Anthropic Releases Claude Sonnet 5.5 and Advances Opus 5.5

Anthropic Releases Claude Sonnet 5.5 and Advances Opus 5.5

While OpenAI hit pause, Anthropic kept its foot on the gas. Claude Sonnet 5.5 shipped September 28, a week after Opus 5.5, and it's the kind of release that actually moves the needle on cost-per-output: $2/$10 per million tokens, over 30% faster than its predecessor, and scoring 70.6% on Terminal-Bench 4.0 — beating the flagship Opus 5.5's 66.4% [1].

Opus 5.5 itself dropped September 22 at $4/$20 pricing, 20-40% cheaper to run than prior versions, with meaningful gains in coding and knowledge-work safety [2][3]. That a mid-tier model is now outperforming the flagship on real-world coding benchmarks says a lot about where the gains are actually coming from — not raw scale, but efficiency and post-training refinement.

The irony wasn't lost on X: Dario Amodei has been one of the loudest voices calling for a development slowdown, yet Anthropic just shipped two frontier models in a week. Builders don't seem to mind the contradiction — they're too busy migrating workloads to the cheaper, faster Sonnet.

NVIDIA Launches Open Agent Safety Platform with 100+ Partners

NVIDIA used the same week to address the elephant in the room: agents misbehaving in production. The new Open Agent Safety Platform pairs an open-source sandboxing runtime (OpenShell) with Sentry, a hardware watchdog running on BlueField-4 DPUs that can quarantine a rogue agent in milliseconds [1][2].

More than 100 partners signed on at launch, including Anthropic, Microsoft, Perplexity, Salesforce, and SpaceXAI — a rare moment of industry-wide agreement that agent safety needs to be infrastructure, not an afterthought. Jensen Huang's framing was blunt: "Safety and security require full-stack engineering" [3]. Translation: you can't patch this at the model layer alone: it has to be baked into silicon.

The timing next to the Astra cancellation is not a coincidence. The industry is quietly admitting that "trust the model" isn't a strategy — you need a kill switch that works even when the model itself is lying about what it's doing.

SUSE Integrates NVIDIA Open Agent Safety Platform for Enterprise AI Agents

SUSE moved fast to fold NVIDIA's OpenShell and Sentry into its SUSE AI Factory offering, giving enterprises runtime controls, sandboxing, and hardware-enforced policy for agents out of the box [1][2]. This is the practical, unglamorous layer that actually determines whether "agentic AI" is deployable in a regulated European enterprise or just a demo.

It's a small story on its own, but it's a useful signal for the Nordic/EU market specifically: sovereignty and governance conversations happening at events like Bits & Pretzels in Munich are no longer purely theoretical. Vendors are shipping the actual guardrails European compliance teams have been asking for.