Key Signals

• GitHub Copilot received its largest single-day platform expansion, adding one-million-token context windows, configurable reasoning levels, an Agent tasks REST API, and one-click Actions failure fixing across Pro, Pro+, and Max tiers. The context window increase lets developers work across larger codebases without losing coherence, while the reasoning dial unlocks extended thinking for hard architectural challenges. The REST API is particularly significant: it enables programmatic orchestration of Copilot cloud agent tasks, opening the door to fan-out migrations, automated repo scaffolding, and release pipelines driven by a simple script. Combined with the "Fix with Copilot" button on failing Actions runs, this positions Copilot as a continuously operating background agent rather than just an in-editor assistant. [1][2][3]

• The AI coding industry is rapidly shifting from flat-rate subscriptions to token-based billing, with Cursor restructuring its pricing and introducing enterprise governance controls in the same week GitHub retired Copilot's fixed subscription model. Cursor cut Teams seats by 20% to $32/month while adding a $120 Premium tier, introduced ringfenced usage pools for its cheaper first-party Composer model, and launched an org-level dashboard with per-department budgets, model access controls, and spend alerts. The Linux Foundation also announced the Tokenomics Foundation backed by Google, Microsoft, and others to build open standards for AI token economics. This convergence signals that unpredictable inference costs have become a top-of-mind concern for enterprise CFOs across the industry. [4]

• OpenAI Codex 0.137.0 shipped enterprise credit limits, remote-control client pairing, and multi-agent v2 with per-thread runtime selection. The enterprise features indicate Codex is maturing toward production enterprise deployment. Multi-agent v2 now keeps runtime choice per thread and exposes cleaner follow-up semantics for spawned agents, enabling more sophisticated orchestration of parallel coding tasks. Hosted web and image tools are also now available in more code-mode flows with parallel standalone web searches. [5]

• Kiro v2.6.0 and Claude Code v2.1.163 both advanced their CLI agent ergonomics with session portability and developer workflow improvements. Kiro added transcript export, auto-derived terminal window titles, a --effort launch flag, and persistent model/effort preferences. Claude Code shipped managed version range settings for enterprise deployments, a /plugin list command, clipboard copy from /btw, and critical fixes for Bedrock/Vertex/Foundry auth and $TMPDIR handling. Both releases show CLI coding agents converging on similar UX patterns around session management and effort configuration. [6][7]

• Snowflake launched CoCo, a model-agnostic enterprise coding agent deeply integrated with governed data infrastructure, signaling that data platform vendors are entering the AI coding space. CoCo features a native desktop IDE, mobile app, Slack integration, agent SDK, and cloud agents for autonomous task execution—all built on Snowflake's permissioning and governance layer. The product is model-agnostic and works across Snowflake, Postgres, AWS Glue, dbt, Airflow, and Databricks data sources. This represents a distinct "governed data coding" niche that pure-play AI coding tools don't currently serve. [8]

AI Coding News

• The era of flat-rate AI coding tool pricing is ending as GitHub, Cursor, and the Linux Foundation's new Tokenomics Foundation all converge on token-based economics within the same week. GitHub retired Copilot's fixed subscription model in favor of consumption-based billing, prompting backlash from users who saw projected bills jump tenfold. Cursor responded with restructured pricing, ringfenced pools for its cheaper Composer 2.5 model ($0.50/M input vs Claude Opus 4.8's $5.00/M), and enterprise spend dashboards. JetBrains also open-sourced Mellum2, a 12B-parameter model targeting the infrastructure layer of agentic systems for on-premises deployment—an alternative approach that puts inference costs entirely in the team's hands. [4]

• Snowflake rebranded Cortex Code as CoCo and positioned it as the enterprise coding agent for governed data environments, with a desktop IDE, agent SDK, and partnerships with Vercel, Superblocks, and Retool. The announcement at Snowflake Summit positions CoCo against general-purpose coding tools by emphasizing deep integration with data governance, permissioning, and cross-platform data access. CoCo also introduced Snowflake Datastream, a Kafka-compatible streaming service enabling real-time AI app development. [8]

• Anthropic's annualized revenue crossed $47 billion in May 2026, up from roughly $9 billion at the end of 2025, as the company prepares for its IPO. President Daniela Amodei addressed skeptics questioning AI's returns, noting the company's breakneck growth trajectory. This 5x revenue growth in under six months reflects the explosive demand for Claude-family models powering tools like Claude Code and third-party integrations. [9]

• OpenAI published a case study on enterprise services company Endava using AI agents, ChatGPT Enterprise, and Codex to redesign their entire software delivery pipeline. The article describes how Endava is building an "agentic organization" that uses AI agents throughout the development lifecycle—from requirements analysis to workflow automation—representing the enterprise adoption pattern OpenAI is promoting for Codex. [10]

