AI model release watch: what has shipped and what is still unconfirmed in 2026

The short answer: Of the five most-discussed upcoming models from our August edition, three have now shipped: Claude Fable 5.1 (1 September 2026), GPT-6 Astra (announced 3 September, generally available 4 September) and Grok 4.7 (21 September). Two still have no general release: Google announced Gemini 4 Argon on 30 September but it is not on sale yet, and Z.ai has never announced GLM-5.5. On 22 September three more arrived: Claude Opus 5.5 from Anthropic, and GPT-6 Sol and GPT-6 Luna from OpenAI, a week before DevDay.

Update, 1 October 2026: since our last edition, Claude Sonnet 5.5 shipped on 28 September and GPT-6.1 Sol on 29 September, at OpenAI's DevDay, where OpenAI also announced dots, always-on agents for ChatGPT Pro and Business Premium. Google announced Gemini 4 Argon on 30 September in a limited release to vetted cyber defenders; you cannot buy it yet. Claude Haiku 5.5 is announced, not released.

Correction, 23 September 2026: the 22 September version of this page called GPT-6 Sol a rumour. OpenAI released it, with GPT-6 Luna, the same day. The GPT-6 section below is rewritten with sources.

Updated 1 October 2026 (first compiled 14 August 2026). This page tracks announcements and releases, 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 We do not publish expected benchmark scores, expected prices, or expected release dates for models that do not have them.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

Status at a glance (1 October 2026)

Anthropic Claude Fable 5.1 — Shipped 1 September 2026

In August this was a rumour: no launch post, no model ID, no pricing entry. On 1 September 2026 Anthropic released it, available the same day on the Claude API, AWS, Google Cloud and Azure.

• Price: $10 / $50 per million input/output tokens, unchanged from Fable 5. Cache reads were cut 75% to $0.25 per million.
• Anthropic-reported benchmarks: 55.8% on Terminal-Bench 4.0 and 73.4% on CursorBench 3.2.0. These are Anthropic's own figures.
• Claude Mythos 5.1, the same underlying model with lighter cyber safeguards, was announced alongside it. It is gated behind verification programmes and limited to a set of US organisations, so most buyers cannot get it.

The July rumour got the price right (unchanged $10/$50) and the timing roughly right (August was predicted, it landed on 1 September).

What it means for a business: the per-token rate card did not move, so the real change is cheaper cached input. That helps only if your workload re-sends the same context repeatedly. Details on our model tracker.

OpenAI GPT-6 Astra — Shipped 4 September 2026

OpenAI first disclosed Astra on 1 August 2026, inside a post about mathematics results, and then paused some internal work on it after an evaluation found large advances in agentic coding and cybersecurity. It shipped under the name GPT-6 Astra: first to enterprises in OpenAI's Trusted Access Program on 3 September 2026, then generally available to the API and the paid ChatGPT plans from 4 September.

• Price: $10 / $50 per million input/output tokens, the same rate card as Claude Fable 5.1. Prompts over 272K input tokens bill at 2x input and 1.5x output for the whole request.
• Context: 922,000 input tokens and 128,000 output tokens on the API model card.
• OpenAI-reported benchmarks: 72.6% on OSWorld 2.0 (computer use) and 57.7% on Terminal-Bench 4.0. These are OpenAI's own figures.
• Its advanced cybersecurity capabilities remain gated to a tester group.

What it means for a business: it is built to operate a computer through long chains of clicks and forms. If your work is drafting and answering, it is an expensive way to do it.

xAI Grok 4.7 — Shipped 21 September 2026

In August, Grok 4.7 existed only in Elon Musk's public statements. xAI released it on 21 September 2026, on the xAI API, OpenRouter, Cursor and Grok Build.

