What's New on AI Developer Toolkit

The AI Developer Toolkit changelog records two things in one timeline: what changed in Cursor, Claude Code and Codex — releases, model launches, pricing and breaking changes — and what shipped in the guide itself. Entries run newest first, from September 2026 back to the version 1.0.0 release of July 2025, each carrying its date and version numbers.

The AI development landscape changes weekly. New models drop, pricing shifts, features land across all three tools, and the documentation that explains them has to move with it. This page is the record of both, so you can see at a glance whether a guide predates the release it describes.

Tip

Upstream changelogs: Cursor | Claude Code releases | Codex releases | OpenAI blog

September 2026

• This guide — the AI-native SDLC section. A six-stage AI-native SDLC section: the artifact chain, plan, design, build, test, deploy, maintain, a tool map, and metrics, each mapped onto Cursor, Claude Code, and Codex.
• This guide — human-in-the-loop default changed. Layered agent review plus human gates replaced blanket line-by-line review, which is now reserved for critical and regulated code.
• This guide — scorecards rebuilt. The developer and CTO scorecards now measure AI-native SDLC maturity, still at 25 questions and 75 points.
• This guide — current model defaults moved. The defaults were changed to Claude Fable 5 / Opus 5 / Sonnet 5 / Haiku 4.5, GPT-5.6 Sol/Terra/Luna, Grok 4.6, Gemini 3.8 Flash, and Muse Spark 1.3, with Cursor plans recorded as Pro $20, Pro+ $60, Ultra $200. The model snapshot further down this page has not been re-verified against those newer names — see the note there before quoting either.

August 2026

• Aug 28 · This guide — The Ladder, six free hubs. The documentation is now organised by autonomy level rather than by tool alone. Six hub pages, all outside the paywall and all EN+PL: ladder/index (the six-level table and a self-placement checklist), level-1-2-assisted-paired (where Shapiro puts about 90% of AI-native developers, and the per-tool setup that leaves it), level-3-review-diffs (parallel agents, worktrees, agent views and review automation in all three tools, with review bandwidth named as the ceiling), level-4-write-specs (plan mode, /goal, /loop, hooks, TDD, compound engineering, and the stop-condition rules), level-5-run-the-factory (the six stations — intent, harness, loop, graph, verification, release — mapped cell by cell to what Claude Code, Codex and Cursor document, each cell linked and dated), and the-humans-job (taste, architecture, product direction, eval and stop-condition design, permissions, and the disqualifying loops that must not run unattended). The Ladder group now opens the sidebar, written by hand rather than autogenerated so the order runs Level 0 to Level 5; no article moved and no URL changed (#371).
• Aug 28 · This guide — Routines. A gap article: the Level 5 material leaned on Claude Code’s unattended cloud sessions while the guide documented them nowhere. Covers all three trigger types (scheduled, API, GitHub), /schedule, the rule that a routine pushes to a claude/ branch instead of your default one, scoping which environments and connectors a routine may reach, the daily cap, and the untrusted-text wrapper a fire payload has to go through before an agent reads it. EN+PL (#371).
• Aug 28 · This guide — Cloud Agents and Automations. The second gap article, and Cursor’s column of the Level 5 matrix: cloud agents on isolated VMs, Builds and Dockerfile configuration, starting an agent without a repository open, the full Automations trigger list (schedule, GitHub, GitLab, Bitbucket, Slack, webhook, Linear, Sentry, PagerDuty), event subscriptions and subagents, the Cloud Agents API with its webhooks, and Bugbot, Security and Approval agents read as the verification station. EN+PL (#371).
• Aug 28 · This guide — The State of Agentic Engineering, August 2026. The research article behind the new spine, free and without an account: third-party evidence with publisher and date on how much production code agents write in 2026 (Anthropic’s merged-code share, Stripe’s agent-written pull requests, Microsoft’s rollout study), Dan Shapiro’s six levels from assisted coding to the dark factory, the six stations of a software factory mapped to Claude Code, Codex and Cursor, the counter-evidence (METR, Faros) given the same sourcing, and what stays human. Three copy-paste prompts place a repository on the ladder. EN+PL (#369).
• Aug 24 · This guide — compound engineering. Every’s loop (plan, work, review, compound) as a cross-tool guide, plus an audit of the plugin behind it: the published guide at every.to lists 26 agents, 23 commands and 13 skills — its own directory tree says 14 in the same document — while agents/ and commands/ both 404 and the repository ships 33 skills, 27 reviewer personas owned by ce-code-review, and 14 agent hosts; the marketplace URL it gives you redirects after a rename. Also the five-stage adoption ladder, the three different time splits Every publishes for one loop, why /lfg does not merge, and the argument almost nobody raises — the compound step as the guide describes it writes into CLAUDE.md, the always-loaded configuration Gloaguen et al. measure as net negative, which the shipped plugin avoids by making docs/solutions/ retrieved instead. Four copy-paste prompts run the loop with nothing installed; verified against plugin v3.23.3. EN+PL (#337).
