English · 7 min read
AI model release watch: what is actually confirmed for the rest of 2026
Gemini 4, OpenAI Astra, Grok 4.7, GLM-5.5, Claude Fable 5.1. Which of these labs have actually announced, which are rumour, and what a business should plan around. Updated as each ships.
The short answer: Of the five most-discussed upcoming models, two are officially announced (OpenAI's Astra, Google's Gemini 4), one was announced by the company's founder without published specs (Grok 4.7), and two have no official announcement at all (GLM-5.5, Claude Fable 5.1) despite being written about as if they were dated products. Do not restructure your AI spend around any of them.
Compiled 14 August 2026. This page tracks announcements, not predictions. We do not publish expected benchmark scores, expected prices, or expected release dates for models that do not have them — those numbers do not exist, and inventing them is how a page becomes unciteable. Each entry below is labelled with what the source actually is.
How we classify
| Label | Means |
|---|---|
| Confirmed | The lab itself published or stated it, on the record |
| Announced by principal | A founder or executive stated it publicly, but the lab has published no model card, docs, or pricing |
| Reported | Credible outlets or analysts report it; the lab has not confirmed |
| Rumour | Community, leaks, or social posts only. No institutional source |
A model can be Confirmed to exist and still have nothing you can plan around. That is the situation with most of this list.
OpenAI Astra — Confirmed
OpenAI disclosed Astra on 1 August 2026, describing it as its "next major model." The announcement arrived unusually: inside a post about mathematics, presenting ten claimed solutions to long-standing open problems in maths and theoretical computer science, with manuscripts, Lean certificates, and discovery walkthroughs. The problems spanned high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography, and extremal combinatorics — several untouched for a decade or more.
OpenAI subsequently paused some internal activities involving Astra after an internal evaluation found significant advances in agentic coding and cybersecurity.
Not announced: release date, pricing, context window, API availability, public benchmarks.
What it means for a business: nothing operational yet. A model whose own lab has paused internal work pending safety review is not a procurement decision. The math results are a capability signal, not a product.
Google Gemini 4 — Confirmed, but thin
Google stated on 21 July 2026 that it had "already started our most ambitious pre-training run yet, for Gemini 4." Sundar Pichai described it as a significantly larger frontier model, and said Google intends to move to an almost monthly release cadence with Gemini 4 as the baseline.
Not announced: release date, architecture, context window, pricing, any benchmark.
Note what "pre-training has started" means: it is the beginning of the expensive part, not the end. A late-2026 preview is plausible; it is not scheduled. Prediction markets have priced a 2026 release optimistically, but a market price measures what traders believe, not what Google has said.
What it means for a business: if you are on Gemini today, the stated monthly cadence is the more actionable news than Gemini 4 itself.
xAI Grok 4.7 — Announced by principal
Elon Musk publicly described a rapid cadence: Grok 4.6 followed weeks later by Grok 4.7, characterised as a 2.1-trillion-parameter model, better than 4.6 "in every way" but slightly slower to serve, with better token efficiency. Timing pointed at late August or early September 2026. Grok 4.5 shipped 8 July; Grok 4.6 followed.
Caveat that matters: the parameter count and the capability claims come from the founder, not from a model card, docs page, or evaluation. xAI has shipped on roughly this cadence recently, which makes the timing credible; it does not make the specs verified.
What it means for a business: Grok 4.6 is real, priced, and available now. That is the one to evaluate. Treat 4.7 as likely-soon, not as a spec sheet.
Z.ai GLM-5.5 — Reported, not announced
Z.ai has not announced GLM-5.5. No model card, no benchmarks, no pricing, no confirmed name.
What exists: an August 2026 launch window reported by Reuters and CGTN on 30 June 2026, citing JPMorgan research, and a >1 trillion parameter estimate from that same note and from Chinese AI outlet AIBase (23 June). Community discussion in mid-July described a 1M-token context window carried over from GLM-5.2, open weights, and a focus on long-running coding agents — that portion is unsourced.
Relevant context: Z.ai did just ship GLM-5.3 on 14 August — reusing GLM-5.2's base model, with all gains from post-training, and weights promised two weeks out. A lab that just shipped a post-training-only release, and has not yet published the weights for it, is not obviously two weeks from a new trillion-parameter base model.
What it means for a business: an analyst note is not a roadmap. Evaluate GLM-5.2, which is published, priced, and open-weight.
Anthropic Claude Fable 5.1 — Rumour
Anthropic has not announced Claude Fable 5.1. As of 12 August 2026 there is no launch post, no model card, no API model id, and no pricing-page entry. Anthropic's catalogue and developer release notes list Claude Fable 5.
The circulating claims — that the model is finished internally, due in August, at unchanged $10/$50-per-million pricing, focused on long-horizon reasoning and agent workflows, and timed against OpenAI's next release — trace to two X posts dated 26 July 2026 from industry watchers.
What it means for a business: treat "Fable 5.1" as a name people are using, not a product. Plan on Fable 5.
The pattern worth noticing
Four of these five are pitched around agentic coding and long-horizon autonomy, not chat quality. Two of them — Astra and GLM-5.3 — have had capability gated or staged behind safety and cybersecurity review. That is new. It is now a normal reason for a launch to arrive without weights or an API.
The practical consequence: announcement-to-availability lag is growing. A model being announced tells you less about when you can use it than it did a year ago. Budget against what is shipping and priced today.
What we do not publish
We do not put expected scores in a table. We do not average rumoured prices. We do not give an unannounced model a release date because a prediction market implies one. When any model here publishes real numbers, it gets a real entry on our model tracker, and an independent score gets it onto our benchmark page.
If you want your model choice decided by what is measurably available rather than by launch calendars, a Business AI Plan ($99 one-time, optional $29/month to keep it current) does exactly that. The free 2-minute AI plan finder is the fast first pass.
Sources: OpenAI Astra announcement (1 August 2026) and subsequent pause reporting; Google statement on Gemini 4 pre-training (21 July 2026); Musk public statements on Grok 4.6/4.7 cadence; Reuters/CGTN citing JPMorgan (30 June 2026) and AIBase (23 June 2026) on GLM-5.5; Anthropic model catalogue and release notes as of 12 August 2026.