xAI's Grok 4.3 Now Generally Available on Amazon Bedrock

xAI's Grok 4.3 is now available on Amazon Bedrock, officially making xAI a model provider on the platform.

The key features of Grok 4.3 include:

million token context window**: Supports long-document processing and multi-turn conversations

Configurable reasoning effort**: The depth of reasoning can be adjusted per request across four levels: `none` / `low` / `medium` / `high`. You can use `none` for low-latency simple classification and `high` for deep reasoning like contract analysis

Text and image input support**: Capable of multimodal processing

Strong tool use**: Excels at function calling and instruction following, making it well suited for building agents

According to xAI, the model achieves the lowest hallucination rate on the Artificial Analysis Omniscience benchmark and also ranks first on the Tau2 Telecom benchmark (tool use and customer support). The company also claims it delivers "2-10x more intelligence per dollar" compared to other frontier models.

Notably, Grok 4.3 runs on Mantle, Amazon Bedrock's next-generation inference engine. Mantle uses an OpenAI-compatible API, so it is accessed differently from the existing Amazon Bedrock Runtime API. It can be called directly via the OpenAI SDK or the Chat Completions API.

Enterprise use cases envisioned for the model include contract review, credit agreement analysis, and Q&A over financial documents.

The "Compute Gap": Enterprise AI Infrastructure Investment Outpacing Cost Visibility

A survey conducted by VentureBeat of 107 enterprises (with 100 or more employees) has highlighted accelerating AI infrastructure investment paired with lagging cost visibility — a phenomenon it calls the "compute gap."

Key findings include:

Only 21% run AI at scale in production**, yet investment appetite far exceeds that figure

% of companies have GPU utilization at or below 50%**: Compute resources are significantly underutilized

Fewer than 44% rigorously track AI compute costs**

% of companies plan to switch or add infrastructure providers within 12 months** (38% within a quarter)

The top criteria for provider selection are "integration with the existing stack" (41%) and "total cost of ownership" (35%). Token price accounts for a mere 8%

Particularly noteworthy is that "AI-specialized clouds" (45%) top the list of areas enterprises plan to evaluate next. This is a layer that most companies do not currently use, suggesting that the present landscape of hyperscalers and model provider APIs could shift dramatically in the near future.

The survey also found that only about one in five companies are aware of the critical inflection point at which, as inference scales, the bottleneck shifts from GPU compute to memory bandwidth.

Russian Elite Hacker Group "Sandworm" Adopts ClickFix Attacks

Ukraine's CERT (Computer Emergency Response Team) has warned that Sandworm, a sophisticated hacking unit within Russia's military intelligence agency GRU, has adopted an attack technique known as ClickFix.

ClickFix works by displaying a fake CAPTCHA on attacker-controlled websites and prompting visitors to copy and paste text. That text contains a malicious script that installs malware on the device or steals data.

While this technique has mostly been used by financially motivated criminals, its adoption by a state-level hacking group raises the threat level. The attacks began in the spring and continued through the summer, with network compromise confirmed at at least one organization. Devices infected with Sandworm's custom malware "FreakyPoll" were discovered, and ten compromised websites containing PowerShell commands displayed as fake CAPTCHAs were identified.

Although not an attack targeting AI, this is a telling example of how social engineering attacks are growing more sophisticated and diverse as AI adoption spreads.

Google Gemini Next Version Launch May Be Delayed

A report that the launch of Google's next Gemini version is being delayed due to missed internal targets has been drawing attention in the LocalLLaMA community on Reddit. At this point, details are limited — only a link to a Reddit post has been shared and technical information is scarce — but amid the recent momentum of Google's AI rollout, including Gemini Spark, a delay to the next model could affect the competitive landscape.

We'll cover this separately as more details emerge.

ChatGPT Ads Begin Serving in Japan

ChatGPT ads have been served to users in Japan since June 19, 2026. The audience is Free plan users, with ads appearing in ChatGPT's response area, sidebar, and elsewhere.

OpenAI began serving ads in the US at the end of 2024, and Japan is positioned as one of the destinations of that expansion. A Japanese-language article summarizing the ad delivery mechanism and placement details has been published on Qiita.

How this new format of advertising on AI assistants will affect user experience and marketing practices is a key point to watch going forward.

---