Google shipped Gemini 3.6 Flash yesterday at $1.50/M input and $7.50/M output tokens, down from the previous Flash generation’s $9 output price. The bigger number is buried under the pricing: it produces the same output using roughly 17% fewer tokens, and that difference compounds fast in agentic workloads where a single task chains dozens of tool calls. Fewer tokens per step means real latency and cost savings that a benchmark score won’t show you.
If you’re routing production traffic across model tiers, this is the one to re-test against whatever you’re currently using for high-volume, low-complexity calls (classification, extraction, routing decisions). Knowledge cutoff also jumped from January 2025 to March 2026, which matters if you’ve been hitting stale-answer issues on anything referencing recent libraries or APIs.
What to check this week: re-run your eval suite against 3.6 Flash if you have anything on the old Flash tier. The token efficiency gain alone may justify a swap even before you factor in the price cut.
Supported by Cube
Ask three tools for last quarter's revenue and you'll get three answers — every disconnected dashboard, spreadsheet, and AI bot calculates it its own way. Cube ends the drift. Cube is the agentic analytics platform where metrics are defined once and every surface reads from them: reports, dashboards built by AI or drag-and-drop, analytics embedded in your own product, and AI agents 400+ companies like Wix, Patagonia, Brex, and Webflow build on it.
Patch now: five separate agent-tooling CVEs landed this week
A rough week for anyone running AI coding agents with real filesystem or shell access:
Cursor: workspace-controlled hook config vulnerability (CVE-2026-48124), patched in v3.0.0. If you’re on an older version and use shared/team workspaces, update before you open anything from an untrusted repo.
Codex CLI v0.95.0: allowlist bypass, high-severity bounty paid out. Affects anyone relying on the CLI’s permission allowlisting to restrict what the agent can execute.
Gemini CLI and Cursor: both hit by a shared Docker socket flaw. If either tool is running inside a container with the Docker socket mounted, treat that as a live escalation path until you’ve confirmed a patch.
Antigravity: macOS Seatbelt denylist bypasses. Sandbox restrictions on macOS were not holding as expected.
ServiceNow AI Platform: a pre-auth RCE (CVE-2026-6875) is being actively exploited in the wild, days after the July 14 patch. Self-hosted instances are the exposure; hosted instances were patched by ServiceNow directly. If you’re self-hosting, patch today, not this sprint.
The pattern across all five: agent tooling with filesystem, shell, or container access is getting real security scrutiny for the first time, and it’s turning up real bugs. Worth an afternoon auditing what permissions your team’s agent setups actually have versus what you assumed they had.
Quick hits
GitHub Copilot shipped its first open-weight coding model, available now for enterprise admins to enable under usage-based billing.
Claude Cowork added “Record a Skill.” Screen-record a task, narrate it, and it turns into a reusable skill without hand-writing agent instructions. It’s a direct answer to Codex’s workflow-capture feature.
AMD and Anthropic announced a $5B equity investment and 2GW MI450 GPU deployment, with Anthropic using Claude to help accelerate ROCm development. Relevant if you’re evaluating non-Nvidia inference infrastructure over the next 12 months; first gigawatt online H1 2027.
One link worth keeping
Terence Tao published a walkthrough of an AI-assisted counterexample to a longstanding conjecture this week, including the actual chat log where he used a model to check his calculations. Regardless of where you land on AI-assisted math, it’s a rare, concrete look at how a top mathematician actually uses these tools day to day.
That’s it for today.