H2 2026: Six confirmed dates, four rumored launches, one $60B option.

Agentic coding in H2 2026 is not a vague horizon — it has six vendor-confirmed dates already on the calendar, from GitHub Copilot's usage-based billing transition on June 1 to the EU AI Act's Article 73 incident-reporting obligations on August 2, with Sonnet 4 and Opus 4 API retirements and the Gemini CLI sunset stacked in between. This forecast tiers every upcoming event — Confirmed, Rumored, or Speculative — each with a confidence percentage and a primary-source citation, so development and engineering teams can plan budgets, toolchain migrations, and model commitments against what is actually known.

The week ending May 24 already moved the goalposts: Composer 2.5 launched at $0.50/$2.50 per million tokens on May 18, Antigravity 2.0 and Gemini 3.5 Flash hit general availability on May 19, GitHub quietly removed all Gemini models from Copilot Chat on the web on May 20, and Cursor announced it was training a new model on Colossus 2 using 10× more compute — all within a 72-hour stretch. H2 2026 begins with a landscape already in motion.

This guide covers: the six confirmed H2 dates and what each means operationally; the June cliff of four deadlines in 18 days; a Polymarket odds snapshot retrieved May 24, 2026; the Colossus 2 compute calendar (Cursor's new model is one of seven parallel training jobs currently running); the IDE consolidation signals that fired this week; a model-vs-ship-window decision matrix translating each forecast into a buyer action; and the EU AI Act binary with its Digital Omnibus wild card. For the broader model-release context, see our Q3 2026 frontier-model release forecast and the 30 agentic AI predictions for H2 2026.

Every forecast starts with what is actually known. The table below covers only events where a vendor or official body has stated a date in a primary-source document — no rumor-site aggregators, no prediction markets, no single-analyst calls. Confidence percentages reflect whether the date is formally scheduled (100%) or planned-but-not-company-confirmed (~85% where sourced from Reuters / Yahoo Finance reporting rather than SpaceX itself).

For the Sonnet 4 / Opus 4 retirement context, see our Opus 4.7 complete guide for what teams migrate to, and our Opus 4.6 to 4.7 migration playbook for the breaking-changes checklist.

The density of June is unusual: Copilot billing changes June 1, Microsoft Build runs June 2–3, Sonnet 4 and Opus 4 retire June 15, and Gemini CLI deprecates June 18 — four distinct toolchain events in 18 calendar days. Engineering managers who have not mapped these to sprint cycles should do so this week.

The June 1 Copilot billing change is the broadest impact. Every team currently on a Copilot Pro or Pro+ monthly plan will see a different cost model starting June 1. The $0.01 per credit rate is confirmed; what is not yet published is the per-task credit consumption rate for specific operations like Copilot Cloud Agent runs or Spark completions. Budget for variance in June until GitHub publishes that breakdown. The GitHub blog announcement and the GitHub Docs migration guide are the canonical preparation resources.

Microsoft Build on June 2–3 is the single event most likely to generate net-new toolchain announcements — specifically around Copilot's /fleet and autopilot mode and the Foundry Agent Service MCP integration. For the broader enterprise Copilot context, see our Microsoft Copilot Cowork enterprise workflows guide.

The Sonnet 4 / Opus 4 retirement on June 15 is the most operationally dangerous item for API-integrated teams. The distinction matters: Sonnet 4.0 and Opus 4.0 (the May 14, 2025 base versions, identified by the model IDs claude-sonnet-4-0 and claude-opus-4-0) retire. Sonnet 4.6 and Opus 4.7 are NOT retiring — they are the recommended migration targets. Any pipeline that hard-codes the old model IDs will break at 9 AM PT on June 15.

Prediction markets are a useful but noisy signal. Their value here is not precision — it is calibration. A Polymarket market at 80–89% for GPT-5.6 by June 30 does not mean GPT-5.6 ships in the next 37 days; it means sufficiently many market participants with real money on the line believe it probably will. Treating this as a planning input is rational. Treating it as a fact is not. Every figure below was retrieved on May 24, 2026 — odds change daily and the retrieval date is the provenance.

