Key Signals

• GitHub Copilot's switch to usage-based billing triggers sticker shock across the developer community. As of June 1, all Copilot plans now bill based on AI Credits consumed (1 credit = $0.01), replacing the previous flat request-based model. Early reports show power users burning through their entire monthly allotment in a single day, with model choice dramatically affecting cost — GPT-5.5 frontier output costs 24× more than GPT-5.4 nano. This pricing shift may accelerate fragmentation in the AI coding market as developers explore alternatives with more generous or subsidized limits. [1][2]

• The agentic coding tool market has converged on a shared architectural blueprint after six months of parallel evolution. Claude Code, Cursor, Codex, and Antigravity all now feature terminal surfaces, explicit planning, approval gates, MCP integration, and parallel agent delegation — a pattern so consistent it signals discovery rather than imitation. SWE-bench scores have narrowed to a tight band, effectively demoting the model as the differentiator and shifting competition to workflow, distribution, and pricing layers. The AGENTS.md convention is emerging as a cross-tool standard for repository-level agent configuration. [3]

• OpenAI Codex reaches 4 million weekly developers and expands to AWS as a generally available service. Codex's bundling inside ChatGPT plans drove rapid scale (3M in April, 4M+ by late May), and its June 1 availability on AWS gives enterprises a new deployment path through existing cloud relationships. The 0.136.0 release adds session archiving, Windows sandbox provisioning, Python SDK beta documentation, and image generation — further establishing Codex as a full-featured agent platform rather than a simple coding assistant. [4][5]

• Anthropic files confidentially for an IPO at nearly $1 trillion valuation while shipping multi-agent orchestration in Claude Code. The IPO filing follows a $65B Series H that pushed Anthropic's valuation to $965B, with revenue run-rate now at $47B (up from $9B at end of 2025). Simultaneously, Claude Code's Dynamic Workflows feature enters research preview, allowing users to orchestrate tens to hundreds of parallel agents for complex tasks like large migrations and security audits — a capability that could justify the tool's premium pricing tier. [6][7]

• JetBrains open-sources Mellum2, a 12B-parameter MoE coding model designed for the infrastructure layer of agentic systems. With only 2.5B parameters active per token (via 64 experts), Mellum2 achieves 21% faster inference than Qwen2.5-7B under concurrent load while scoring 78.4% on EvalPlus. Released under Apache 2.0 with base, instruct, and thinking variants, it targets enterprise teams that need to own their AI inference layer for routing, retrieval pipelines, and sub-agent tasks — positioning itself as the self-hosted alternative to API-dependent tools like Claude Code and Codex. [8]

• Kiro CLI ships thinking display and self-correcting subagent review loops, making autonomous agent work more transparent and reliable. Version 2.5.1 introduces real-time streaming of model reasoning, subagent pipelines that can review and revise their own output before returning results, and configurable display settings for accessibility. This pattern of agents that internally verify their work before presenting results to humans represents a maturing approach to reducing the review burden on developers. [9]

AI Coding News

• SkipLabs launched Skipper, a closed-loop coding agent that generates complete backend services without developer iteration. Created by Julien Verlaguet, Skipper takes a plain-language prompt or OpenAPI spec and produces a fully validated, running service in Docker — including routes, validators, TypeScript types, and unit tests. Unlike Claude Code or Cursor which keep developers in the review loop, Skipper runs its own internal review-and-refine cycle with up to eight attempts, leveraging a reactive runtime from Skip to handle state management and concurrency where AI typically fails. The $8M seed was led by Amplify Partners with angels including Yann LeCun. [10]

• GitHub Copilot now serves evaluation models to individual users through auto mode. Non-enterprise users may have pre-release or experimental models served during auto model selection for testing and feedback purposes. Users who prefer not to participate can opt out through their Copilot model settings. This approach gives GitHub a mechanism to A/B test new models at scale before promoting them to general availability. [11]

• OpenAI frontier models and Codex are now generally available on Amazon Web Services. This gives enterprises standardized on AWS infrastructure a direct path to integrate OpenAI's capabilities without managing a separate provider relationship, significantly expanding Codex's enterprise distribution beyond OpenAI's own platform and ChatGPT. [4]

Feature Update

• GitHub Copilot billing and plans overhaul goes live for all users. Usage-based billing via AI Credits is now active across all Copilot plans. New capabilities include user-level budget controls, configurable default Actions runners for code review, and Copilot Max as an upgrade path for power users. New sign-ups remain paused across Student, Pro, Pro+, and Max plans. [1]

• Copilot CLI v1.0.57 ships with process management and plugin improvements. Canceling shell commands now terminates the entire process tree instead of leaving orphans. Plugin slash commands show immediate feedback during operations. Default networking transport switches to HTTP/1.1 for reliability (HTTP/2 available via opt-in). Other fixes include tmux key handling, diff mode mouse selection, session resume after crashes, and proper preToolUse hook error handling that now denies tool calls on failure. [12]

• Copilot SDK v1.0.0-beta.12 introduces breaking Go changes and cross-language improvements. All Go identifiers now use idiomatic uppercase initialisms, and session boolean flags become tri-state *bool to distinguish explicit opt-out from unset. Java callers gain slash command response access via RPC, Python/Rust get stable public namespaces replacing internal generated sub-modules, and Node fixes the suppressResumeEvent wire mapping. [13]

• OpenAI Codex 0.136.0 adds session archiving, image generation, and Python SDK beta docs. Sessions can now be archived/unarchived from the TUI or CLI. Remote execution supports CODEX_API_KEY registration, Windows gets alpha sandbox provisioning via codex sandbox setup --elevated, and a standalone image generation extension connects to the native artifact pipeline. Security hardening prevents /diff from executing repository Git hooks and rejects browser-origin websocket handshakes. Python SDK documentation now presents the standard pip install openai-codex installation path. [5]

• Kiro CLI v2.5.1 adds thinking display, subagent review loops, and display controls. Real-time thinking display streams model reasoning as it works, enabled by default. Subagent pipelines gain self-correcting review stages where a reviewer can send work back to the implementer until quality meets the bar. New /settings display controls toggle animations, ASCII art, and icons for accessibility. Message and shell output streaming is noticeably faster. [9]

• Claude Code Dynamic Workflows enters research preview for parallel agent orchestration. The feature allows Claude Code to dynamically create orchestration scripts, break complex tasks into subtasks, distribute them across specialized agents, and validate results before presenting a final answer. Activation is via explicit prompt or the new "ultracode" setting. Available on Max, Team, and Enterprise plans, plus Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. Anthropic warns the feature consumes substantially more tokens than typical sessions. [7]