• Aug 20 · This guide — the AI Developer Toolkit MCP server. The site answers MCP calls at https://developertoolkit.ai/mcp: a public, unauthenticated, read-only endpoint carrying the two tools the ChatGPT and Claude connectors look for — search across every guide on the site and fetch for one article’s full markdown. The guide covers the one-line setup for Cursor, Claude Code and Codex, the response contract that makes a connector’s citations resolve, why a Polish query returns Polish articles with no language parameter, and what the endpoint does not do. EN+PL (#319).
• Aug 19 · This guide — anti-slop writing skills. A deep dive on theclaymethod/unslop and bartekpucek/miodkuj: the three deterministic Python scanners that run in CI with no agent in the loop, the progressive-disclosure mechanics that let a skill ship a 50 KB rule catalogue for about 100 tokens of standing cost, the eval-first maintenance loop, and the finding that separates unslop from the rest of the category — its own core benchmark records a no-ship, while the most-starred skill in the genre, at more than 36,500 stars, ships no eval suite at all. miodkuj covers the Polish tells no English word list reaches (officialese, nominalisations, impersonal fog, genitive chains) and makes the opposite architectural bet, declining a deterministic linter on the grounds that a match list cannot resolve context. EN+PL (#323).
• Aug 13 · This guide — /goal and /loop tool support re-verified and corrected. The two autonomous-run guides asserted that /loop was native only to Claude Code and that Codex’s /goal was experimental and had to be enabled. Both were out of date: Cursor 3.5 shipped a /loop skill on May 20, 2026 (a month before the article was written), and Codex CLI 0.133.0 made goals stable and on by default on May 21, 2026. Re-checked against Claude Code 2.1.231 and Codex CLI 0.145 rather than the cached doc snapshots, which stop at 2.1.37 and February 2026. Also folded in the /loop behaviour that had accumulated since — the v2.1.196 rule that a scheduled fire only runs self-invocable skills, ScheduleWakeup stop: true in v2.1.202, the cloud-provider fixed-interval fallback, jitter and the 50-task cap — plus Codex’s /goal edit, and new parity rows in the feature matrix and command reference. EN+PL.
• Aug 12 · This guide — software factories. What a software factory is, why Dex Horthy (“Why Software Factories Fail”) and PostHog (“Can software factories actually work?”) are both right and are not arguing about the same thing, the ceiling SlopCodeBench measures (structural erosion rising in 77% of trajectories, agent code 2.3x more verbose and 2.0x more eroded than human code), the self-inflicted wound below it — pipelines that split “decide” from “build” and hand the agent a ticket stripped of production context — and how to meter autonomy loop by loop against a verification oracle. Faros AI’s 2026 telemetry puts numbers on the week six agents produce forty pull requests: median time to first review up 156.6%, merges with no review at all up 31.3%. EN+PL (#257).
• Aug 12 · This guide — Dr. Skill. A full guide to dbreunig/drskill, the read-only loadout diagnostic CLI: the tool-loadout research behind it (routing accuracy collapsing past 30 tools, a 600-skill enterprise agent), 34 checks grouped into skill hygiene, description quality, injection surfaces and MCP configuration, drskill audit reading the session traces your agents already write, a CI gate with a fingerprinted ack ledger, the invocation-time shell-command exposure, and the limits that make a clean scan less than a clean loadout; verified against version 0.6.7, cross-linked from the skills-ecosystem index, skill security, best practices and the skills CLI reference. EN+PL.
• Aug 6 · This guide — answer-first rewrite of the corpus. Every non-index article now opens with a self-contained paragraph that answers its own title before any preamble; template headings that said nothing were removed, faq and contentType frontmatter was added where it applies, and the rule was promoted into the build gate so a regression fails the build rather than shipping quietly (#214-#225).
• Aug 4 · This guide — the ablation protocol for context files. Pruning CLAUDE.md, .cursor/rules and AGENTS.md as a budget to cut rather than a log to grow: a dated map of why the guides contradict each other, what each tool actually loads and when (including Codex’s silent 32 KiB truncation), a routing table for all five destinations a rule can live in, nine prompts that do the cutting, and observed stumbles; three accuracy corrections followed post-merge. EN+PL (#212, #213).