GPT-5.6 carries the highest near-term probability. Its presence in OpenAI's Codex routing logs in mid-May (a canary-test pattern consistent with imminent public release) plus OpenAI's 30–60 day cadence since GPT-5 launched in August 2025 both point toward a June landing. Polymarket's June 15 bucket carries 66% implied probability; the June 8 bucket is at 42%. GPT-5.6 is expected to ship before GPT-6, not alongside it. For the OpenAI model-lineage context, see our GPT-5.5 1M-context complete guide and the OpenAI next-model roadmap analysis.

Claude 5 carries a 66% modal probability for September 30, not June. The April 7, 2026 Project Glasswing “Claude Mythos” preview is a cybersecurity-restricted research preview — not a public Claude 5 release. For the Mythos and Capybara-tier analysis, see our Mythos / Capybara deep dive.

The Gemini 4.0 figure (2–15%) deserves a note: the wide spread reflects market illiquidity, not genuine uncertainty in the range. When a Polymarket market has thin order books, a small trade moves the price significantly. The operative takeaway is: Gemini 4 by June 30 is not a high-probability event on Polymarket. Gemini 3.5 Pro, expected in June 2026 per multi-source coverage, is the more likely near-term Google model event. For the Gemini 3.5 Flash context from I/O 2026, see our Gemini 3.5 Flash benchmarks and API guide.

No competitor publication connects the H2 model-release rumors to the underlying compute supply story. The connection matters because model launches don't happen without GPU-hours, and the GPU-hours are concentrated at a single facility.

Colossus 2 — xAI's AI training cluster — reached gigawatt-scale operation in January 2026, reportedly the first coherent AI training cluster to reach that threshold. The cluster is currently running seven distinct parallel model training jobs simultaneously. Cursor confirmed on April 21, 2026 that its new, larger model is one of those seven jobs, trained using 10× more total compute than its previous models. The other six jobs include Grok 5 (xAI's own flagship). This means Cursor's H2 model and Grok 5 are competing for the same compute window on the same cluster.

The financial backdrop explains why this matters beyond the technical detail. xAI's FY2025 capex was $12.7B; its Q1 2026 capex alone was $7.7B — a quarterly rate that annualizes mechanically to ~$30.8B for 2026 (this is a mechanical extrapolation, not a stated plan). xAI's FY2025 segment burn was $6.4B operating loss on $3.2B revenue. The pressure to monetize Colossus 2's compute by delivering Grok 5 and the Cursor model on a tight H2 timeline is real. The SpaceX S-1 audited stats post has the full segment-economics breakdown from the SEC filing.

Anthropic occupies a separate cluster arrangement: a $1.25B/month compute commitment through May 2029 (~$40–45B total) with a 90-day mutual termination clause. This is Colossus 1, the first-generation cluster. The termination clause is the H2 wild card: if either party triggers it, Anthropic's inference capacity — and with it, the pace of Claude 5 training — shifts significantly. For the Cursor financial and acquisition context, see our Cursor $29B valuation analysis.

The week of May 18–20 was not a random cluster of launches. Five events fired in a 72-hour window and collectively shifted the power dynamics of the agentic IDE market: Cursor launched Composer 2.5 at $0.50/$2.50 per million tokens (10× cheaper than Opus 4.7) on May 18; Antigravity 2.0 and Gemini 3.5 Flash hit GA on May 19; GitHub removed all Gemini models from Copilot Chat on the web on May 20; Google announced the June 18 Gemini CLI sunset on May 20; and Cursor publicly disclosed its Colossus 2 training partnership on May 18. Framing these as separate announcements misses the aggregate signal.

