Key Signals

• Claude Sonnet 5 launches as the new agentic price/performance baseline, simultaneously becoming the default model across Claude Code, GitHub Copilot, and OpenCode. Priced at a promotional $2/$10 per million input/output tokens through August 31 (then $3/$15), Sonnet 5 scores 63.2% on agentic coding benchmarks versus Opus 4.8's 69.2% — close enough to displace larger models for most daily coding workflows. Anthropic increased rate limits for Claude Code users to accommodate higher token usage at elevated effort levels. The coordinated rollout across every major AI coding tool signals that agentic capability at the Sonnet price tier is now table stakes. [1][2][3]

• GitHub Copilot CLI ships two major releases (v1.0.66 and v1.0.67) in a single day, delivering over 90 changes including self-paced /pr auto loops that drive pull requests to green autonomously. The /pr auto command now fixes one thing per CI run and paces itself around CI checks; /pr automerge continues until the PR merges. Other highlights include configurable subagent concurrency/depth limits, the new /chronicle skills review command, desktop notifications, and the compact "highlight reel" timeline view for all users. The release also adds Claude Opus 4.8 Fast as a model option (deprecating 4.6 Fast) and Claude Sonnet 5 support. [4][5]

• GitHub Copilot Agent becomes a first-class option in JetBrains AI Assistant, deepening the IDE integration between GitHub and JetBrains through the Agent Client Protocol. Developers can now select Copilot in the AI Assistant agent picker, choose supported models, tune reasoning depth, and hand off multistep coding tasks — all without leaving their JetBrains IDE. Upcoming features include Next Edit Suggestions, reusable Skills, and deeper orchestration for complex development tasks. [6]

• The MCP ecosystem accelerates significantly: X launches a hosted MCP server, AWS WorkSpaces for Agents uses MCP for connectivity, and Cursor enables centralized Team MCP distribution. X's read-only hosted MCP server lets AI tools like Claude and Cursor access platform data without developers building custom integrations — joining GitHub, Slack, Notion, Stripe, and Salesforce in the growing roster of official MCP providers. AWS WorkSpaces for Agents, now generally available, uses MCP as the primary mechanism for agents to interact with cloud desktops. Cursor's new Team Marketplaces feature allows admins to configure MCP servers once and distribute them across cloud agents, IDE, and CLI. [7][8][9]

• Copilot SDK v1.0.5-preview.1 introduces MCP OAuth host token callbacks and Java annotation-based tool definitions, simplifying enterprise integrations and reducing boilerplate. The new onMcpAuthRequest handler allows host applications to supply OAuth tokens when connecting to protected MCP servers — critical for enterprise environments with complex auth flows. The Java SDK's new @CopilotTool annotation with compile-time metadata generation via JSR-269 eliminates manual JSON schema construction entirely. [10][11]

• AWS makes WorkSpaces for Agents generally available, giving AI agents their own persistent cloud desktops with Active Directory identity, human-in-the-loop controls, and combined MCP + computer vision interaction. Agents can now operate legacy desktop applications without custom integrations, with MCP tool forwarding handling most tasks efficiently and vision-based interaction as fallback. AWS positions this as a way for enterprises to automate workflows in applications that lack APIs, with full CloudTrail auditability and IAM access management. [8]

AI Coding News

• Anthropic positions Claude Sonnet 5 as confirmation that agentic capability is the new baseline expectation at every price tier, with the differentiator shifting from capability to cost and reliability. The model plans, uses browsers and terminals, and runs autonomously at a level that previously required larger, more expensive models. On benchmarks, it slightly outperforms Opus 4.8 on knowledge work tasks despite being significantly cheaper. Safety improvements include lower hallucination rates, better prompt-injection resistance, and reduced sycophantic behavior compared to Sonnet 4.6, though the model cannot develop full working exploits and poses low cybersecurity risk according to Anthropic's testing against Firefox 147. [1][2]

• X launches a hosted MCP server making its real-time data accessible to AI tools like Claude, Cursor, and Grok Build without developers building custom integrations. The server uses OAuth with users' own account permissions and is explicitly read-only — no Write API access — addressing spam concerns while positioning X as a data network for AI applications. The pricing model discourages automated posting with $0.015 per post and $0.20 per link via API, designed to curb misuse. [7]

• OpenClaw, the open-source AI agent that went viral earlier in 2026, is now available on iOS and Android with gateway-based agent routing. Users pair their phone with the OpenClaw Gateway to run AI agents from mobile, covering use cases from coding to task automation. OpenClaw's creator Peter Steinberger joined OpenAI in February following the viral MoltBook marketing stunt, and the mobile launch reflects the broader trend of agentic AI expanding beyond desktops. [12]

