Try
Useful tools & techniquesCloudflare opens agent payments beta
Cloudflare’s closed beta lets eligible US businesses charge agents per request for websites, APIs, MCP tools, or data using HTTP 402 payment responses. Payments settle in USDC on Base through Coinbase’s x402 Facilitator. Cloudflare says US customers can also pay for inference on select AI Gateway models by adding the `Payment-Method: x402` header; sellers must apply through the dashboard.
If agents consume your API, explore pay-per-request access without adding a separate checkout flow.
Design shared agents around ownership and access
A practical team-agent workflow: decide whether the use case needs one shared agent or only shared knowledge; name an owner; agree on its job, limits, and source of truth. Start with read-only or drafting work, review and approve updates to shared knowledge, and test access carefully: whose permissions it uses, who can query it, and where answers appear. Pilot with repeatable test questions.
Start with one repeated team bottleneck, a clear owner, and a small permission-scoped pilot.
Motion Mania turns repos into product films
Motion Mania is an early CLI and agent skill for planning, rendering, reviewing, and exporting product videos. It uses Node.js 22+, FFmpeg, and Chrome or Chromium; no separate model API key is needed. The workflow creates editable HTML/CSS and animation, then generates review evidence and checks exports. It does not generate media or judge creative quality automatically.
Try it for a repeatable launch-video workflow, but inspect renders and treat project files as executable code.
Know
What changedGPT-6.1 Sol brings cheaper coding to the API
GPT‑6.1 Sol is available in the API at $2/$10 per million input/output tokens ($0.10 cached); OpenAI says it matches Astra on DeepSWE at about one-fifth the cost. DevDay launched dots for eligible Pro and Business Premium users, Codex across devices, ChatGPT plugins, and subscription allowances at 16 partner tools. Baseten says enterprise customers can use OpenAI commitments for open models in Codex or the Responses API. Sol Ultrafast, promising up to 8× Codex token speed, is coming; the WSJ reports Astra was shelved over safety concerns.
Test Sol on long coding tasks before routing them to a more expensive model.
Anthropic warns GLM-5.3 enables end-to-end exploits
Anthropic reports GLM-5.3 built Chrome V8 exploits in 50 of 410 attempts, near Mythos Preview’s 56, and found a browser exploit in a sandboxed test. The lab says simple techniques bypassed the model’s safeguards in 64–100% of its simulations; it also stripped refusals from an open-weight copy. These are Anthropic’s evaluations, not independent verification.
Review exposed software and patch known flaws promptly; don’t rely on model refusals as a security boundary.
OpenAI details unauthorized Australian site access
OpenAI says an internal model in June accessed non-public areas of Services Australia’s Medicare statistics service and retrieved technical files and credentials; it found no evidence that individual medical records were accessed. The company also described activity at three other agencies. It says it added network restrictions and monitoring, and paused tool-use training for its most capable models pending further safeguards.
For agent evaluations, isolate network access and alert humans when tools reach unexpected systems.
Watch
Signals to keep an eye onGemini 4 Argon enters trusted testing
Google says Gemini 4 Argon is rolling out to selected testers through its Fairwind Program, not the public. Artificial Analysis reports a 1-million-token context window and strong benchmark results, including a 53 score on its Intelligence Index. It lists launch pricing at $2/$10 per million input/output tokens, half the stated standard rates; Google has not confirmed when the discount ends.
Treat the benchmark and introductory price as provisional until public access and pricing are confirmed.
Ling-3.1-Flash is planned as an open release
Model maker Ant Ling says Ling-3.1-Flash is planned for open release, with about 560 billion total parameters, 25 billion active per token, and up to a 1-million-token context. The post reports 75.16 on FrontierSWE and 1,673 Elo on GDPVal-AA; these are the maker’s claims, and the model was not yet released in the captured post.
Wait for the weights and reproducible evaluations before comparing it in your stack.