• Price: $2 / $6 per million input/output tokens for prompts under 200K tokens, and $4 / $12 above that. That is the same rate card as Grok 4.6.
• Context: 500,000 tokens, per xAI's developer release notes. Reasoning effort can be set to low, medium, high or xhigh.
• xAI-reported benchmark: 38.0% on Terminal-Bench 4.0, up from Grok 4.6's 20.3%.
• Independent score: Artificial Analysis puts it at 46 on its Intelligence Index at xhigh effort, against 53 for both Claude Fable 5.1 and GPT-6 Astra on the same scale. It also flags Grok 4.7 as very verbose, so the cost per task is higher than the rate card suggests.

Correction to our August edition: we reported Musk's description of Grok 4.7 as a "2.1-trillion-parameter" model. xAI has not published a parameter count for Grok 4.7, so we no longer repeat that figure.

What it means for a business: if you already use Grok 4.6, moving to 4.7 is a one-line model change at the same price. An independent index still ranks it well behind the two $10/$50 flagships.

Google Gemini 4 Argon — Announced 30 September 2026, not on sale yet

Google stated on 21 July 2026 that it had "already started our most ambitious pre-training run yet, for Gemini 4," and Sundar Pichai described it as a significantly larger frontier model. On 30 September 2026 Google announced Gemini 4 Argon, a frontier model for long-horizon software engineering, enterprise knowledge work such as legal and finance, and cyber defence.

• Availability: not generally available. It is rolling out only to vetted cyber defenders through Google's Fairwind Program, while Google goes through the US government's voluntary pre-release access process. Google says the broad release comes "as soon as possible", starting with paid API customers and Google AI Ultra subscribers, with no date.
• Announced price: an introductory $2 / $10 per million input/output tokens, with cached input 95% off, then $4 / $20 after the introductory period.
• Output limit: 1M tokens, up from 64K. Google's post does not state the input context window.
• Google-reported benchmarks: 77.9% on DeepSWE v1.1, which Google calls a new state of the art, and 51.3% on AutomationBench, ranked first. These are Google's own figures.
• Independent score: Artificial Analysis puts it at 53 on its Intelligence Index at high effort, #8 of 223, behind Claude Opus 5.5 at 58.

What Google did ship to everyone in September was Gemini 3.8 Flash (2 September), its third Flash release in six weeks.

What it means for a business: Argon looks strong on paper, but you cannot buy it. If you are on Gemini today, Gemini 3.8 Flash is what you can actually use; re-test when Argon reaches the Gemini API or Google AI Ultra.

Z.ai GLM-5.5 — Reported, still not announced

Z.ai has still not announced GLM-5.5. There is no model card, no benchmarks, no pricing and no confirmed name. The August launch window came from a JPMorgan research note reported by Reuters and CGTN on 30 June 2026, and that window has now passed without a release.

What Z.ai shipped instead: GLM-5.3 on 14 August, built on GLM-5.2's base model, and GLM-5.3-Flash on 26 August, an open-weight model that Z.ai reports as a large gain over GLM-5.2 on coding benchmarks (vendor-reported).

What it means for a business: an analyst note is not a roadmap. Evaluate the GLM-5.3 models, which are published and priced.

Anthropic Claude Opus 5.5 — Shipped 22 September 2026

Anthropic released Opus 5.5 on 22 September 2026 as the successor to Opus 5, on the Claude API as claude-opus-5-5 and on AWS, Google Cloud and Azure.

• Price: $4 / $20 per million input/output tokens, down from Opus 5's $5 / $25. Cache reads $0.20 per million.
• Context: 1M input tokens and 128K output tokens, per Anthropic's models overview.
• Anthropic-reported benchmarks: 66.4% on Terminal-Bench 4.0 against Fable 5.1's 55.8%, and 57.8% on CursorBench 4.0.
• Independent score: Artificial Analysis puts it at 58 on its Intelligence Index, first of 212 models, ahead of Fable 5.1 and GPT-6 Astra at 53.
• Anthropic now lists Opus 5, Fable 5 and Opus 4.8 as legacy models. They are still available, not retired. Sonnet 5.5 followed on 28 September (see Claude Sonnet 5.5 vs Opus 5.5 vs GPT-6 Sol); Haiku 5.5 is still promised "in the coming weeks".

