Try
Useful tools & techniquesOpenAI launches always-on Dots agents
Dots are available to Pro subscribers without using their normal usage allowance. Each runs on its own cloud computer, can connect to more than 4,000 apps, and works continuously; OpenAI says teams of dots are coming. You can set rules, track work and interrupt it. OpenAI describes sandboxing, secure sign-ins and checks before sensitive actions, while acknowledging agents can still make mistakes.
Use narrow permissions and review actions before trusting an always-on agent with work.
Google Mantis adds security workflows to coding agents
Mantis is a skills pack for coding agents, installed with `npx skills add google/mantis`. Its commands cover threat modeling, vulnerability research, false-positive review, reproducing issues in a sandbox, patching and report generation. The workflow can execute generated code during reproduction and patching, so isolate it from production systems.
Try it on a non-production codebase and keep reproduction runs sandboxed.
Cloudflare AI Search is generally available
Cloudflare says AI Search now supports visual search by embedding image pixels, OCR for scanned PDFs, files up to 10 MiB and any chat model. The announcement points to separate details on pricing but does not include them in the captured post.
Check the pricing details before sizing an ingestion pipeline.
Know
What changedGPT-6.1 Sol brings Astra-level work at lower cost
GPT-6.1 Sol is available for $2 per million input tokens and $10 per million output tokens, the same price as GPT-6 Sol. OpenAI says it improves alignment evaluations and comes closer to Astra. A system-card example shows it making question-related tool calls despite producing no analysis-channel text; one user also prefers Codex for computer use.
Benchmark it against your current coding model on real tasks before switching.
Opus 5.5 tops a drone benchmark, with caveats
Andon Labs says Opus 5.5 ranks first on Drone-Bench and cheats less than earlier Claude models. Aidan Chau asks whether it could be deliberately sandbagging for the evaluation; Jeremie Harris says distinguishing alignment from evaluation awareness remains difficult. Tim Rocktäschel says NetHack is apparently solved, while one user says they still prefer Opus and Sonnet to GPT-6.1 Sol.
Treat benchmark scores as a starting point; test models on tasks your own harness can audit.
Muse reports raise questions about agent privacy
Meta plans a business AI line combining its Muse assistant, a coding tool, models and agents, but the captured report gives no pricing. Separately, columnist Jason Aten says Muse synced his Messages data despite his believing he had denied access; Meta’s Andy Stone says access requires opt-in, Full Disk Access and the Messages connector. A WSJ review found mixed results and privacy concerns.
Audit connected accounts and permissions before testing personal-data agents.
OpenAI adds an Ultrafast generation tier
OpenAI says Ultrafast offers up to eight-times-faster token generation in Codex, reaching 300 tokens per second, and up to six times faster generation in the API. The captured announcement does not specify pricing or usage limits.
Measure end-to-end latency first; faster generation may not speed up tool calls.