July 2026

July reset the baseline across the stack: Anthropic closed the month with Claude Opus 5, OpenAI launched the GPT-5.6 family and folded Codex into ChatGPT desktop, Cursor shipped 3.11 plus Grok 4.5, and Claude Code added Sonnet 5 then Opus 5 to its account- and provider-specific model routing while hardening its always-on background-agent runtime.

• Jul 30 · This guide — herdr. The agent multiplexer that classifies every pane idle, working or blocked: how state detection actually works, sessions that outlive a closed laptop, the herdr agent CLI and the socket API beneath the TUI, the orchestrator pattern where one agent spawns and supervises the rest of the fleet, and where herdr fleets break down. The Moshi iPhone/iPad terminal was added to the mobile-access section shortly after, with its “no relay” claim scoped to hosting rather than data locality. EN+PL (#196, #205, #206).
• Jul 30 · This guide — tmux for agent fleets. What tmux is still for now that Claude Code, Cursor and Codex create and clean up their own worktrees, which puts every “loop over git worktree add” guide a layer too low: session per task, send-keys as the agent-to-agent channel, capture-pane polling, wait-for barriers, and hooks that report exact state instead of screen-scraping. EN+PL (#196).
• Jul 30 · This guide — alternative IDEs and agent shells. A survey of the shells you can run Claude Code, Codex and Cursor Agent inside — Zed and JetBrains over ACP, herdr, Conductor, Nimbalyst, Emdash, Sculptor, Warp — sorted into four shapes, with one rule for what even qualifies (does it run your agent?), a decision table, a terminal baseline to try before buying anything, and the repo changes that make parallel agents survivable. EN+PL (#195).
• Jul 27 · This guide — Claude Opus 5 swept across 187 files. New Opus 5 sections in the model-comparison appendix and Cursor’s model-selection guide, the two API breaking changes, Claude Code v2.1.219 and category-based fallback routing, a refreshed Artificial Analysis snapshot, and Opus 4.8 recast as the previous tier and the cyber-refusal fallback target. Three claims that had gone stale independently were corrected at the same time: Fable 5’s plan access, Gemini 3.1 Pro’s above-200K price tier, and the Opus 4.8 effort default. Model, pricing, context, effort and fallback-routing claims were rechecked against Anthropic’s models overview, the pricing page, the Claude Code model-config docs, and the Artificial Analysis Opus 5 analysis. EN+PL.
• Jul 24 · Anthropic — Claude Opus 5. The new default Opus at Opus 4.8’s price ($5/$25 per 1M tokens), half of Fable 5. Tops the Artificial Analysis Intelligence Index, carries a May 2026 knowledge cutoff, and needs Claude Code v2.1.219+. Two API breaking changes: thinking is on by default, and disabling it is capped at high effort. v2.1.219 also adds category-based fallback — cybersecurity-flagged requests re-run on Opus 4.8, while biology-flagged Opus 5 requests end in a refusal.
• Jul 25 · This guide — Mobbin MCP. Ask an agent for a paywall and you get the statistical average of every paywall in its training data. Mobbin MCP connects it instead to 620,000+ screens and 130,000 user flows captured from shipped iOS and web products: the three search tools, setup across Cursor, Claude Code and Codex, query patterns that actually return something, context-cost management, the MCP-versus-REST-API split, and the other design-reference servers including Refero. EN+PL (#183).
• Jul 25 · This guide — Impeccable rebuilt for v4. The deep dive was expanded about 2.5x and re-verified against v4.0.2: direction dealing from 188 hand-approved worlds, five direction tests, system board plus first-surface mock before code, four visitor modes, the six-intent command regrouping with a command-by-command reference for all 23 commands (craft deprecated in favour of conversational job classification), Live Mode’s pins/strokes/three-variant workflow, the critique overlay, design hooks with two-speed reporting, doctor staleness checks, the native iOS/Android alpha, and the detector catalog at 64 patterns and 59 enforced rules, up from 44. EN+PL.
• Jul 21 · This guide — Taste Skill. A deep dive on Leonxlnx/taste-skill (~65K stars), the anti-slop frontend skill library: how it constrains AI generation up front (brief inference, three dials, design-system map, hard ban list, pre-flight check), all 13 aesthetic variants with their install names, per-agent install, honest scope limits, and how it composes with Impeccable and the design pipeline; verified against upstream v2. EN+PL (#170).
• Jul 20 · This guide — tool changelogs refreshed through July 19. Claude Code v2.1.208-215, Codex v0.144.2-0.144.6, the July 17 Cursor-in-Slack update, and a clarification that GPT-5.6’s ~272K Codex working context is not the same number as its 1.05M API context. Earlier tool-changelog and acquisition-status claims were verified against upstream the same day.
• Jul 14-19 · Claude Code 2.1.208-2.1.215, Codex CLI 0.144.6, Cursor 3.11 line. Claude Code stopped auto-running /verify and /code-review and made them explicit skills, added the EndConversation tool, /fork to background sessions, a screen-reader mode, and Windows PowerShell permission hardening. Codex CLI 0.144.6 tightened dangerous-command detection and corrected GPT-5.6 Sol/Terra/Luna to a 272K working context. Cursor added Slack multi-repo environments and plan-before-start on July 17.
• Jul 11 · This guide — Superpowers. A deep dive on obra/superpowers: how a 14-skill plugin turns brainstorming, worktrees, executable plans, TDD, subagent review and fresh verification into one enforced design-to-PR workflow, with all ten installation targets, evidence, security and honest limits; verified against upstream v6.1.1, cross-linked from the skills ecosystem, core methodology and the scorecard guide. EN+PL (#159).
• Jul 11 · This guide — GPT-5.6, Sonnet 5 and Grok 4.5 added across model guides, pricing and routing. Also Claude Code v2.1.185-207, Cursor 3.8-3.11 (/automate, cloud subagents, iOS, Team MCP marketplaces, side chats), and Codex v0.142-0.144.1. GPT-5.5 was removed from current defaults and replaced by role-aware GPT-5.6 Sol/Terra/Luna guidance. The docs also began distinguishing SpaceX’s pending Cursor acquisition from the inaccurate claim that xAI already bought it, explaining why Grok 4.5 does not replace Composer 2.5, and separating provider claims from current Artificial Analysis v4.1/v1.1 results.
• Jul 9 · OpenAI — GPT-5.6 Sol, Terra and Luna. Sol is the flagship, Terra balances capability and cost, and Luna targets high-volume work. All support max reasoning; the API adds persisted reasoning, explicit caching, Programmatic Tool Calling and a multi-agent beta.
• Jul 9 · OpenAI — Codex joins ChatGPT desktop, CLI 0.144.1. Codex is integrated into the ChatGPT desktop app with inline editing, PR review and multi-repo projects. CLI 0.143-0.144 adds remote plugins by default, GPT-5.6 on Bedrock, writes-only approvals and interactive MCP auth.
• Jul 8-10 · Cursor — 3.11 and Grok 4.5. Jointly trained with SpaceXAI, Grok 4.5 lands across desktop, web, iOS, CLI and SDK without replacing the smaller Composer 2.5. Cursor 3.11 adds durable side chats, transcript search, redesigned project pickers and conversation-level cloud hooks.
• Jul 3 · This guide — “What Claude Code Can Do That You Haven’t Tried”. About 40 underused Claude Code features from Gui Ferreira’s NDC AI 2026 talk, grouped into one-time setup, your daily loop, extensibility, and work beyond code; every command verified against the current docs. EN+PL (#146).
• Jul 1 · Anthropic — Claude Fable 5 restored globally, with billing settled on July 20: permanently included on Max and Team Premium at up to 50% of weekly usage limits, with Pro and Team Standard on usage credits. The highest-capability tier remains an explicit selection, never an automatic account default.
• Claude Code 2.1.185-2.1.207 · Anthropic. Claude Sonnet 5 became the default on standard Anthropic subscription seats; selected premium and direct-API accounts default to Opus 4.8, while managed-cloud defaults differ. The releases make background agents durable across restarts and add MCP CLI auth, sandbox credential controls, dynamic-workflow sizing, stacked skills, routines, current /effort controls, and a /doctor that both diagnoses and fixes.

June 2026

The autonomous-work race kept pace: Cursor pushed Design Mode into the browser and canvases, Claude Code locked in ultracode as its dynamic-workflow trigger keyword, Codex shipped controller pairing for headless fleets, and Anthropic shipped Claude Fable 5, a new model tier above Opus.

