Get tomorrow's brief in your inbox
Today: Moonshot AI's Kimi K3 (2.8 trillion parameters) outperforms Claude Opus 4.8 and GPT-5.5 on coding benchmarks at half the cost of Opus. 1Password shipped zero-exposure credential access for Claude agents so they can book flights and manage accounts without exposing your passwords. Enterprises are deploying AI agents faster than they can secure them, with 54% reporting security incidents and most agents still sharing credentials.
Moonshot AI releases Kimi K3, China's largest AI model, priced competitively with Claude Sonnet
Chinese startup Moonshot AI unveiled Kimi K3, a 2.8 trillion parameter model that challenges the dominance of U.S. AI labs. The model beats Claude Opus 4.8 and GPT-5.5 on coding and agent benchmarks, though it still trails Claude Fable 5 and GPT 5.6 Sol overall. At $3/million input tokens and $15/million output tokens, it matches Claude Sonnet's pricing while delivering performance that approaches top-tier U.S. models.
For small businesses, this matters on two levels. First, competitive pressure from Chinese AI labs is driving down prices across the board. A year ago, this level of performance would have cost 3-4x more. Second, the model shows that pre-training scaling paired with architectural innovation can deliver step-change gains even under compute constraints. Moonshot raised $2 billion at a $20+ billion valuation in May, backed by Alibaba and Tencent.
The open-weight version promises to launch by July 27, which could make this capability accessible for self-hosted deployments. If you've been priced out of Claude Opus for specific tasks, K3 offers a middle ground. Bank of America analysts noted that despite China's hardware constraints, K3 demonstrates that innovation can still drive gains at the frontier. Watch how this pricing pressure shapes the next wave of AI tool costs.
1Password ships zero-exposure credential access for Claude agents
1Password for Claude lets Claude access your stored credentials without exposing passwords to Anthropic's AI models. The system works by injecting credentials through a secure channel that Claude can't view. You approve each request with a biometric prompt, and 1Password scans the page after every autofill to ensure nothing remains exposed before returning control to Claude.
This is agentic commerce infrastructure. If you use Claude to book travel, manage subscriptions, or automate multi-step workflows, this removes the friction of manually entering passwords without compromising security. Available now on Mac for business, family, and individual plans. You need the 1Password desktop app, browser extensions, and Claude desktop app. Support for payment cards and identity details is coming after launch.
The "zero-exposure security framework" is the unlock here. Agents get task-specific access to credentials, but nothing else in your vault is reachable. For businesses running AI agents that need to interact with third-party services, this pattern shows how to grant access without creating credential sprawl.
Thinking Machines Lab releases Inkling, a 975B open-weights multimodal model
Mira Murati's Thinking Machines Lab released Inkling, an Apache-2.0 licensed model with 975B total parameters (41B active). It's a Mixture-of-Experts transformer trained on 45 trillion tokens of text, images, audio, and video. The lab also promises Inkling-Small (276B/12B active) once testing completes.
Inkling isn't a frontier model by their own admission. It's designed as a strong base for fine-tuning on their Tinker training platform. If you've been waiting for a viable U.S. open-weight alternative to Chinese models, this joins NVIDIA Nemotron and Gemma 4 as a contender. The Apache-2.0 license means you can use it commercially without restrictions. Multimodal capabilities, efficient thinking, and availability for customization make it worth evaluating if you need domain-specific AI without vendor lock-in.
DoorDash CLI: Order food from your AI agent
DoorDash launched a command-line tool (dd-cli) that lets developers integrate DoorDash ordering into AI agents. The beta is open to U.S. and Canadian macOS developers via waitlist. You can search stores, find deals, and check out programmatically.
On the surface, this is funny (sudo make me a sandwich). Underneath, it's agentic commerce infrastructure. DoorDash is exposing its ordering platform to AI agents, which means developers can build custom tools for ordering food, groceries, or finding deals. If you're building AI workflows for teams, this is an example of how third-party services will expose APIs designed for agents, not just humans. DoorDash already has an AI chatbot ("Ask DoorDash") and integrations with ChatGPT and Claude. This CLI extends that pattern to developer-built agents.
Cars24 handles 1M+ monthly conversation minutes with OpenAI-powered agents
Indian automotive marketplace Cars24 deployed AI voice and chat agents across buying, selling, financing, and support workflows. The agents handle 1 million+ conversation minutes per month, recovered 12% of previously lost seller leads through AI-powered re-engagement, and reduced turnaround time by 80% across key service workflows.
The ROI story is worth studying. Buying or selling a car in India takes days or weeks across calls, document checks, and follow-ups. Cars24 built agents that ask about budget, family size, commute needs, and preferred car type, then recommend cars, book test drives, and help with financing. The agents follow up before test drives, check in after visits, and re-engage customers who dropped out after 10 days.
For sellers, agents collect vehicle details, schedule inspections, send reminders, and gather competitive insights if a customer sold elsewhere. The key lesson: AI agents work best for high-touch, multi-step processes where conversation drives conversion and no transaction happens in a single session. If your business fits that pattern (professional services, complex B2B sales, real estate), this is the playbook.
