Tue, Oct 13
8 itemsThis week saw a dense cluster of updates to AI coding tools: GitHub Copilot rolled out Claude Haiku 5.5 and pushed forward with local sandboxes and billing controls, Claude Code shipped multiple releases in a row and adjusted Sonnet 5.5 cache pricing, and Meta also launched a new model series aimed at coding and agents.
Claude Haiku 5.5 launches on GitHub Copilot and Claude Code
Anthropic's latest lightweight model, Claude Haiku 5.5, is now generally available in GitHub Copilot and has become the default Haiku model for Claude Code and the Anthropic API, supporting a 1M context and priced at $0.10/$0.50 per million tokens.
Why it mattersA generational swap in lightweight models directly affects the cost and speed of subagents, quick edits, and terminal tasks, and developers need to assess the billing changes that come with the default model switch.
Anthropic halves Sonnet 5.5 cache read pricing
Anthropic cut the cache read price for Sonnet 5.5 in half, but its own tools still showed the old price for two days.
Why it mattersCache reads are a major component of high-frequency call costs, and both the price cut and the display delay affect teams' cost accounting and budget decisions.
GitHub Copilot local sandbox is now generally available
GitHub Copilot's local sandbox feature is now generally available in Copilot CLI, the Copilot app, and VS Code sessions using Agent Host, giving developers a safer execution environment.
Why it mattersSandboxes make the permission boundaries for agents executing code clearer, and are a key prerequisite for enterprises evaluating how far to open up agent automation.
Copilot code review adds org billing and telemetry attribution controls
GitHub added organization billing and license controls for Copilot code review administrators, and now lets enterprise administrators attribute Copilot activity to users via telemetry.capture.identity without collecting prompts and response content.
Why it mattersBilling attribution and telemetry granularity directly affect enterprise procurement, compliance audits, and internal cost allocation.
Copilot for JetBrains adds default model and MCP controls
GitHub Copilot for JetBrains was updated with more control over the default model and MCP servers, plus improvements to chat navigation, account controls, and diagnostics.
Why it mattersJetBrains users can manage models and tool integrations more precisely, which affects configuration strategies in their daily development environments.
Claude Code ships multiple releases in a row
Claude Code released v2.1.292 through v2.1.296, adding plugin marketplace installation, an auto-compact window for subagents, hook failure blocking, terminal status protocol support, and a fix for instruction-type hooks being incorrectly allowed through.
Why it mattersHook and subagent behavior is now more controllable, and teams using automated workflows need to watch for policy and script compatibility after upgrading.
Claude Code isolates policy variables and supports shareable Artifacts
Claude Code isolated nine policy variables from installed plugins and introduced a shareable, real-time Artifacts feature.
Why it mattersPolicy isolation reduces the risk of plugins overstepping their permissions, while shareable artifacts change how teams collaborate and deliver results.
Meta launches Muse Wonder series to take on Anthropic and OpenAI
Meta introduced the Muse Wonder series of AI models aimed at coding and agent work, seen by observers as a challenge to Anthropic and OpenAI.
Why it mattersThe coding model race has gained another major competitor, giving developers one more option to evaluate when choosing tools and models in the future.