SpaceXAI has released Grok 4.6 across Grok Build, Cursor, Grok Bot and its API, priced at $2 per million input tokens and $6 per million output tokens. The 36-page model card describes it as the latest model in a 1.5-trillion-parameter family, developed with Cursor and supplemented with anonymized Cursor workflow data to improve coding and agentic performance. The rollout prioritizes work before consumer surfaces: Grok 4.6 is the default model in Grok Build, Word, PowerPoint and Excel add-ins, available on every Cursor tier and reachable through gateways including OpenRouter, Vercel, Cloudflare, Snowflake and Databricks.

The model card offers a more concrete example than a benchmark chart. In five hours, an early checkpoint tested 297 chat-inference optimizations, discarded changes that failed to produce end-to-end gains and opened seven pull requests. SpaceXAI says three of those PRs now serve production traffic in Grok Chat, producing a 1.5% decode-throughput gain and a 3.1% prefill-throughput gain. Distributing the model through Cursor while learning from Cursor workflow data creates a notable loop: an agent can test, measure, reject false wins and move verified improvements into production.

Capability did not improve evenly across reliability. In internal tests, self-harm compliance rose from 0.5% to 3.7%, dishonesty from 0.67% to 3.8% and hallucination rate from 0.98% to 1.7%, even as broad harmful-request compliance and jailbreak resistance improved. Grok 4.6 is therefore most interesting as a package of model, agent surfaces, pricing and a distribution-data flywheel, while deployment teams still need to test failure modes against their own workflows instead of reading only the headline benchmarks.

Bloomberg reports that DeepSeek has created a verified WeChat account for a new Harness Team. The account belongs to a Beijing entity identified by Qichacha as controlled by DeepSeek and verified by Tencent. The company is hiring several roles to turn its models into agentic products. A harness is the software scaffolding around a model that lets an agent manage long tasks, call tools, preserve context and handle errors across multiple steps. It is also where Claude Code creates much of its difference from a direct model call.

DeepSeek has also updated V4 Pro with better agentic capabilities, but this is an update to a model released in April, not a completely new launch. Official documentation says V4 Pro supports a 1-million-token context, up to 384,000 output tokens, tool calls, the Responses API, and both OpenAI and Anthropic API formats. Current pricing is $0.435 per million uncached input tokens, $0.87 for output and $0.003625 for cache hits. DeepSeek also warns that pricing could rise significantly soon.

Taken together, the strategy has three layers: a low-cost model to reduce loop costs, API compatibility to lower switching friction, and a harness to control how the agent executes work. DeepSeek has not shown that it already has a complete Claude Code rival, but the direction is clear: token economics become harder to replace when they are packaged into a workflow users return to every day.

Voice AI agents don’t usually fail because of noise. They fail when another person starts talking. Denoising can’t remove background voices. When speech overlaps, speech-to-text picks up fragments of both, and the agent acts on a bad transcript.

Krisp Voice Isolation 2.5 sits before speech-to-text, tuned for what machines read, not just what sounds good to people. It separates the primary speaker from competing voices, background chatter, and room echo.

Across six STT engines and five conditions, it cuts word errors by 46% on average. With competing voices, errors drop from 36% to 11%.

Lightweight, CPU-based, and 15ms latency. Ships in the Krisp VIVA 2.5

Eric Zakariasson came away with a simple concept: do not try to make an agent “try harder” with longer prompts or motivational language. Design a verification loop that lets the agent prove the output is done. Start with the outcome and turn “production-ready” into observable criteria for both function and design.

A verification loop for agent work:

Short prompts can work when you are willing to delegate more taste and decision-making to the model. Acceptance criteria and a verification loop are what turn that autonomy into more trustworthy output.

A Google DeepMind study of 10,101 participants separates manipulative process from outcome and finds that effects vary by domain and geography. Finance was affected most and health least. The controlled, text-only experiments do not prove mass manipulation in the real world, but they offer a better framework for evaluating risk.

CNBC argues that AI infrastructure is pushing up electricity, chip, software and data-center costs before productivity gains arrive evenly. The lens helps separate costs that already exist from a payoff that still depends on whether companies actually redesign their workflows.

FT Lex views Nvidia through a paradox: customers increasingly cannot ignore AI compute, but they also struggle to finance all of that demand themselves. As the vendor helps build a financing layer for its ecosystem, growth and risk distribution become more tightly connected to financial engineering.

The engineering director for Claude Code and Cowork describes just-in-time planning, targeted human review and metrics such as onboarding ramp and PR cycle time as code generation accelerates. It is Anthropic's own team experience, but it offers concrete evidence of a broader shift: human expertise moves toward verification, security, taste and trust boundaries.