State of AI Coding Agents — June 2026: Opus 4.8 Lands, Google Goes Closed-Source, and the Money Gets Silly

The 30-second version

• Anthropic shipped Claude Opus 4.8 (May 28) — same base API price as 4.7, a much cheaper fast mode, and Claude Code gained dynamic workflows: orchestrating parallel subagents at migration scale. Three days later Anthropic filed confidentially for IPO after a $65B Series H at a $965B valuation.
• Google is sunsetting the open-source Gemini CLI (announced at I/O, May 19). Its replacement, the Antigravity CLI, is closed-source; Gemini CLI and Code Assist stop serving requests June 18, 2026. Migrate or fork.
• GitHub Copilot swapped its whole pricing model (live June 1): every plan moved to usage-based billing on “AI Credits,” a premium Copilot Max tier appeared, and — remarkably — new sign-ups are paused for most paid tiers.
• OpenAI’s Codex had no new model this window, but Goal Mode and remote computer use went GA (May 21), and a Sites plugin preview means Codex now wants to host your deploys too.

What’s new this month

Claude Opus 4.8 — and Claude Code learns to run fleets

Update, June 9: Anthropic has since shipped Claude Fable 5 — a new tier above Opus at 2× the price. Our launch-day verdict covers what’s real, what’s Mythos, and who should pay.

Update, June 18: Claude Code also gained Artifacts — a session’s work published as a live, shareable page (Team/Enterprise beta), not to be confused with the older Claude.ai Artifacts.

Update, June 27: Anthropic published the month’s biggest data point — a study of ~400,000 Claude Code sessions finding that domain expertise, not a coding background, predicts who succeeds with the agent, and that the gap widens with experience rather than closing.

Claude Opus 4.8 (May 28) is the headline model release of the window. The pricing is the strategic part: the base model holds at $5 / $25 per million tokens — same as Opus 4.7 — while fast mode drops to $10 / $50, which Anthropic bills as 3× cheaper than fast mode was, at ~2.5× the output speed. Frontier capability is no longer the price lever; speed is.

The Claude Code side is the part working engineers will feel. Dynamic workflows let one session script the orchestration of parallel subagents — fan out over a large migration or audit, verify findings adversarially, synthesize — instead of you babysitting each worker. There’s also a new maximum-effort mode (you’ll see it called “ultracode”) for tasks where you want the model to spend tokens like it means it. Benchmarks-wise the announcement leans on agentic evals and reports Terminal-Bench through a public harness; we wrote a whole companion piece this week on why that harness footnote matters more than the headline number.

Honest take: orchestration-as-a-feature is the real shift. The 12-toolkit workflow ecosystem we mapped in our convergence piece existed precisely because users had to hand-build multi-agent pipelines. Some of that scaffolding just moved into the product.

Google retires the open-source Gemini CLI — Antigravity CLI is closed

The biggest strategic move of the window: at I/O on May 19, Google announced the transition from Gemini CLI to Antigravity CLI. The new CLI is Go-based, shares the agent harness with the Antigravity 2.0 desktop app, supports background multi-agent orchestration, and carries over Agent Skills, Hooks, Subagents, and Extensions (rebranded as plugins). It is also closed-source — and the open-source Gemini CLI plus the Code Assist IDE extensions stop serving Pro, Ultra, and free-tier requests on June 18, 2026.

The community noticed. Within a day, the new CLI’s tracker filled with requests for the interop the old one had (the top issue: Agent Client Protocol support, so editors like Zed can drive it). A one-month sunset on a tool that was a flagship open-source project eighteen months ago is a sharp turn.

Honest take: if your team standardized on Gemini CLI, you have two weeks at publication time. The pragmatic options: move to Antigravity CLI and accept the closed model, or treat this as the moment to re-evaluate — the open-source slot in your stack is now Cline’s and OpenClaw’s to claim (both shipped steadily all window). What you shouldn’t do is nothing; June 18 is a hard date.

