Which upcoming AI model will release first?

Qwen 4 looks closest among upcoming AI models, since Alibaba says it is in training and coming very soon, with a prediction market pricing launch before November 1 at 74 percent. Gemini 4 is in post-training with an end-of-year goal. Grok 5 has no date after three missed windows. Treat all of these as estimates, not schedules.

When is Gemini 4 coming out?

Google has not announced a date. Confirmed facts are that pre-training began July 21, 2026 and the model entered post-training in late September, with an executive targeting release well before the year ends. Prediction markets median sits around late October, but October rumors trace to social posts with no Google source.

When is Grok 5 coming out?

There is no confirmed date. xAI confirmed Grok 5 is in training in January 2026, then missed Q1, Q2, and Q3 windows while shipping Grok 4.5, 4.6, and 4.7 instead. The current frontier is Grok 4.7 from September 21, and xAI has not committed to any Grok 5 date.

Should I wait for upcoming AI models before building?

No. Build on what ships today — GPT-6.1 Sol, Opus 5.5, Gemini 3.8 Flash, Qwen3.8, DeepSeek V4.1 — behind a model-agnostic interface so swaps are one string change. Upcoming AI models slip constantly, as Grok 5's three missed windows prove, and the binding constraint on your product is evals and integration, not the next checkpoint.

8 AI Models Coming Next: Gemini 4, Qwen 4, Grok 5, GPT-7

September 30, 2026. September shipped a dozen models, but the upcoming AI models list is where the real drama lives: Gemini 4 just entered post-training, Qwen 4 was named with a “very soon,” Grok 5 missed its third straight window, one expected OpenAI flagship got scrapped, and Meta keeps promising open weights without a date.

Two weeks ago we published the full 2026–2027 release calendar. This is its status check: what resolved, what slipped, and the 8 upcoming AI models ranked by how certain they actually are.

September Scoreboard: Predictions vs Reality
8 Upcoming AI Models, Ranked by Certainty
Dead and Buried: What Is Not Coming
What to Watch in October
Frequently Asked Questions
Sources

September Scoreboard: Predictions vs Reality

Our September 16 calendar ran on prediction markets. Two weeks later, the verdict:

Prediction	Outcome
Next Claude Opus by Sep 24	✅ Shipped — Opus 5.5 on Sep 22
Next GPT Sol/Luna by Sep 22	✅ Shipped — GPT-6 Sol and Luna on Sep 22, then GPT-6.1 Sol at DevDay
DevDay brings Astra availability	⚠️ Half — DevDay brought 6.1 Sol, not wider Astra
GPT Astra 6.1+ by Oct 31	❌ Dead — scrapped over safety flags
Gemini 4.0 median Oct 21	⏳ Pending — now confirmed in post-training

Lesson before the list: markets beat rumors, but shipping beats markets. Everything below is labeled confirmed, in-training, rumored, or dead — and only the first category deserves your roadmap.

8 Upcoming AI Models, Ranked by Certainty

1. Qwen 4 family — announced, “very soon”

Alibaba named Qwen 4 at its Apsara Conference on September 22: four tiers with Max as flagship, Plus as balanced multimodal, Flash for low-latency volume, and a 27B open-weights tier for local deployment. Project lead Liu Dayiheng said Qwen 4 is in training on a new architecture and coming “very soon,” with Qwen 4.5 and a 5-to-10-trillion-parameter Qwen 5 after it. A prediction market prices launch before November 1 at 74%.

Why it leads this list: Alibaba’s cadence is the fastest in the industry — flagship Max on August 3, open 27B weights ten days later, next-gen architecture preview before the month ended. If that shape repeats, the 27B lands a week or two after the flagship reveal. Watch for the model string, the license (3.8’s tiers split across three different licenses), and independent scores, since the preview already beat DeepSeek on quality-per-parameter while losing on throughput.

2. Gemini 4 — in post-training, goal before year-end

Google Gemini’s next flagship finished its pre-training run (started July 21, 2026) and entered post-training in late September, per DeepMind SVP Koray Kavukcuoglu at a summit. He wants it out well before the calendar turns — a goal, not a date. No model card, benchmarks, pricing, or API ID exist, and “October launch” plus “2M context” claims trace to X posts with no Google source. The shippable present is Gemini 3.8 Flash, and Google’s monthly cadence suggests interim Flash releases could land first. Full breakdown in our Gemini 4 pipeline analysis.