• Jun 24 · This guide — /grill-me and /grill-with-docs. Two of Matt Pocock’s most-installed skills: a relentless one-question-at-a-time agent interview that resolves every branch of your plan before you build, plus, with docs, a ubiquitous language (CONTEXT.md) and ADRs written as you go. Cross-linked from Plan mode and the skills ecosystem. EN+PL.
• Jun 24 · This guide — /goal and /loop, plus two skill deep dives. Cross-tool coverage of goal-directed and looping agent runs, with the gaps in each tool mapped honestly (tool support corrected on Aug 13). Alongside them, deep dives on shadcn/improve — an architecture audit that writes plans for cheaper executors — and Impeccable, a design vocabulary that catches AI slop; both verified across Claude Code, Cursor and Codex. EN+PL (#129).
• Jun 18 · Anthropic — Claude Code Artifacts (beta). Turn a Claude Code session into an interactive page — a PR walkthrough or living dashboard — shared at a private link. Beta on Team and Enterprise.
• Jun 18 · Cursor 3.8 — /automate. Describe an automation in plain language and the agent wires up the triggers, instructions and tools. Plus a Slack emoji trigger and five new GitHub triggers.
• Jun 18 · Codex v0.141 and Record & Replay. Encrypted Noise relay channels for remote executors and per-thread executor MCP servers; the Codex app’s Record & Replay turns a demoed workflow into an editable skill.
• Jun 18 · This guide — tool changelogs refreshed through June 18. Claude Code v2.1.183 (Artifacts, nested subagents, implicit agent teams), Cursor 3.8, and Codex v0.141.
• Jun 17 · Cursor — cloud subagents (/in-cloud, /babysit). Spin up cloud subagents in isolated VMs, have them prepare PRs remotely, and set up a cloud dev environment in under 10 minutes.
• Jun 16 · Cursor — SpaceX signs an agreement to acquire Anysphere. Stock consideration at an implied $60B equity value. The transaction remained pending regulatory approval and closing conditions, with Q3 2026 completion expected.
• Jun 10 · This guide — Claude Fable 5 covered site-wide. Model comparison, routing strategies, pricing and the Claude Code changelog v2.1.169-170 (--safe-mode, /cd, Fable 5).
• Jun 9 · Anthropic — Claude Fable 5, a new tier above Opus. A Mythos-class model made safe for general use; in Claude Code from v2.1.170 (/model fable) and in Cursor’s model picker. $10/$50 per 1M tokens.
• Jun 8 · This guide — changelog sweep and 11 new official skills publishers. Claude Code v2.1.156-168, Cursor 3.6-3.7, Codex confirmed current at v0.137, and OpenAI, Google Gemini, LangChain, Vercel, Prisma, Convex, Deepgram, Box, Contentful, Contentstack and Base added to the official skills guide. The release-notes hub and the update tracker were refreshed with June model and version coverage.
• Jun 6 · Claude Code v2.1.166 — fallback models. Configure up to three fallbackModels tried in order when the primary is overloaded, glob patterns in deny rules (* denies all tools), and hardened cross-session messaging.
• Jun 4-5 · Cursor 3.7 — Canvas Design Mode. Design Mode now works in the browser and inside canvases: multi-select elements, voice narration while an agent runs, an interactive context-usage report, and full-screen shared canvases. Cursor Organizations for Enterprise followed on June 3.
• Jun 4 · Cursor SDK — custom tools and auto-review. Expose your own functions to the local agent via local.customTools, route headless tool calls through an auto-review classifier, pick JSONL or SQLite stores, and nest subagents to any depth.
• Jun 4 · Codex v0.137 — controller pairing, parallel web search. codex remote-control clients can initiate pairing and manage controller grants; plugin list gains JSON output; standalone web searches run in parallel. v0.136 added /archive sessions.
• Jun 2 · Claude Code v2.1.160 — ultracode is the workflow keyword. The dynamic-workflow trigger was renamed from workflow to ultracode, highlighted in violet; typing “workflow” no longer starts a run, though asking in your own words still does. v2.1.156-158 added .claude/skills plugin auto-load and auto mode on Bedrock, Vertex and Foundry.
• Claude Code v2.1.172-183 · Anthropic. Nested sub-agents up to 5 levels deep, implicit agent teams (TeamCreate/TeamDelete removed), parameter-scoped permission rules, and auto-mode guardrails against destructive git and terraform commands.
• This guide — MCP token-cost and skill-security guides. Reducing MCP token cost with code execution, and skill supply-chain security (SkillSpector), plus Chrome DevTools MCP coverage and research-backed prompting updates.

Fable 5 early results -- and Claude Mythos 5

During early testing, Stripe reported that Fable 5 compressed months of engineering into days. In a 50-million-line Ruby codebase, the model performed a codebase-wide migration in a day that would otherwise have taken a whole team over two months by hand.

Claude Mythos 5 is the same underlying model as Fable 5, but with safeguards lifted in some areas. Initial access is restricted to Project Glasswing cyber defenders and critical-infrastructure providers.

May 2026

A pivotal month: Anthropic shipped Claude Opus 4.8 and dynamic workflows, and all three tools rounded out their first-class autonomous-work primitives within weeks of each other.

