New guardrails for AI app spending
Your AI bill should not be a mystery novel. Today’s updates give you cheaper model runs, clearer cost controls, safer code scans, and more ways to connect agents to the tools your team already uses.
In today’s edition
- OpenAI added API-key spend tracking and hard monthly limits alongside a temporary GPT-5.6 Sol price cut.
- Claude Security now uses Mythos 5 to scan GitHub repositories and send suggested patches into Claude Code.
- v0 apps and agents can connect securely to more than 100 services through Vercel Connect.
- Replit added scheduled Routines, persistent Memories, expanded Free Mode, and external-style pen testing.
- DeepSeek V4 Flash Vision Exp is available on OpenRouter with image input at V4 Flash pricing.
OpenAI adds spend limits by API key
OpenAI now lets you track usage and spend by API key in its Usage and Spend dashboards, so you can see which apps and workloads are driving your bill.
- Set monthly spend limits at the organization or project level.
- Use hard limits to stop traffic once a budget is reached.
- The controls are available in the API Platform and through the Admin API for programmatic workflows.
- GPT-5.6 Sol API and credit pricing is down by more than 20% for the next three months.
What it means Tag each workload with its own API key, add hard caps before an agent goes live, and use the lower Sol price window to test higher-volume workflows.
Source: OpenAIDevs
Claude Security gets Mythos 5 scans
Claude Security scans now run on Claude Mythos 5 in public beta for all Claude Enterprise customers, with no separate model access required.
- Point it at a GitHub repo to scan for vulnerabilities across files and component interactions.
- Findings include a CWE category, confidence rating, severity rating, and suggested fix.
- Suggested patches can open in Claude Code on the web using the models your team already uses.
- Mythos runs behind the scan and returns findings without giving users direct model access.
What it means Put repository scanning into your release checklist, then route high-confidence findings into Claude Code so developers can review and patch issues without changing tools.
Source: claudeai
v0 connects agents to 100+ services
Apps and agents built in v0 can now securely connect to Slack, GitHub, Salesforce, and more than 100 other services through Vercel Connect.
- Vercel Connect handles authentication for the connected services.
- Teams can reuse connectors instead of rebuilding auth for each app.
- Connections use short-lived tokens.
- The update applies to both apps and agents built in v0.
What it means Prototype internal agents that can read and act across your company stack, while leaning on reusable connectors instead of spending your first sprint on OAuth plumbing.
Source: v0
Replit turns chats into scheduled work
Replit shipped a broad Agent update: more capacity in Free Mode, Conversations for starting ideas in chat, scheduled Routines, Memories, and black-box pen testing.
- Core subscribers can create up to 30X more in Free Mode, where everyday tasks no longer use credits.
- Routines, now in beta, schedule a task from a conversation and return the result in the same thread.
- A routine can produce a morning summary from Slack, inbox, and calendar data.
- Level 3 scans in the Security Center test an app like an external attacker would.
What it means Turn recurring research and operations work into scheduled Routines, and use the external-style scan before sharing a new app beyond your team.
Source: Replit
DeepSeek Flash adds image input
DeepSeek V4 Flash Vision Exp is live on OpenRouter, bringing image input to the V4 Flash price tier.
- The model supports image input alongside its text capabilities.
- OpenRouter says it matches V4 Flash 0731 on text agent benchmarks.
- OpenRouter says it beats Opus-4.8 on Agents' Last Exam and ZeroBench.
- It is available through OpenRouter now.
What it means Add image understanding to an existing Flash-priced agent workflow, then test it on screenshots, documents, and visual QA tasks before paying for a higher-cost model.
Source: OpenRouter