AI・テクノロジー
xAI "Grok 3.5" Officially Launched — Real-Time Reasoning × Web Integration Reshapes Research Workflow Design
機械翻訳 / Machine-translated
On August 23, 2026, xAI officially launched its new model, "Grok 3.5." The most significant change is the adoption of a "Live Grounding" architecture that calls real-time web searches mid-step during reasoning (Chain-of-Thought). Rather than the conventional flow of "search first, then reason," the design inserts searches only where needed during the reasoning process itself — a direct fit for research and investigative work that simultaneously demands information freshness and reasoning accuracy.
According to xAI's official announcement, Grok 3.5 posted the following benchmark results:
Live Grounding works by dynamically inserting "search nodes" during the generation of thinking tokens, triggering a web search at the moment the model determines "my knowledge is outdated or uncertain." According to xAI's published materials, the hallucination rate on fact-verification tasks has reportedly dropped by approximately 41% compared to the previous Grok 3.
Verification posts flooded X immediately after the release:
"I asked Grok 3.5 for a competitive analysis report and it spontaneously pulled IR documents from multiple companies partway through. The fact that this runs without explicit search instructions means its architecture is fundamentally different."
(Corporate researcher, approximately 12,000 followers)
API access begins today. Pricing is set at $8.00 per million input tokens and $24.00 per million output tokens — on par with GPT-4.5 Turbo.
The timing of Grok 3.5's release is a continuation of the reasoning model race over the past several weeks. OpenAI released o4 and Google released Gemini 2.5 Ultra, leaving benchmark rankings essentially neck and neck across the field. The focus of differentiation has shifted from "accuracy" toward "information freshness" and "suitability for agent integration," and xAI has entered that arena with Live Grounding in hand.
With Grok 3, xAI had emphasized access to real-time data from X (formerly Twitter) as a key strength. With Grok 3.5, by embedding that capability into the model's internal architecture, the scope of search has expanded from X exclusively to the general web.
Conventional RAG (Retrieval-Augmented Generation) and tool-calling agents follow a sequential structure: search → retrieve → input → reason. Because Live Grounding inserts searches mid-reasoning, it operates more like the human research process of "formulating hypotheses while verifying them." Concretely, this is expected to reduce the number of steps required for analysis report generation tasks that involve competitive research, literature review, and fact-checking.
A context window of 512,000 tokens can handle roughly 400 pages of documents in a single pass. This puts it in practical range for tasks like multi-period comparisons of earnings reports, contract review, and cross-referencing long codebases — use cases where you want to hand over everything at once and ask questions. That said, whether reasoning accuracy is maintained over long contexts requires independent verification; relying solely on official benchmark figures would be premature.
At $8/$24 (input/output per million tokens), Grok 3.5 holds an advantage on input cost compared to Claude Opus 4.6 at $15/$75 and GPT-4.5 Turbo at $10/$30. It appears likely to enter consideration for enterprise procurement that prioritizes cost optimization.
The current API specification offers an OpenAI-compatible endpoint, allowing drop-in replacement from major frameworks such as LangChain, LlamaIndex, and Claude Agent SDK. The Live Grounding feature itself is controlled via a parameter at API call time (live_grounding: true), making it straightforward for agent pipeline designers to integrate into existing workflows.
Alongside Grok 3.5, xAI also announced an expansion of the enterprise "Grok for Teams" plan. At $40 per seat per month, it includes priority API access and audit logs. This move reflects awareness of corporate governance requirements and could signal a shift away from the previous "consumer-first" approach.
What stands out most in this announcement is not Live Grounding as a feature in isolation, but rather the design philosophy of "erasing the boundary between reasoning and search." A mechanism by which the model autonomously decides "I should search here, right now" has genuine practical value for agent designers, in that it eliminates one loop from the pipeline.
However, the fact that control over search timing remains a black box is a risk that cannot be ignored. Without guaranteed traceability of what was searched and when, deployment in work requiring auditing — legal, compliance, and financial reporting — will necessarily require caution. Whether xAI moves toward "providing search logs via API" is likely to be the deciding factor for enterprise adoption.
Now that competing models have converged on similar reasoning accuracy numbers, the battleground for differentiation has moved to "information freshness" and "cost of agent integration." From that perspective, Grok 3.5's positioning is clear. Whether Live Grounding delivers the accuracy improvements it claims on real-world tasks, however, is something we would prefer to assess after independent reproduction experiments emerge over the next two to four weeks.
The essence of xAI's "Grok 3.5" lies in a structural shift: the integration of real-time search inside the reasoning loop. Stepping back from the benchmark numbers race, its deliberate incorporation of "information freshness" into its design means the impact on research and investigative workflows is direct. The next focal points are the publication of traceability specifications for the Live Grounding feature and the results of independent evaluations. Enterprise procurement teams would be well advised to wait for the third-party benchmark reports expected to emerge this week before making any decisions.
This article was written by an AI writer (AI News) from the Mirai News editorial team.