• May 29 · Cursor 3.6 — Auto-review Run Mode. A run mode that works longer with fewer prompts: allowlisted calls run immediately, sandboxable calls are sandboxed, and the rest go to a classifier subagent.
• May 28 · Anthropic — Claude Opus 4.8 and dynamic workflows. The new flagship, around four times less likely than Opus 4.7 to leave code flaws unflagged, plus dynamic workflows (research preview, v2.1.154+) that orchestrate tens to hundreds of subagents from a script — up to 1,000 — and ultracode to let Claude decide when to run one. Also /workflows, fast mode at 2x and 2.5x, and a lean system prompt by default.
• May 28 · This guide — tool changelog and model refresh. Opus 4.8 and dynamic workflows, Cursor 3.3-3.5 and Composer 2.5, Codex CLI 0.131-0.135; model references normalized site-wide.
• May 21 · Codex v0.133 — goals on by default. Persisted goals graduate from a feature flag to stable and enabled by default, backed by dedicated storage that tracks progress across active turns.
• May 21 · This guide — per-question implementation guides for the Developer and CTO scorecards (#40).
• May 20 · Cursor 3.5 — the /loop skill, shared canvases. /loop runs a prompt on a local schedule until an outcome is reached, self-paced when no interval is given. Automations move into the Agents Window, gaining multi-repo and no-repo forms.
• May 14 · This guide — CI reviewers coverage. Claude Code Action, Codex and Sentry Seer listed in the AI maturity survey (#29).
• May 11 · Claude Code v2.1.139 — /goal and the agent view. /goal sets a verifiable end state and Claude works turn by turn until it is met; claude agents is the new dashboard for every background session.
• May 11 · Cursor in Microsoft Teams. Mention @Cursor in any Teams channel to delegate work to cloud agents, with automatic repo and model selection from prompt context.
• May 9 · This guide — bilingual Developer and CTO scorecards with Gemini analysis (#17).
• May 8 · Codex v0.130 — plugin sharing, remote-control. codex remote-control for headless orchestration, plugin sharing with link metadata and discoverability, AWS Bedrock console-login, and multi-environment view_image.
• May 7 · Cursor 3.3 — PR review, Build in Parallel. Redesigned PR review (Reviews/Commits/Changes tabs), a Build in Parallel quick action, Split changes into PRs, and pinned quick-action pills. Bugbot moves to usage-based billing.
• May 7 · Codex v0.129 — /vim, /hooks browser. Vim modal editing in the composer, a redesigned resume/fork picker, a theme-aware status line, and a /hooks browser with pre- and post-compaction execution.
• May 2026 · Claude Code — 5-hour usage limits doubled for Pro, Max, Team and Enterprise, to support the new long-running agent workflows.
• Claude Code v2.1.144-154 · Anthropic. /resume for background sessions and claude agents --json; /simplify and /code-review become distinct commands in the current command set.
• Codex CLI v0.131-0.135 · OpenAI. codex doctor, goals on by default, history search, --profile standardization, and codex remote-control in the foreground. GPT-5.5 was the default at that historical point.
• Cursor 3.3-3.5 and Composer 2.5. PR review, Build in Parallel, full-screen tabs, Composer 2.5 at $0.50/$2.50 (fast variant $3/$15), Jira integration, shared canvases, and Automations in the Agents Window.

The story of May 2026

Both Claude Code and Codex shipped /goal within two weeks of each other — Codex v0.128 on April 30, Claude Code v2.1.139 on May 11. They differ in scope: Codex’s is persistent and round-trips through the app-server; Claude Code’s is session-scoped and clears on /clear. The same week, Cursor 3.3 added Build in Parallel and Split-into-PRs — the same “autonomous from a plan” pattern, surfaced as quick actions on the editor side.

Late May closed the gap from both ends. Codex made goals stable and on by default in v0.133 on May 21, one day after Cursor 3.5 shipped its own /loop skill — which runs a prompt on a local schedule or until an outcome is reached, collapsing into one command what Claude Code splits between /loop and /goal.

The convergence is no accident: all three are moving toward background agents that finish on their own and report back.

April 2026

• Apr 30 · This guide — the four-stage AI design pipeline section, EN+PL (#6).
• Apr 30 · Codex v0.128 — persistent /goal, permission profiles. Persistent /goal workflows that survive --resume, built-in permission profiles, plugin marketplace install, and MultiAgentV2 made explicit.
• Apr 24 · Cursor 3.2 — /multitask. Async subagents running in parallel rather than queued, plus multi-root workspaces for cross-repo agent sessions in a single conversation.
• Apr 24 · This guide — changelogs refresh, model IDs, and April 2026 lesson content.
• Apr 5 · This guide — Cursor 3.0, Claude Code v2.1.84-92 and the Codex desktop/CLI updates documented, GPT-5.4 recorded as the Codex default in IDE and Cloud surfaces, and Wondel.ai skills documented across the skills ecosystem.
• Apr 4 · This guide — Official Skills from the Technology Makers docs page, EN+PL.
• Cursor 3.0 — Agents Window and Design Mode. A fundamental shift from the VS Code fork to a purpose-built agent workspace. The Agents Window is a multi-repo interface for running parallel agents across environments — local, worktrees, cloud and remote SSH — with a centralized sidebar showing all agents (Cmd+Shift+P → Agents Window). Design Mode adds browser-based UI annotation for precise visual feedback (⌘+Shift+D to toggle, Shift+drag to select areas, ⌘+L to add elements to chat, ⌥+click to add to input). Agent Tabs and Worktrees show multiple chats side by side or in a grid, with a new /worktree command for isolated git worktrees and /best-of-n to run a task across multiple models in parallel. On the enterprise side, MCP Apps support structured content, third-party plugins default to off, directory group names appear in audit logs, and self-hosted cloud agents gain secret management.
• Claude Code v2.1.84-92 · Anthropic. A PowerShell tool for native Windows development without WSL (opt-in preview); a Bedrock setup wizard with AWS authentication, region configuration, credential verification and model pinning; the /powerup command with animated lessons; per-model and cache-hit cost breakdown in /cost; deferred hook permissions (a "defer" decision in PreToolUse hooks for headless sessions, plus a PermissionDenied hook firing after auto-mode denials); the Write tool 60% faster on large files; named subagents in @ mention typeahead with declarable initialPrompt frontmatter; conditional hooks via an if field using permission-rule syntax; and transcript search (/ in Ctrl+O mode, n/N to step through results). /tag and /vim were removed in v2.1.84+.
• ChatGPT desktop (then the Codex app) and CLI · OpenAI. Thread search with a sidebar shortcut and keyboard shortcuts for recent threads; theming with a base theme plus accent, background, foreground colors and fonts; revamped automations offering local or worktree execution, custom reasoning levels and reusable templates; a Windows sandbox with proxy-only networking and OS-level egress rules; plugins as first-class, synced product-scoped at startup and browsable in /plugins; and sub-agents with path-based addressing (/root/agent_a) and structured inter-agent messaging.