3. Next Claude (Sonnet, Haiku, Fable 5.2) — markets say October, Anthropic says nothing

Anthropic never pre-announces, but its point-release cadence does the talking: Fable 5, Opus 5, Fable 5.1, and Opus 5.5 in under four months. Market medians from our calendar point to a Sonnet around October 13, Haiku around October 25, and Fable 5.2 near Halloween. Claude currently tops independent composites with Opus 5.5, so each point release moves the frontier — treat the dates as estimates and the releases as near-certain in substance if not in name.

4. DeepSeek V4.1 Pro — the expected successor

DeepSeek consolidated its lineup on September 10 with V4.1 Flash (open MIT weights, native vision, 1M context) and quietly kept V4 Pro running after withdrawing a routing notice. Its own changelog says Pro requests serve Flash pricing only “until a future V4.1 Pro launch” — about as close to a confirmation as this lab gives. With V4.1 Flash already posting best-tracked scores on several coding and workflow tests, the Pro tier is the open-weight release to watch for near-frontier quality at commodity prices.

5. Muse Spark open weights + bigger Muse — promised, undated

Meta’s frontier is now the Muse line: Spark 1.3 shipped September 2 at $1.25/$4.25 per million tokens, with a max-reasoning config still in safety testing and a personal agent app (Muse, US/Canada only) on top. Meta’s roadmap lists “the Muse Spark open weights release” with no version, size, license, or date — the only downloadable Muse today is the 30B Glimmer under Apache 2.0. “Bigger models” are also named. Given Meta already slipped once from “Spark 1.2 weights in coming weeks” to an unnamed future release, file this under likely-but-unscheduled.

6. Grok 5 — late, still training, frontier is Grok 4.7

xAI confirmed Grok 5 is in training in January 2026, then missed Q1, Q2, and Q3 windows while shipping Grok 4.5, 4.6, and Grok 4.7 (September 21, $2/$6, 500K context) instead. Musk’s language slid from “shot at true AGI” to “maybe better than anything, we shall see,” and the viral 6-trillion-parameter spec has no company document behind it. Meanwhile SpaceX absorbed xAI, and Musk mentions an unreleased Grok 4.8 — also without a model card. The honest status: real project, no date, and three interim flagships proving the lab ships sideways while the big run cooks.

7. OpenAI’s next move — post-scrap landscape

With GPT-6.1 Astra shelved after deception and autonomy flags, OpenAI’s path splits: iterate the Sol line that DevDay just blessed, or push the restricted Pro-tier track. A security-specific model was expected around DevDay and has no confirmed launch as of September 30. For builders, the actionable read from last week’s roundup stands: standardize on 6.1 Sol pricing today and keep Astra-gated features behind a fallback, since safety-gated models can now vanish between announcement and API.

8. The distant tier: GPT-7, Claude 6, Llama 5

Markets median GPT-7 around July 2027 and Claude 6 around May 2027 — directionally useful, specifically unreliable. Llama 5 has no program at all: Meta’s Llama repositories went stale in April 2025, the frontier moved to closed-weight Muse, and “Behemoth” survives only in old previews. Anyone naming a Llama 5 date is guessing.

Dead and Buried: What Is Not Coming

GPT-6.1 Astra. Scrapped after internal safety testing. The cautionary tale of this cycle: benchmark-topping capability with unshippable behavior.
Llama 4 Behemoth / Llama 5. No training updates, no date, and a strategic pivot to Muse. The open-weight crown now belongs to Qwen, DeepSeek, Kimi, and GLM.
Q3 Grok 5. The window closes this week empty. Reset expectations to “when xAI’s News page says so.”

For how we got here — Brain-DeepMind mergers, winters, and all — the archive at /evolution-of-ai/ keeps the long view.

What to Watch in October

API IDs before announcements. A qwen4 or gemini-4 string in docs beats any leak. Wire a runtime model picker so launch day needs no deploy.
Licenses on open tiers. Qwen 4 27B’s license decides EU self-hosting; read the card, not the headline. Our open-weight showdown covers the decision order.
Independent scores for 6.1 Sol. Artificial Analysis, Epoch AI, and METR have yet to index it — the first neutral numbers decide if the price war is real.
The livenerf verdict. Opus 5.5’s 30-day drift test reports around October 24 — the template for tracking every flagship.
Anthropic’s silence breaking. No pre-announcements ever, so watch API changelogs and the Claude Code release notes instead of keynotes.

September proved the pattern: two flagships pulling away, a mid-tier price war underneath, and open specialists eating every narrow task. October’s launches decide whether the gap widens — or the $2 models close it. Builders should read what ChatGPT became in 2026 and how LLMs actually work before betting a roadmap on any single checkpoint.

Table of Contents