GitHub Copilot: everyone’s billing just changed

GitHub moved every Copilot plan to usage-based billing (announced May 29, live June 1): monthly AI Credit allowances per plan, overage budgets you set, code review that now also consumes Actions minutes, and a new top-end Copilot Max tier. Most striking: new sign-ups for Student, Pro, Pro+, and Max are paused while the transition lands. A June 2 follow-up wave added a GA Copilot SDK, prompt scheduling and voice input in the CLI, and cloud/local sandboxes in preview.

Honest take: this is the largest install base in the category swapping from flat seats to metered consumption — the same direction Codex (token billing, April) and Claude (rolling usage windows) already went. The flat-rate era of AI coding assistance is functionally over; if your team budgets per-seat, your finance model is now wrong on every major vendor. (We did the AI Credits arithmetic the announcement skipped — base + flex, and a 15× swing by model — and tracked the enterprise spend reckoning it triggered: Uber’s $1,500 cap, Microsoft’s internal Claude Code cut.)

The old guard checked in

Codex (OpenAI). No new model in-window (GPT-5.5 predates it), but a busy changelog: Goal Mode hit GA in CLI 0.133.0 (May 21) — a persistent /goal that survives session breaks and budget resets — alongside remote computer use that keeps driving desktop apps after your Mac locks, including from Codex Mobile. June brought Amazon Bedrock support (June 1) and a Sites plugin preview (June 2): build and host sites and dashboards inside Codex. One sour note: a well-upvoted issue over the Desktop app dropping its visible context-usage indicator — small thing, but context visibility is exactly the wrong place to regress.

Cursor. Four notable ships in the window, per the changelog: Cursor in Jira (May 19 — assign a ticket to @Cursor, get a PR back), 3.5’s shared canvases and upgraded Automations (May 20), 3.6’s auto-review run mode (May 29 — allowlisted calls run instantly, sandboxable ones run sandboxed, everything else goes to a classifier subagent: a genuinely thoughtful middle ground between approve-everything and YOLO), and Organizations for enterprise multi-team management (June 3).

Grok Build (xAI). Last edition’s debutant widened access in late May: the beta now covers all SuperGrok and X Premium Plus subscribers (it launched gated to the $300 Heavy tier), with intro pricing at $99/mo per xAI. Parallel sub-agents running in isolated git worktrees remains the signature trick. Still a beta; the cadence is fast.

The open-source bench. Cline shipped continuously (VS Code 3.85→3.87 plus a maturing standalone CLI; added GPT-5.5, DeepSeek V4, Gemini 3.5 Flash routing). OpenClaw cut two stable releases. Aider stayed quiet — no release since August 2025, increasingly conspicuous. And the MCP spec had no in-window milestone (next release candidate is dated late July).

The money, briefly

Two financings in 48 hours define where this market thinks it’s headed. Anthropic closed a $65B Series H at a $965B post-money valuation (May 28) — the most valuable AI startup — then confidentially filed a draft S-1 on June 1. And Cognition (Devin, Windsurf) closed a reported $1B+ round at a $26B valuation (May 27). Whatever you think of the numbers, the capital is now priced on the assumption that agents — not chat — are the product.

The landscape after the window

Where this matters for your stack

1. On Gemini CLI? Move before June 18. That’s not a deprecation notice, it’s a shutoff date for Pro/Ultra/free requests. Test Antigravity CLI now or pick your open-source fallback deliberately.
2. Re-check your AI spend math. Copilot went metered, Codex is token-billed, Claude is windowed. Flat per-seat budgeting no longer describes any major vendor; set overage caps before the first surprise invoice, not after.
3. If you’re on Claude Max, try the new orchestration before building more scaffolding. Dynamic workflows absorb a chunk of what custom multi-agent setups did. And if you do run parallel fleets: more agents means more memory and transcript files on disk — we keep ours legible with AI Memory Reader (disclosure: we make it), and wrote up how to run long-lived fleets without the memory bleeding together.

---

This is part of an ongoing monthly series — May 2026 edition here. Every dated claim above was verified against the linked primary source (vendor changelog, official blog, or GitHub release) on June 3, 2026; where only secondary reporting exists (the Cognition round), it’s marked as reported.