Fireworks hits $17.5B valuation on open-source model demand
Cloud startup Fireworks exceeded $1 billion in annualized revenue (5x last year) and raised $1.5 billion at a $17.5 billion valuation. The company hosts open-source models that developers can fine-tune with proprietary data. Fireworks' pitch: specialized intelligence at 5-10x lower cost than equivalent-quality closed models.
For small businesses, this is evidence that open-source models are getting good enough for production use. Finance executives are directing employees to consider open-source alternatives as frontier model costs climb. Fireworks gives developers access to models from DeepSeek, MiniMax, Z.ai, and OpenAI's open-weight releases. You bring your own data, fine-tune the model, and get state-of-the-art performance for specific tasks without paying for generalized intelligence.
The partnership with Microsoft (announced in March) extends reach: you can access Fireworks models through Microsoft Foundry. If you've been priced out of Claude Opus or GPT-5 for specific workflows, Fireworks is worth evaluating. Coinbase CEO Brian Armstrong wrote in June that the company is adopting cheaper models where it makes sense.
EU orders Google to open Android and Search to rival AI assistants
The European Commission ordered Google to give rival AI assistants and search engines comparable access to Android and Google Search data under the Digital Markets Act. Google has until January 2027 to share search data and July 2027 to implement Android changes.
For businesses, this means Android users could eventually choose ChatGPT, Claude, Perplexity, or other assistants as deeply integrated system assistants instead of Gemini. Rival assistants would get access to voice commands, app interactions, and device hardware. Competing search engines and AI chatbots would gain access to Google Search data, which could boost their ability to compete. If your business builds on Google's ecosystem, watch how this interoperability requirement shapes the landscape. Non-compliance could cost Google up to 10% of annual worldwide turnover (tens of billions of dollars).
Over half of enterprises report AI agent security incidents
VentureBeat research found that 54% of enterprises have experienced a confirmed AI agent security incident (18%) or a near-miss caught before harm (36%). Only 32% give every agent its own scoped, managed identity. Most agents share credentials or run on shared API keys and human or service-account credentials. Only 30% isolate their highest-risk agents in sandboxes.
If you're deploying AI agents, this is a warning. Shared credentials mean a single compromised or over-permissioned agent carries a wide blast radius. The security stack is overwhelmingly provider-native (OpenAI's guardrails at 51%, Google and Microsoft cloud controls), and satisfaction is high (4.2 out of 5), yet 64% plan to switch or add infrastructure providers within 12 months.
The takeaway: enterprises are comfortable inside the security gap. They're buying provider-native controls while preparing to replace them. If you're running AI agents in production, give each agent its own scoped identity, isolate high-risk agents in sandboxes, and audit credential sharing. The structural weakness is identity, not tooling.
Half of enterprises shipped agents that passed evals and then failed customers
VentureBeat found that 50% of enterprises deployed an agent or LLM feature that passed internal evaluations and then caused a customer-facing failure in the past year. A quarter saw it happen more than once. Only 5% fully trust automated evaluation today. The most-cited limitation: evaluations align poorly with real-world outcomes (29%).
Two-thirds of organizations already permit or are engineering toward fully automated, zero-human-in-the-loop deployment for low-risk agents. The evaluation stack is fragmented: model providers' native evals are tied with having no dedicated tooling at all (17% each). Only about a quarter run real-time quality checks on live production traffic.
If you're shipping AI agents, don't trust passing evals alone. Test against real customer interactions, monitor production traffic in real-time, and keep a human in the loop for high-stakes workflows. The autonomy is arriving faster than the assurance.
16 AI prompt templates for better agent outputs
Zapier published 16 prompt templates designed to get usable outputs on the first try. Three worth trying immediately:
Context-setting template: Front-load background, audience, constraints, and success criteria before your request. This gets you an answer to your version of the request, not the average version from the model's training data.
Standing instructions template: Move your preferences (tone, format, length, always/never rules) into custom instructions or system prompts so the model sees them on every run. This eliminates inconsistency from re-explaining preferences slightly differently each time.
Placeholder template: Provide a template with CAPITALIZED PLACEHOLDERS for content. The model fits its response into the placeholders and preserves formatting. Use this for business communications, recurring reports, and any agent whose output needs to follow the same format every run.
Copy, paste, and customize these templates to reduce the trial-and-error tax on AI prompting.
Three patterns across today's stories: competitive pressure from Chinese AI labs is driving down frontier model costs, enterprises are deploying AI agents faster than they can secure or evaluate them, and agentic commerce infrastructure is maturing (1Password for Claude, DoorDash CLI). Watch how the Moonshot K3 release shapes pricing at the Sonnet tier and whether U.S. labs respond with price cuts or capability jumps. If you're running AI agents in production, audit credential sharing and eval trust before scaling further. The security and evaluation gaps are structural, not temporary.