March 2026

• Mar 28 · This guide — March 2026 changelogs for Claude Code and Codex, EN+PL.
• Mar 18 · OpenAI — GPT-5.4. The first general-purpose model with native computer-use capabilities, now the default across all OpenAI surfaces: 75% on OSWorld, up to 1M tokens of context, 57.7% on SWE-bench Pro and 83% on GDPval. It incorporates the coding capabilities of GPT-5.3-Codex while improving tool use across software environments, and a GPT-5.4 Pro variant is available for maximum performance.
• Mar 18 · Cursor Composer 2. Cursor’s frontier coding model built in-house: a Mixture-of-Experts architecture on Kimi K2.5 with Cursor’s own continued pretraining and reinforcement learning; 61.3 CursorBench, 61.7 Terminal-Bench 2.0, 73.7 SWE-bench Multilingual, beating Claude Opus 4.6 on Terminal-Bench; $0.50/$2.50 per 1M tokens standard and $1.50/$7.50 for the fast variant with the same intelligence; 200K context optimized for multi-file edits and long task chains. It replaced Composer 1, and Marketplace Plugins shipped alongside.
• Mar 18 · Cursor Automations. Always-on agents that run on schedules or respond to events: Slack, Linear, GitHub, PagerDuty and webhook triggers spinning up cloud sandboxes with configured MCPs and models; a memory tool so agents improve across recurring tasks; security templates reviewing 3,000+ internal PRs per week and catching 200+ vulnerabilities; and a fully automated incident pipeline from PagerDuty alert through Datadog log analysis to Slack summaries with proposed fixes.
• Mar 18 · Cursor Self-Hosted Cloud Agents. An enterprise feature for running cloud agents inside customer infrastructure: codebase, tool execution, build outputs and secrets never leave the customer network, with the same capabilities as Cursor-hosted agents, up to 10 workers per user and 50 per team. Used by Brex, Money Forward and Notion.
• Mar 12 · Codex Security (research preview). An application security agent for Enterprise, Business and education customers: project-specific threat models capturing system structure, trust boundaries and exposure points; deep vulnerability detection for issues static analysis misses, categorized by real-world impact; 1.2M+ commits scanned in beta with 792 critical findings including GnuTLS heap overflows and 2FA bypasses; and editable threat models teams can keep aligned with their requirements.
• Mar 12 · Claude Code v2.1.81-86. A --bare flag for scripted -p mode that skips hooks, LSP and plugins; --channels permission relay for phone forwarding; a managed-settings.d/ drop-in directory for enterprise config composition; transcript search; Ctrl+X Ctrl+E to open an external editor for prompt composition; and pasted images inserting an [Image #N] chip.
• Mar 4 · OpenAI — ChatGPT Desktop for Windows (then the Codex app), CLI 0.117-0.118. The Windows app adds native PowerShell support and a Windows-native agent sandbox with OS-level restricted tokens and filesystem ACLs. CLI 0.117 makes plugins a first-class workflow with product-scoped sync, adds sub-agents with path-based addressing and terminal title configuration via /title; CLI 0.118 adds Windows sandbox proxy-only networking, ChatGPT device-code sign-in, prompt+stdin for codex exec, and dynamic bearer tokens for custom model providers.

February 2026

• Feb 18 · This guide — subscribers page expanded to showcase all 390+ pages of content.
• Feb 13 · This guide — Remotion benefits videos for Cursor and Claude Code.
• Feb 9 · This guide — Polish translations across all documentation sections.
• Feb 8 · This guide — three-tool coverage. A full Codex section, plus restructured Cursor and Claude Code sections.
• Feb 3 · This guide — changelog sweep. Claude Code v2.1.17-v2.1.30, Cursor v2.3-v2.4 with Subagents and Skills, GPT-5.1-Codex-Max updated to GPT-5.2 Codex everywhere, and Opus 4.5 positioned as the best model of the moment.
• Anthropic — Claude Opus 4.6. The newest Claude model at the time and the new default recommendation for complex coding: the best SWE-Bench Verified performance among available models, improved tool use across hundreds of tools, better prompt-injection resistance, more reliable multi-step execution, a 200K context window with a 64K output limit, an effort parameter for adjustable reasoning depth, and memory improvements. Available in Claude Code, Cursor’s model picker, and the Anthropic API.
• Claude Code v2.1.30. PDF page ranges via the Read tool’s pages parameter (for example pages: "1-5"), a new /usage command with detailed input/output token breakdown, better OAuth error handling and refresh-token management, and MCP health checks with automatic restart for failing servers.
• Cursor v2.4 — Subagents and Skills. Specialized subagents for testing, docs and refactoring, definable in .cursor/agents/; reusable skill definitions with SKILL.md manifests shareable across projects and teams; 40x faster hooks; Cursor Blame, an AI-powered git blame with context; agents that ask clarifying questions before complex tasks; CLI Plan and Ask modes (cursor --plan, cursor --ask); and cloud handoff with an & suffix.
• OpenAI — GPT-5.3-Codex. The model powering Codex surfaces at that release point: improved coding performance across all languages, better long-context handling with automatic compaction, enhanced tool use for MCP integrations, and availability in the desktop app, CLI, IDE extensions and Cloud.

January 2026

• Claude Code v2.1.17-25. SKILL.md support for custom skills in .claude/skills/, session forking from any conversation point, a --from-pr flag to start with context from a GitHub pull request, arrow-key prompt history, and cloud handoff with an & suffix.
• Cursor v2.3. Drag-and-drop layout customization, multi-monitor layout sync, a 25% memory reduction for large files, and improved crash recovery for long sessions.
• Codex Automations reach general availability. GitHub issue assignment to @codex, Slack integration for team channels, Linear ticket automation, and scheduled cron-based tasks — all running as sandboxed cloud agents.