• Microsoft launches Copilot Autofix for GitHub Advanced Security in Azure DevOps, extending AI-powered vulnerability remediation to Azure Repos teams. CodeQL identifies vulnerabilities, and Copilot's coding agent generates context-aware fixes as pull requests for human review — addressing what Microsoft calls the "last mile" bottleneck between detection and remediation. The system may include coordinated changes across multiple files, but developers remain responsible for validating every proposed fix. [13]

• A blog post investigating Claude Code's steganographic request marking sparked discussion about transparency in AI coding tools. The article examines how Claude Code embeds hidden metadata or identifiers within prompts sent to the Anthropic API, raising questions about developer awareness and privacy implications. The discussion highlights broader concerns about AI tools fingerprinting requests without explicit user consent. [14]

Feature Update

• Claude Code v2.1.197 makes Claude Sonnet 5 the default model with a native 1M-token context window and promotional pricing of $2/$10 per million tokens through August 31. Users should update to this version for access to the new model. The combination of the expanded context window and lower pricing makes longer, more complex agentic coding sessions economically viable. [3]

• GitHub Copilot adds Claude Sonnet 5 as a generally available model across all paid plans and every surface including VS Code, JetBrains, Copilot CLI, cloud agent, Xcode, Eclipse, and GitHub Mobile. Internal testing showed particularly strong performance on CLI-style tasks, excellent prompt-cache utilization, and competitive latency at lower effort levels. The model is billed at provider list pricing under Usage Based Billing and operates under Zero Data Retention. [15]

• GitHub Copilot CLI v1.0.66 delivers a massive update with 90+ changes: Claude Opus 4.8 Fast model support, compact timeline view, /pr auto self-paced loops, configurable subagent limits, and extensive MCP improvements. Notable additions include /chronicle skills review for managing draft skill changes, desktop notifications for attention prompts, embeddings-based skill retrieval via --dynamic-retrieval, custom agents with reasoning effort settings, and /worktree improvements that use the active model to name branches from context. MCP OAuth token recovery now works automatically after mid-session expiry. [4]

• GitHub Copilot CLI v1.0.67 adds Claude Sonnet 5 as a supported model along with immediate sandbox disabling, subagent parent tool restriction inheritance, and a 30 AI credit minimum for session limits. Tool calls now continue when hooks time out rather than blocking, and MCP OAuth against Microsoft Entra servers behind tenant vanity domains is fixed for AADSTS9010010/AADSTS90023 errors. [5]

• Cursor expands Team Marketplaces to support Team MCPs and organization-level access groups, enabling centralized MCP server distribution across all agent surfaces. Admins configure Team MCP servers once via Dashboard → Integrations & MCP, then members install approved integrations locally without manual server configuration. Marketplace access can now be restricted to specific organization groups in addition to team-level SCIM directory groups. [9]

• OpenCode v1.17.12 enables adaptive thinking for Claude Sonnet 5, adds a yolo mode for auto-approving permissions, and improves MCP OAuth with reconnect, refresh-token scope, and error display. The SDK gets a live event subscription stream, session access APIs, paged durable session history, and permission request endpoints. Desktop improvements include MCP resource autocomplete and stored token/cost totals in session context. [16]

• Copilot SDK v1.0.5-preview.1 introduces MCP OAuth host token callbacks, sessionId in BYOK bearer-token providers, and the Java @CopilotTool annotation-based tool definition system. The onMcpAuthRequest handler lets host applications supply tokens or cancel OAuth requests when sessions connect to protected MCP servers. The BYOK callback field is renamed from getBearerToken to bearerTokenProvider across all SDKs, now receiving sessionId for per-session token scoping. The Rust SDK adds 9 GitHub-anchored attachment variants. [10]

• GitHub Copilot SDK for Java 1.0.5 ships the annotation-based tool API using @CopilotTool and @CopilotToolParam annotations with a JSR-269 compile-time annotation processor — no runtime reflection required. ToolDefinition.fromObject() registers all annotated methods at once, and @CopilotTool methods can inject a hidden ToolInvocation parameter for runtime context like sessionId without it appearing in the LLM-visible JSON schema. [11]

• GitHub introduces per-user AI credit budgets for cost centers, letting enterprise admins set one budget that applies to every individual in the cost center without configuring thousands of individual limits. The budget counts all AI credits and can stop a user before the pool is exhausted. Precedence: individual user-level > cost center user-level > universal. Currently API-only with billing UI support coming soon. [17]