xAI Releases Grok Build 0.1 API for Agentic Coding at $1/M Tokens

Public beta of the same model powering the Grok Build CLI — priced at $1 per million input tokens and $2 per million output. xAI calls it extremely cost-effective, intelligent, and fast.

xAI launched the public beta of Grok Build 0.1 API, priced at $1 per million input tokens and $2 per million output. The model, designed specifically for agentic programming, is the same one that powers the Grok Build CLI. With 2,938 likes and over 511,000 views on its announcement, the release signals strong developer appetite for coding-specialized models that can iteratively plan, execute, and debug complex software tasks. xAI positions this as a cost-effective alternative in an increasingly crowded agentic coding market.

Claude Opus 4.8 Adds Mid-Conversation System Instructions Without Breaking Prompt Cache

Anthropic released Claude Opus 4.8 with a key capability: system instructions can now be added mid-conversation without interrupting the prompt cache. This means more cache hits, lower cost, and reduced latency for API requests. The feature addresses a long-standing friction point for developers building multi-turn agentic applications where context must evolve dynamically. The update arrives alongside broader Opus 4.8 improvements, including increased honesty and a roughly 4x reduction in code defect omission rate.

Step-3.7-Flash: 198B MoE Vision-Language Model Debuts with Day-0 vLLM and NVIDIA Support

Jieyue Xingchen's Step-3.7-Flash launched with immediate ecosystem backing. vLLM announced day-0 support, noting the model packs 198 billion parameters as a sparse mixture-of-experts architecture with roughly 11 billion active parameters per token. It supports native image and text input across a 256K context window — suited for long documents, multi-file repositories, and dense visual interfaces. Simultaneously, NVIDIA confirmed NIM and NeMo acceleration endpoints are ready, while the model also appeared on HuggingFace in GGUF quantized format for local hardware. Multiple integration partners — OpenRouter, Kilocode, ModelScope, and ZenMux — have already onboarded the model into their stacks.

Visa Invests in AI Coding Platform Replit to Power Agentic Payments

Visa has invested in Replit, exploring how the AI coding platform can enable agentic payments. The partnership aims to help developers build payment applications more efficiently by leveraging Replit's AI-powered development environment. The move signals growing financial sector interest in agentic AI infrastructure as payment workflows become programmable and autonomous.

Step-3.7-Flash GGUF Lands on HuggingFace for Local Hardware

Jieyue Xingchen released the GGUF quantized version of Step-3.7-Flash on HuggingFace, enabling users to run the model on their own hardware. The release aligns with the broader push toward local-first AI, allowing developers to download and run frontier models without API keys or cloud dependencies.

Anthropic Reports Annualized Revenue Run-Rate of $47 Billion

Simon Willison relayed Anthropic's self-reported annualized revenue growth, noting that Axios founder Jim VandeHei said he could not find any company in any industry in any era that has scaled organic revenue this quickly at this level. When Anthropic was at $30 billion, the claim already seemed extraordinary; at $47 billion, it defies comparison. The figure underscores the breakneck commercialization of frontier AI models.

GPT-5 Pro Series Remains Unbeaten on Single-Shot Hard Problems Since Last Summer

Ethan Mollick observed that GPT-5 Pro series models have consistently been the best at solving the hardest problems in a single attempt since summer 2025, with no real competition emerging in all that time. The observation highlights OpenAI's sustained lead in frontier reasoning despite an increasingly crowded model landscape.

Claude Dynamic Workflows Can Launch Hundreds of Subagents for Large-Scale Tasks

Commentator op7418 noted that Claude's newly released dynamic workflows may be more significant than the Opus 4.8 model update itself. The system extends concurrent subagent logic, potentially launching hundreds of subagents to tackle massive tasks such as researching an entire codebase or generating comprehensive reports in a single session.

DeepSeek's Infrastructure Engineering Is So Good the Industry Politely Pretends It Doesn't Exist

Teortaxes Tex remarked that DeepSeek is so excellent at infrastructure engineering that the rest of the industry has to pretend they are operating at a loss or that it simply is not happening. The observation points to the competitive tension around DeepSeek's cost efficiency and its underappreciated operational excellence in serving large-scale AI workloads.

France Releases Advanced Open-Source LLM Under Apache 2.0 License

An advanced large language model has been released by France under the permissive Apache 2.0 license, targeting both personal and enterprise use cases. The move represents a significant European contribution to the open-weight AI ecosystem, offering a sovereign alternative to American and Chinese frontier models.