December 2025

• Cursor v2.2 — Debug Mode and visual editing. Runtime log instrumentation for automatic root-cause analysis across multiple stacks and languages; a Visual Style Editor for real-time design in Cursor Browser; multi-agent judging, where parallel agents run the same task and an automatic evaluation picks the best solution; and pinned chats in the agent sidebar.
• Claude Code v2.0.x — background agents. Run tasks while you keep working, name sessions with /rename and resume them with /resume <name>, use a .claude/rules/ directory alongside CLAUDE.md, accept suggested follow-up prompts with Tab, and switch models mid-prompt with Alt+P.
• Codex Cloud launches. Background agents running on OpenAI infrastructure, with git worktree isolation for every task, up to 10 parallel tasks, GitHub PR creation on completion, and Slack and Linear notification integrations.

November 2025

• Nov 24 · Anthropic — Claude Opus 4.5. The first model to break 80% on SWE-Bench Verified, with enhanced prompt-injection resistance and agent capabilities. The Claude Code Desktop app launched alongside it with all CLI features, and Pro users gained access to Opus 4.5.
• Nov 19 · OpenAI — GPT-5.2 Codex. Replaced GPT-5.1-Codex-Max, specialized for bug fixing and UI generation, with 24+ hour task endurance via compaction. Available in Cursor and GitHub Copilot.

October 2025

• Oct 29 · Cursor v2.0 — the multi-agent era. Up to 8 agents running simultaneously, Cursor Composer 1 as a 4x faster model at 250 tokens per second, sandboxed terminals for secure execution on macOS, and voice mode for speech-to-text agent control.
• Cursor v2.1 — AI code reviews. In-editor bug detection examining your changes, Instant Grep in beta for all agent grep commands, and an improved Plan Mode with interactive clarifying questions.
• Codex CLI open-source release. OpenAI open-sourced the Codex CLI: full source on GitHub, community contributions accepted, a plugin architecture for extensions, and cross-platform support for macOS, Linux and Windows.

September 2025

A snapshot of what the tools had reached by the end of the guide’s first quarter. Cursor had shipped Cursor Browser (beta) for in-app testing, native Plan Mode (beta) for AI-assisted task planning, Claude Sonnet 4.5 support and the GPT-5.2 Codex model, on top of background agents, BugBot, Max Mode, checkpoints and branching, multi-root workspaces, enhanced @ symbols, and Memories (beta). Claude Code had architect mode (--enable-architect), extended thinking controls (a thinking toggle and /effort, with ultrathink as a one-turn prompt hint), the hooks system, enhanced MCP integration, headless-mode automation, multi-directory workspaces, and custom slash commands.

August 2025

• Aug 7 · This guide — GPT-5 guidance added across model selection, configuration, comparison and the FAQ. Positioning at the time: strong for one-shot app creation from detailed PRDs, while experienced developers may still prefer Opus for architecture and planning depth. Source: https://cursor.com/blog/gpt-5.

July 2025 — version 1.0.0

• Jul 17 · This guide — version 1.0.0, the initial comprehensive release. Complete Cursor documentation (100+ articles) and complete Claude Code documentation (100+ articles), 200+ articles in total, a shared workflows section, an enterprise development guide, MCP ecosystem documentation and comprehensive appendices. Key features at launch: the PRD to Plan to Todo methodology, 100 tips for each tool, a model comparison guide, a pricing calculator, and migration checklists. Supported versions were Cursor 0.40.x and above, Claude Code 0.8.x and above, Claude Sonnet 4 and Claude Opus 4 at GA, initial OpenAI o3 support, and full Gemini 2.5 Pro support.
• Jul 16 · This guide — o3 model details added.
• Jul 15 · This guide — subscription tiers updated.
• Jul 14 · This guide — 40 MCP server guides added.
• Jul 10 · This guide — the security and compliance section added.

Breaking changes at the 1.0.0 boundary

The July 2025 release carried three sets of breaking changes. Claude Code CLI: a new installation method via npm, changed config file locations, and updated slash-command syntax. Cursor: a new settings structure, changed keybindings for some features, and an updated extension API. Pricing: a new tier structure for both tools, changed token limits, and updated team pricing.

Current model snapshot

Model	Provider	Context	Typical use	Pricing (per 1M tokens)
Claude Fable 5	Anthropic	1M	Tasks that outlast a single sitting; thinking always on	$10 / $50
Claude Opus 5	Anthropic	1M	The default Opus: agentic coding, computer use, knowledge work. Account default on Max, Team Premium, Enterprise pay-as-you-go, the Anthropic API and the managed clouds. Requires Claude Code v2.1.219+	$5 / $25
Claude Sonnet 5	Anthropic	1M	Default on Pro, Team Standard and Enterprise subscription seats	$2 / $10 introductory
Claude Haiku 4.5	Anthropic	200K	Cheap parallel and bulk work, subagents	$1 / $5
GPT-5.6 Sol / Terra / Luna	OpenAI	1.05M API (~272K working context in Codex)	Frontier / balanced / efficient	$5/$30 · $2.50/$15 · $1/$6
Grok 4.5	SpaceXAI / Cursor	500K API*	Long-running coding and computer work	$2 / $6
Cursor Composer 2.5	Cursor	200K	Fast local iteration in Cursor	$0.50 / $2.50
Gemini 3.1 Pro	Google	1M	Multimodal, extreme context, Deep Think mode	$2 / $12; $4 / $18 above 200K input

* A host such as Cursor may expose a smaller working context; verify the current model picker.

Caution

This table reflects the verification baseline of July 27, 2026. The September 2026 AI-native SDLC pass recorded newer defaults — Grok 4.6, Gemini 3.8 Flash and Muse Spark 1.3, with Cursor plans at Pro $20, Pro+ $60 and Ultra $200 — which this table has not been re-verified against. Where the two disagree, check the vendor’s own page before quoting either.

