# Claude Opus 5.5

Claude Code's changelog says sandboxed Bash commands now follow the normal permission flow by default instead of running without prompts; keeping prompt-free execution requires setting sandbox.autoAllowBashIfSandboxed. The same batch makes read-only git commands ask first in repositories that are not the project's or trusted, and unpacks plugin MCP bundles under the plugins directory so read-only plugin folders load. These changes concern the permission boundary for an agent running commands and loading plugins locally, not model capability.

Claude Code's changelog says sandboxed Bash commands now follow the normal permission flow by default instead of running without prompts; keeping prompt-free execution requires setting sandbox.autoAllowBashIfSandboxed. The same batch makes read-only git commands ask first in repositories that are not the project's or trusted, and unpacks plugin MCP bundles under the plugins directory so read-only plugin folders load. These changes concern the permission boundary for an agent running commands and loading plugins locally, not model capability.

Weight-space programming lets a language model read and edit a trained network's weights directly, and had been shown on game-playing networks and language models. Applied to EEG and ECG models, Claude Opus 5.5 wrote all 2,730 weights of TinyEEGNet with no gradient step and only five practice subjects, reaching 0.834 balanced accuracy on 65 unseen subjects, on par with 0.806 for the same architecture trained on 30 subjects and above 0.704 trained on five; it could not infer a network's task from its weights, nor improve an already well-trained network.

Weight-space programming lets a language model read and edit a trained network's weights directly, and had been shown on game-playing networks and language models. Applied to EEG and ECG models, Claude Opus 5.5 wrote all 2,730 weights of TinyEEGNet with no gradient step and only five practice subjects, reaching 0.834 balanced accuracy on 65 unseen subjects, on par with 0.806 for the same architecture trained on 30 subjects and above 0.704 trained on five; it could not infer a network's task from its weights, nor improve an already well-trained network.

Anthropic released message threads in beta on the Claude Platform Messages API, keeping the conversation on the server so later requests send only new messages, which makes requests smaller and lowers time to first token for long conversations and agent loops. The feature is enabled with the message-threads-2026-08-12 beta header, a thread expires 5 minutes after its last request and is deleted no more than 1 hour after its last use, and it is available on the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed), and Claude in Microsoft Foundry.

Anthropic released message threads in beta on the Claude Platform Messages API, keeping the conversation on the server so later requests send only new messages, which makes requests smaller and lowers time to first token for long conversations and agent loops. The feature is enabled with the message-threads-2026-08-12 beta header, a thread expires 5 minutes after its last request and is deleted no more than 1 hour after its last use, and it is available on the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed), and Claude in Microsoft Foundry.

Anthropic integrated Project Glasswing and its Cyber Verification Program into three access tiers that relax model cyber safeguards according to defensive, authorized red-team and specialized testing work. Previously, generally available models blocked most cyber work, and the CVP offered a single reduced-safeguard channel. On CyScenarioBench, with five attempts at each of 10 challenges per tier, no CVP access blocked every task at the first prompt, Defense Access blocked 46 of 50 trials with four successes, and Red Team Access had no blocks and completed 34 of 50, comparable to the 67.6% success rate with no safeguards.

Anthropic integrated Project Glasswing and its Cyber Verification Program into three access tiers that relax model cyber safeguards according to defensive, authorized red-team and specialized testing work. Previously, generally available models blocked most cyber work, and the CVP offered a single reduced-safeguard channel. On CyScenarioBench, with five attempts at each of 10 challenges per tier, no CVP access blocked every task at the first prompt, Defense Access blocked 46 of 50 trials with four successes, and Red Team Access had no blocks and completed 34 of 50, comparable to the 67.6% success rate with no safeguards.

Anthropic now offers an IP allowlist for organizations in the Claude Console, in beta, and requests to covered services from addresses not on the list are refused, which limits where a leaked API key can be used. Admins can add IPv4 and IPv6 addresses or CIDR ranges, up to 1,000 per organization. Previously an organization could not narrow the entry point for requests by source address in the console; this setting shrinks the usable range after a key leak from any address to the listed ones.

Anthropic now offers an IP allowlist for organizations in the Claude Console, in beta, and requests to covered services from addresses not on the list are refused, which limits where a leaked API key can be used. Admins can add IPv4 and IPv6 addresses or CIDR ranges, up to 1,000 per organization. Previously an organization could not narrow the entry point for requests by source address in the console; this setting shrinks the usable range after a key leak from any address to the listed ones.

Anthropic opened a research preview of Claude Duet, a proactive agent app built around Claude Opus 5.5. It carries over memory, skills and connected apps from a Claude account, sends a morning brief across those apps, and keeps working on tasks in a background virtual machine. Before sending a message, making a payment, signing a document or deleting anything, it shows the action and waits for approval; the preview starts with US Pro and Max users via a waitlist.

Anthropic opened a research preview of Claude Duet, a proactive agent app built around Claude Opus 5.5. It carries over memory, skills and connected apps from a Claude account, sends a morning brief across those apps, and keeps working on tasks in a background virtual machine. Before sending a message, making a payment, signing a document or deleting anything, it shows the action and waits for approval; the preview starts with US Pro and Max users via a waitlist.

Anthropic's Claude Platform release notes announce that Claude models are available on Google Cloud in an Anthropic-managed form: the API is still operated by Anthropic but reached through a customer's Google Cloud project, with access controlled by Google Cloud IAM and billing handled by Google Cloud Marketplace. The same notes say fast mode for Claude Opus 5.5 is available in that managed form. For teams already running workloads on Google Cloud, the route puts model calls inside the cloud provider's existing identity and billing systems.

Anthropic's Claude Platform release notes announce that Claude models are available on Google Cloud in an Anthropic-managed form: the API is still operated by Anthropic but reached through a customer's Google Cloud project, with access controlled by Google Cloud IAM and billing handled by Google Cloud Marketplace. The same notes say fast mode for Claude Opus 5.5 is available in that managed form. For teams already running workloads on Google Cloud, the route puts model calls inside the cloud provider's existing identity and billing systems.