Feature Update

• GitHub Copilot now supports one-million-token context windows and configurable reasoning levels in VS Code, Copilot CLI, and the GitHub Copilot app. Users can work across larger codebases, longer documents, and complex multi-file projects without losing context. Reasoning levels let developers dial between speed and depth, unlocking extended thinking for architectural and debugging challenges. Higher settings consume more AI credits per interaction. [3]

• GitHub launched the Agent tasks REST API in public preview for Copilot Pro, Pro+, and Max users, enabling programmatic orchestration of Copilot cloud agent tasks. The API supports starting tasks, tracking progress, and authenticating via personal access tokens or OAuth. Example use cases include fanning out refactors across repositories, scaffolding new repos from internal portals, and automating weekly release preparation. [2]

• GitHub Copilot's "Fix with Copilot" button is now available on failing Actions workflow runs for Pro, Pro+, and Max subscribers. Clicking the button triggers Copilot cloud agent to investigate the failure from its own cloud-based dev environment, push a fix to the branch, and tag the user for review—handling common CI fixes like test failures and linter errors without manual intervention. [1]

• GitHub Copilot in Visual Studio received its May 2026 update featuring a new Plan agent, Skills panel, multi-file summary diff, and context window usage management. The Plan agent explores the codebase with read-only tools, asks clarifying questions, and drafts implementation plans as markdown files before handing off to Agent mode. The Skills panel provides a single view of all discovered agent skills, while the summary diff allows accepting or undoing Copilot changes at the all-files, per-file, or per-chunk level. [11]

• Copilot Chat on github.com now provides richer pull request context and is generally available to all Copilot license holders. The feature enables side-by-side code and chat views, adds PR understanding and summary abilities for faster context injection, and is accessible via "Ask about this diff" buttons, the Copilot nav button, or code highlighting. [12]

• Cursor released Canvas Design Mode, allowing users to select and annotate UI elements directly in canvases to guide edits visually. Instead of describing changes in text, developers can point to elements, provide feedback, and iterate more quickly. The release also adds a context usage report as an interactive canvas, breaking down token allocation across system prompts, tool definitions, rules, and skills with a "Debug with Agent" action. [13]

• Claude Code v2.1.163 shipped managed version range enforcement, plugin management, and critical platform fixes. New requiredMinimumVersion/requiredMaximumVersion managed settings let enterprises lock Claude Code to approved versions. The /plugin list command supports --enabled/--disabled filters, and Stop/SubagentStop hooks can now return additionalContext to keep turns going. Key fixes address claude -p hanging with backgrounded commands, Bedrock/Vertex/Foundry auth when CI=true, $TMPDIR regression from v2.1.154, and Windows EEXIST errors with OneDrive. [7]

• Copilot CLI v1.0.60-0 added vim-style navigation in /diff view, git worktree creation from the pull requests screen, and LSP server configuration improvements. New features include a billing help topic, Mission Control sharing status in session info, -r shorthand for --resume, and rewind picker with diff stats. Security fixes block web_fetch from loopback/private/metadata addresses, prompt approval for LD_PRELOAD-style commands, and properly enforce MCP allowlists for npm scoped servers and Azure API Center registrations. [14]

• OpenAI Codex v0.137.0 introduced enterprise monthly credit limits, cloud-managed config bundles, remote-control client pairing via app-server v2 RPCs, and multi-agent v2 per-thread runtime selection. Plugin workflows gained codex plugin list --json output and cached remote catalog suggestions. Hosted web and image tools expanded to more code-mode flows with parallel web searches. Platform reliability improved for macOS app launches, Windows SQLite startup, and thread resume. The Python SDK now publishes runtime wheels pinned to glibc-compatible packages. [5]

• Gemini CLI v0.45.1 shipped a patch release, while the v0.47.0 nightly updated auto mode to use Gemini 3.5 Flash when the experiment flag is enabled. The v0.45.1 patch cherry-picks a fix to the stable release. The nightly also adds EBUSY fallback and TOML parse recovery for policy handling, optimized PR size labeling CI, and pull_request_target triggers for write access on fork PRs. [15]

• Kiro v2.6.0 added transcript export, terminal window titles, effort-at-launch flags, and persistent model preferences. The /transcript save command exports conversations as markdown, plaintext, or JSON for sharing or archiving. Terminal window titles auto-derive from session topics. The --effort flag for kiro-cli chat sets initial reasoning depth, and /model and /effort choices now persist automatically across sessions without requiring explicit default-setting commands. [6]