The current paid-seat IDE market is fragmented: Cursor approximately 30%, GitHub Copilot approximately 28%, Claude Code approximately 17–22%, Windsurf approximately 7–9%, with Codex CLI, Aider, Continue, Cody, and Kiro in the long tail. Cursor, Windsurf, Trae, and Antigravity all rest on the VS Code base. As StackSpend's agentic IDE comparison frames it: “Microsoft tightening that base — licensing terms, marketplace access, extension APIs — is the single highest-impact lever on the entire IDE-first segment.” Microsoft Build on June 2–3 is the next public moment where Microsoft signals whether it tightens or loosens that lever.

For the full IDE comparison and market-share data, see our Cursor vs Windsurf vs Antigravity comparison and the Cursor 3 deep dive. The AI coding tool-consolidation forecast covers the specific market-share trajectories in detail.

Most H2 forecast posts stop at prediction — they don't translate forecasts into buyer action. The matrix below maps each currently shipping model to the most relevant H2 displacement risk and the recommended operational response. All Polymarket figures are retrieved May 24, 2026.

The key framing principle: a high Polymarket probability for a new model does not mean the current model becomes useless — it means the surface area you are integrating against changes again in the near term. The teams hurt most by rapid model cycles are those who build tight abstractions around specific model IDs or pricing assumptions rather than treating the model layer as a swappable dependency.

Most H2 coverage presents August 2, 2026 as a hard EU AI Act deadline. It is — and it is also contingent. The binary is: if the EU Digital Omnibus is formally adopted before August 2, the high-risk obligation deadlines shift by 12–16 months. If it is not formally adopted before August 2, the original deadline stands.

The state of play as of May 24, 2026: EU lawmakers reached political agreement on the Omnibus revisions on May 7, 2026. Political agreement is not formal adoption. The formal adoption process — European Parliament vote, Council adoption, publication in the Official Journal — has a typical lead time of several months. That lead time puts formal adoption of the Omnibus after August 2 as the more likely outcome, meaning the current default is that August 2 obligations stand.

The practical implication for technology teams: stand up incident reporting workflows under Article 73 now, as if August 2 is a hard deadline, because it likely is. Track Omnibus formal adoption monthly — if it slips, August 2 is the operative date. Penalties under Article 99 for high-risk violations are up to €15M or 3% of global turnover. For a full compliance checklist keyed to the August 2 obligations, see our EU AI Act compliance checklist by risk tier.

The pattern in H2 2026 is not a clean story of “AI keeps getting better.” It is a story of model-layer fragmentation and compute-layer concentration happening simultaneously. On the fragmentation side: within 18 days of June 1, three of the four dominant IDE toolchains change their billing models, deprecate a CLI, or retire model IDs. The engineering team that has abstracted its model-layer dependencies cleanly — treating model IDs as config, not code — will navigate this in hours. The team that hard-coded Opus 4.0 or Sonnet 4.0 everywhere will spend a sprint on mechanical fixes.

On the concentration side: the most consequential compute cluster in the world is running seven parallel training jobs at a gigawatt of power. Two of those jobs — Cursor's next model and Grok 5 — will compete for the same cluster time and could both land in H2 2026 within weeks of each other. When they do, the current pricing surface (Composer 2.5 at $0.50, Grok 4.3 at $1.25, Opus 4.7 at $5) gets disrupted again. The team that locks to today's price points in annual contracts will be renegotiating in Q4.

Looking forward through December 31: if Polymarket's 82% probability for GPT-6 by that date is correct, every current frontier model — Opus 4.7, GPT-5.5, Gemini 3.5 Flash, Grok 4.3 — becomes a “previous generation” model by Q4 2026. That is not necessarily a problem; previous-generation models often carry the best price-performance ratio for production workloads. But it does mean the “buy the best model available today” strategy has a six-month shelf life, not a twelve-month one.

The single most valuable infrastructure investment for an engineering team entering H2 2026 is not picking the right model — it is building the evaluation harness that makes model-swapping a 30-minute operation rather than a two-week project. For the broader agentic toolchain strategy and platform-shakeout context, see our agent-stack platform-shakeout forecast and our H1 2026 AI coding retrospective for the data on what actually shifted in the first half. Our AI transformation advisory practice helps teams build exactly this kind of model-agnostic evaluation infrastructure.