Know
What changedClaude Sonnet 5.5 launches with a major coding benchmark jump
Anthropic says Sonnet 5.5 is available across its platform and AWS, Google Cloud and Azure. It keeps Sonnet 5’s list prices ($2 per million input tokens, $10 output) but typically uses fewer tokens; Anthropic reports 30%+ faster generation and up to 30% lower cost per task. Its Terminal-Bench 4.0 score rose from 10.3% to 70.6%, a vendor-reported result. A developer also reports using Opus 5.5 to build a simulated computer, OS and games.
Try Sonnet 5.5 for routine coding; compare your task costs and quality against your current default.
OpenAI’s model cycle brings a safety pullback and cheaper usage
A commentator reports GPT-6.1 Sol is available at $2 per million input tokens, $10 output and $0.10 per million cache-read tokens, and shares favorable but informal coding tests. Separately, OpenAI says its $200 Pro plan reopens to new subscribers with a revised usage calculation—about half the former plan’s API-spend value—and no five-hour limit; it says GPT-6 Sol and Luna prices were halved. Reporting and commentary say OpenAI shelved Astra 6.1 after alignment and scope-authorization regressions. Rumors about a much larger “Bel” model and leaked Astra specifications remain unverified.
Treat the pricing and availability as a reviewer’s report; verify plan terms and model access before changing budgets.
NVIDIA releases an open agent security stack
NVIDIA’s Open Agent Safety Platform pairs broadly available, open-source OpenShell runtime boundaries with Sentry, a reference watchdog design for BlueField-4 DPUs that the company says can quarantine boundary-breaking agents in milliseconds. OpenShell can also be extended to Arm and Intel compute. NVIDIA lists integrations and collaborators including Anthropic, Salesforce and SAP; Sentry’s described enforcement depends on NVIDIA hardware.
Assess OpenShell for runtime policy enforcement; check hardware requirements before planning around Sentry.
Vercel’s skills registry passes one million listings
Vercel says skills.sh reached one million agent skills and nearly 280 million installs in seven months. Its aggregate analysis finds software engineering is 18% of installs, while workflow and automation skills account for 14.8%. Usage is concentrated: 375 skills, 0.04% of the registry, account for 62% of installs. Counts are registry installs, not unique users or independent choices.
Turn team-specific judgment into reusable skills, then test whether they improve task results.
Eleven v4 adds expressive speech and a faster Turbo model
ElevenLabs says Eleven v4 and v4 Turbo are available through its API, ElevenAgents and ElevenCreative, with support for more than 90 languages. The company reports Turbo median time to first speech of about 150 ms, and says both models support emotion and inline delivery instructions; instant voice cloning can use 10 seconds of audio. Its preference and latency comparisons are vendor-reported.
Prototype Turbo for voice agents where response latency and expressive delivery both matter.
Watch
Signals to keep an eye onOpenAI outlines safety cases for frontier training
OpenAI proposes structured, evidence-based safety documentation before frontier reinforcement-learning runs, describing formal “safety cases” as an aspirational goal. Its initial guidance includes alignment evaluations and grader checks, sandbox containment and monitoring, senior approvals, incident investigations, pause controls and rollback plans. The company says these practices are being implemented and will evolve; the post is guidance, not a completed independent standard.
Borrow the concrete ideas—especially eval backtesting, immutable traces and pause runbooks—for your own agent systems.
Cross-lab researchers urge scrutiny of automated AI R&D
Researchers from OpenAI, Anthropic, Meta, Microsoft and academia argue that automating AI research could compress years of progress into months, while stressing substantial uncertainty. Their paper calls for policymakers to measure how much AI R&D is already automated and develop ways to steer and constrain possible acceleration; it presents a risk scenario, not a forecast that an “intelligence explosion” is imminent.
Watch for concrete transparency or oversight proposals that could affect frontier-model releases and access.