Roundup: Akamai Inference, Mistral Large 4, Gemini 4 Carbon”

This week's coverage points to more choice in how teams serve and operate large models: managed inference is expanding, open weights are becoming easier to deploy, and cloud providers are adding accelerator variety. The same reporting also highlights the operating economics behind agentic workloads, from token budgets and API design to CPU orchestration and rack-level efficiency.

Akamai launches managed inference: Akamai launched a beta inference service with OpenAI-compatible APIs, serverless or dedicated deployment options, and 12 open-weight models.
Mistral Large 4 weights released: Mistral released the weights for its roughly trillion-parameter Mistral Large 4 model for download and deployment.
Oracle widens OCI accelerator choice: Oracle said OCI will add Google TPUs, Cerebras hardware and SambaNova accelerators for AI workloads.
Gemini 4 Carbon tops a benchmark index: Google's Gemini 4 Carbon led Artificial Analysis' Intelligence Index v5, while remaining limited to selected users.
Anthropic's managed-agent architecture: Latent Space examined how Anthropic separates agent harnesses from execution sandboxes in managed agents.
AWS plans big infrastructure spending: AWS outlined a major infrastructure-spending plan tied to cloud and AI demand.
Kimi K3.1 Pro lowers listed prices: Moonshot's Kimi K3.1 Pro combined a high benchmark position with lower listed token prices than its predecessor.
GitHub Copilot Fusion routes models at runtime: Copilot Fusion used runtime model orchestration to balance coding-agent performance and task cost.
Anthropic adds enterprise spend reporting: Anthropic introduced reporting and budgeting tools to help enterprises understand token spend and workflow results.
Measuring agentic infrastructure end to end: An Arm methodology highlighted end-to-end task completion, concurrency and rack economics for agentic AI infrastructure.

Akamai adds managed inference at the edge

Data Centre Dynamics reported that Akamai launched Akamai Inference as a beta cloud service with OpenAI-compatible APIs, serverless pay-per-use inference and dedicated deployments. The initial catalogue includes 12 open-weight models ranging from 8 billion to 2.8 trillion parameters, plus embedding models for retrieval-augmented generation. Akamai also announced an agent sandbox and expanded its GPU fleet, while its official release dated the announcement 27 October. [1] [2]

Mistral makes a trillion-parameter model deployable by its users

TechCrunch reported that Mistral released the weights for Mistral Large 4, allowing businesses and governments to download and deploy the roughly trillion-parameter model themselves. Mistral said ML4 used 47% fewer tokens per task than GLM-5.3 and made the weights available through Hugging Face and its API. The company positioned the release for organisations that want on-premises or sovereign control of cybersecurity workloads, while the performance and token-efficiency figures remain company claims. [3]

Oracle broadens the accelerator menu in OCI

The Next Platform reported that Oracle plans to add Google TPUs, Cerebras inference hardware and SambaNova SN50 RDUs to OCI in 2027. The article said the options are aimed at AI training, post-training and inference, including reasoning and agentic workloads, giving enterprise and HPC customers more hardware choices. It also described Oracle's wider data-centre investment and its effort to embed enterprise trust controls around AI applications. [4]

Gemini 4 Carbon leads a benchmark index but is not yet broadly available

Artificial Analysis reported that Google's Gemini 4 Carbon scored 59 and led its Intelligence Index v5, while ranking second on its Coding Agent Index behind Claude Opus 5.5. The model was being rolled out to selected users rather than made publicly available; standard pricing was listed at $4 per 1M input tokens and $20 per 1M output tokens, with a 50% launch discount and no confirmed end date. Artificial Analysis also reported a 1M-token context window and a 95% cached-input discount at standard pricing. [5]

Anthropic separates an agent's brain from its hands

In a Latent Space discussion with Anthropic engineering and product leaders, the platform work behind Opus 5.5 and managed agents was described as separating a persistent harness and session log from the execution sandbox. Credentials are held separately and attached through a proxy, so generated code can run without directly receiving the service credential. The discussion presented this separation as a way to isolate untrusted execution while pursuing lower latency and better reliability. [6]

AWS ties aggressive infrastructure spending to AI-led cloud demand

TechCrunch reported that Amazon expects about $220 billion in 2026 cash capital expenditure, with most of it allocated to AWS, as it expands infrastructure for cloud and AI demand. Andy Jassy said AWS had reached a $184 billion annualised revenue run rate and could grow by more than 40% in both 2027 and 2028, while investors remained concerned about the spending required. These are Amazon's forecasts and management statements, not a guarantee of future revenue or returns. [7]

Kimi K3.1 Pro pairs a high evaluation result with lower listed prices

Artificial Analysis reported that Moonshot's Kimi K3.1 Pro scored 55 on its Intelligence Index, ranking fourth overall in the evaluation and achieving the highest score among the evaluated models on Terminal-Bench 4.0. The 2.8T-parameter model has 104B active parameters, a 1M-token context window and listed pricing of $2/$10/$0.20 per 1M input, output and cache-hit tokens. The report said the weights had not yet been published and that the first-party prices were 33% lower than Kimi K3's. [8]

Copilot Fusion uses runtime routing to manage coding-agent cost

Artificial Analysis reported that GitHub Copilot Fusion selects and coordinates Anthropic and OpenAI models at runtime to balance performance, cost and latency. In its Coding Agent Index v2.0 test, Fusion scored 71 on Terminal-Bench 4.0, cost $5.59 per task and used 7.2 million tokens per task—about half the token use reported for Claude Code with Opus 5.5. The figures are Artificial Analysis' benchmark measurements, and token charges are based on the underlying models used. [9]

Anthropic adds reporting and budget controls for enterprise AI use

TechCrunch reported that Anthropic launched enterprise reporting tools showing team output, token cost and workflow breakdowns over a selected date range, with recommendations for improvement. The release also includes pooled group budgets, and Anthropic said the usage data would not be used for model training by default. The product is aimed at helping businesses evaluate the results and cost of deploying Claude-based tools. [10]

Agentic infrastructure needs end-to-end measurement

In a sponsored feature, The Next Platform described an Arm methodology for evaluating agentic AI infrastructure at core, system and rack levels. It emphasised successful task completion, concurrency, orchestration, retrieval, tool execution and sandboxing alongside component benchmarks, and reported an Arm-sanctioned comparison showing approximately 2x the performance per rack versus AMD EPYC 9755 for the tested CPU-visible workflow. The methodology and comparison are presented by Arm through the sponsored feature and should be treated as vendor-backed claims. [11]

Conclusion

Managed serving, open-weight deployment and accelerator diversity are expanding at the same time, giving infrastructure teams more ways to trade control, speed, cost and operational complexity. For agentic workloads, the evidence this week also supports measuring the whole path—from model calls and token spend through orchestration, sandboxing and rack-level capacity—rather than relying on a single model or component metric.

Host	LLM hosting plans	Model families
Google Cloud	Model Garden, Google Kubernetes Engine (GKE), Compute Engine	Llama, Mistral, DeepSeek, Qwen, gpt-oss, Gemma
Amazon AWS	Amazon Bedrock, Amazon SageMaker, Amazon EC2	Llama, DeepSeek, Qwen, Gemma, gpt-oss, Mistral

LLM Hosting Roundup: Akamai Inference, Mistral Large 4 weights and Gemini 4 Carbon (30 Oct 2026)

Akamai's new managed inference service, Mistral Large 4's open weights, Oracle's wider OCI accelerator menu, Gemini 4 Carbon topping the Intelligence Index, Kimi K3.1 Pro's lower prices and new tools for tracking agentic AI spend.