not make unsuccessful attempts operationally free. Retries still consume time, model usage, and system capacity. Agent evaluations need to measure total attempts, successful calls, completion latency, and downstream correction work—not just the listed tool charge.
This is not merely another model endpoint. It is xAI’s move toward an agent runtime, and production adoption requires controls around every action the runtime can take.
Grok Voice 3.0: Speech Expands Without a Price Reset
xAI’s release notes say Grok Voice 3.0 became available at the same price as Grok Voice 2.0. That gives teams an upgrade path without a vendor-reported generational increase in the base price.
Meanwhile, API Directory’s tracking says xAI’s Speech-to-Speech API reached general availability. Speech-to-text also gained an adjustable voice-activity-detection threshold, allowing developers to tune when the system interprets speech as starting or stopping.
That threshold is a practical configuration control. A setting that is too sensitive can treat noise or pauses as speech boundaries, while a less sensitive setting can delay responses. Voice applications should test the control against their actual acoustic conditions rather than accepting one default across contact centers, mobile devices, and quiet office environments.
General availability changes the procurement conversation, but it does not replace workload testing. Voice systems still need latency measurements, interruption handling, transcript review, and escalation paths before they carry customer-facing interactions.
API Lifecycle: Build Retires as 4.20 Appears
xAI’s release notes state that grok-build-0.1 reached end of life on November 19 and is no longer served. The notes indicate that existing requests require no other changes beyond moving away from the retired model.
There is no grace period in that wording. Any deployment still naming grok-build-0.1 needs an immediate inventory and migration, including hard-coded identifiers, environment variables, test fixtures, and fallback routes.
Separately, API Directory’s release tracking indicates that Grok 4.20 and Grok 4.20 Multi-agent became live during the week. Available reporting provides fewer technical details than xAI’s release notes supplied for grok-4.7, so their precise capabilities and deployment status should not be inferred from the names alone.
Engineering teams should treat the two developments differently. The grok-build-0.1 retirement is an active compatibility issue; the Grok 4.20 listings are candidates for verification and controlled evaluation.
Grok 4.8: Reinforcement Learning Continues
Dealroom reported that Elon Musk said Grok 4.8 is in reinforcement learning, with experimental versions being soft-launched inside Grok Bot to assess practical usefulness. This describes an evaluation stage, not general API availability.
The soft-launch approach suggests xAI is collecting practical signals before a broader release, but no supported launch date, price, context limit, or API identifier was provided. Capacity planning should therefore remain anchored to models that xAI has documented as available.
This distinction matters during a fast model cycle. Experimental exposure can signal direction, but production roadmaps need published interfaces, lifecycle terms, and pricing.
What to Watch
grok-4.7at long context — Enterprise evaluations need evidence on quality, latency, and cost as prompts move toward the 500,000-token limit.Agent permission controls — Web browsing, code execution, files, and custom tools require auditable boundaries before Grok Agents can safely enter production workflows.
Tool-call economics — The $5 per 1,000 successful calls ceiling should be assessed alongside retries, model-token usage, and human correction costs.
Voice threshold tuning — Adjustable voice-activity detection needs workload-specific testing across noise levels, pauses, accents, and interruption patterns.
Grok 4.20documentation — Release tracking indicates availability, but enterprise adoption should wait for sufficient technical and lifecycle detail.Grok 4.8 status — Reinforcement learning and soft-launch testing do not establish general availability, pricing, or an enterprise migration deadline.
Check back next week for the next xAI Weekly.