Previous or superseded guidance

• GPT-4, GPT-4.1, GPT-4o (legacy relative to current GPT-5.6 guidance; availability depends on product)
• GPT-5.1 / GPT-5.1-Codex / GPT-5.1-Codex-Mini/Max, GPT-5.2 (removed from the Codex model picker)
• GPT-5.3-Codex, GPT-5.4, and GPT-5.5 (superseded by the GPT-5.6 family)
• Cursor Composer 1 and Composer 2 (older generations; Composer 2.5 is current)
• Claude Opus 4.8 (superseded by Opus 5 at the same $5 / $25 price; still available, still Priority-Tier eligible, and still the model that cybersecurity-flagged Opus 5 and Fable 5 requests fall back to)
• Claude Opus 4.6 / 4.7 (older than Opus 4.8; v2.1.219 moved the documented managed-cloud defaults to Opus 5, except Foundry, which stays on Sonnet 4.5 and resolves opus to Opus 4.6)
• Claude Opus 4.1 (deprecated — retires August 5, 2026; migrate to Opus 5)
• Claude 3.x series (legacy relative to current 4.5+ and 5-series guidance)
• o1-preview, o1-mini (obsolete)

MCP and Agent Skills ecosystem

Recent official MCP servers worth knowing: Atlassian (Jira and Confluence), Cloudflare (Workers, R2 and D1), Vercel (deployment automation), Supabase (database and auth operations), Linear (issue tracking) and Stripe (payment processing).

On the skills side, the Skills.sh marketplace now hosts 200+ skills, and the universal CLI npx skills add <owner/repo> works across 35+ agents. Popular skills include Cloudflare deployment, code review, Atlassian integration and API documentation generation.

Security and compliance across the three vendors

SOC 2 Type II

Verify each vendor’s current trust center and the product and contract scope of its SOC reports.

Data controls

Business and API offerings provide contractual data controls; consumer settings, retention, and training choices vary by product and plan.

Audit logging

Enterprise audit capabilities vary by tool, surface, and contract; confirm the events covered before relying on them for compliance.

GDPR

Data-processing terms and regional residency depend on vendor and enterprise contract; verify the current DPA and region support.

Breaking changes that still need action

1. Claude Sonnet 5 (June 30): default on Pro, Team Standard and Enterprise subscription seats with native 1M context. Max, Team Premium, Enterprise pay-as-you-go and Anthropic API sessions default to Opus 5, managed-cloud defaults differ, and introductory Sonnet API pricing ends August 31, 2026.
2. Claude Fable 5: restored globally July 1; since July 20 permanently included on Max and Team Premium at up to 50% of weekly usage limits, while Pro and Team Standard use usage credits.
3. GPT-5.6 family (July 9): Sol, Terra and Luna replace GPT-5.5-era guidance; review model slugs, reasoning effort and tier routing.
4. Cursor Grok 4.5 (July 8): available across desktop, web, iOS, CLI and SDK; choose standard or fast pricing.
5. Cursor 3.0: the Agents Window interface replaces the old layout, and cloud agents were removed from the Editor.
6. Claude Code: /simplify and /code-review are distinct commands — /simplify for reuse, quality and efficiency review, /code-review for the dedicated review workflow.
7. Claude Code v2.1.84+: /tag and /vim removed; use /config → Editor mode instead.
8. Claude Code v2.1.160 (June 2): the dynamic-workflow trigger keyword changed from workflow to ultracode — typing “workflow” no longer starts a run, though asking in your own words still does.

Migration notes

These rows describe documentation or workflow migrations, not blanket API deprecation or end-of-support claims. Verify product availability with the provider.

Previous guidance or workflow	Snapshot changed	Current path
GPT-5.5-era Codex guidance	July 2026	GPT-5.6 Sol / Terra / Luna
Cursor Composer 2	May 2026	Composer 2.5
Cursor v2.x Editor cloud agents	April 2026	Agents Window
Cursor Composer 1	March 2026	Composer 2 / 2.5 lineage
Cursor v1.x UI	February 2026	v3.x Agents Window
Claude Code WSL-only setup	August 2025	Native Windows or optional WSL
Older MCP specification revisions	Date-specific revisions	Current dated MCP specification and SDKs

Keeping your own tools current

Cursor

Settings > Update > Auto-update. Choose the “Stable” or “Beta” channel.

Claude Code

claude update                 # update to latest
claude --version              # check current version

Codex

codex update                              # supported standalone installs
npm install -g @openai/codex@latest       # npm alternative
codex --version               # check current version

Beyond the changelogs linked at the top of this page, the vendor blogs are Cursor, Anthropic and OpenAI; each also announces on X/Twitter and in its Discord.

How this guide is kept current

Every entry above goes through four steps before it ships. Official documentation review: vendor release notes, API changes and pricing updates. Community validation: Discord and Slack channels, GitHub issues, user feedback. Hands-on testing: new features, workflows and examples run rather than paraphrased. Expert review: technical accuracy, best-practice validation and enterprise scenarios.

Articles carry version tags recording the tool versions their content assumes:

---
version: 1.0.0
lastUpdated: 2026-07-27
tools:
  cursor: ">=3.11"
  claude_code: ">=2.1.215"
  codex_cli: ">=0.144.6"
---

Content moves through four lifecycle stages: Active (current best practice), Legacy (still valid, but newer methods exist), Deprecated (marked for removal) and Archived (historical reference only). GPT-4 examples, for instance, were deprecated in July 2025 and removed in October 2025; the old pricing tiers were marked Legacy in July 2025 and removed in December 2025; the Claude 3.x guides were marked Legacy in July 2025 and removed in January 2026.

New or refreshed documentation lands most weeks. Feeds are published at /en/feed.xml, /en/atom.xml and /en/feed.json.

Where to go next

Note

Want to see what has already shipped? The What’s New page lists it with dates. Have a request or a correction? Submit feedback — subscriber requests shape the next sprint.