What it means for a business: a cheaper Claude that an independent index ranks above both $10/$50 flagships. If you pay for Opus 5 or Fable 5.1, test it first.

OpenAI GPT-6 Sol and GPT-6 Luna — Shipped 22 September 2026

In our previous edition GPT-6 Sol was a rumour pointed at DevDay. OpenAI released GPT-6 Sol and GPT-6 Luna on 22 September 2026, on the API as gpt-6-sol and gpt-6-luna, and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Free and Go users can use Luna in the desktop app. No GPT-6 Terra was released.

• Price: Sol $2 / $10 per million input/output tokens, Luna $0.10 / $0.50. Both are half of GPT-5.6 Sol and Luna.
• Long prompts cost more: a prompt over 272K input tokens is billed at 2x input and 1.5x output for the whole request, which is $4 / $15 for Sol and $0.20 / $0.75 for Luna.
• Context: 1,050,000 input tokens and 128,000 output tokens for both.
• OpenAI-reported benchmarks: Sol 33.2% on AutomationBench, above Claude Opus 5's 26.9%; Luna 66.6% on DeepSWE v1.1.
• GPT-6 Astra remains OpenAI's flagship.

Do not confuse them with GPT-5.6 Sol and Luna, released 9 July 2026 and still on sale. GPT-5.6 Sol costs a promotional $4 / $20 until at least 21 November 2026.

What it means for a business: Luna now has one of the lowest list prices from a major lab, below DeepSeek V4.1 Flash's off-peak rate. For long prompts, DeepSeek V4.1 Flash off-peak is still cheaper. Our three-way comparison covers which to use for what.

OpenAI GPT-6.1 Sol — Shipped 29 September 2026

A week after GPT-6 Sol, OpenAI released GPT-6.1 Sol at DevDay on 29 September 2026, on the API as gpt-6.1-sol and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. It is not yet in regular ChatGPT chat.

• Price: $2 / $10 per million input/output tokens, the same as GPT-6 Sol. Cached input is $0.10, half of GPT-6 Sol's. The 272K rule still applies: a longer prompt bills the whole request at 2x input and 1.5x output.
• Context: 1,050,000 input tokens and 128,000 output tokens.
• OpenAI-reported benchmarks: matches GPT-6 Astra on DeepSWE v1.1 at roughly one-fifth of the cost, and scores 2.2 points above Claude Opus 5.5 on AutomationBench at medium effort, at roughly a third of the cost. These are OpenAI's own figures.
• Independent score: Artificial Analysis puts it at 52 on its Intelligence Index, up from GPT-6 Sol's 48, at about $0.72 per index task.
• A faster GPT-6.1 Sol Ultrafast for Codex is promised "in the coming days".

At the same DevDay OpenAI announced dots, always-on agents powered by GPT-6 Astra, rolling out to ChatGPT Pro users outside the EEA, Switzerland and the UK and to Business Premium users. They are not on Free, Go or Plus.

What it means for a business: if you call gpt-6-sol, switching to gpt-6.1-sol is a one-line change at the same list price. Our best AI model for coding guide compares it with Claude Opus 5.5 and Sonnet 5.5.

The pattern worth noticing

The models that shipped were pitched around agentic coding and long-running tasks, not chat quality. Two of them gated their strongest cybersecurity capability: Anthropic behind Claude Mythos 5.1's verification programmes, OpenAI behind a tester group for GPT-6 Astra. Launching a model while holding back part of it for safety review is now normal. Gemini 4 Argon goes further: announced, priced, and available only to vetted cyber defenders at launch.

The practical consequence: the gap between announcement and full availability is growing. Astra went from first disclosure to general availability in about five weeks, and parts of it are still not generally available. 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 a 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: Anthropic, Claude Fable and Mythos 5.1 (1 September 2026); OpenAI, GPT-6 Astra and the API model card (September 2026); xAI, Grok 4.7 and developer release notes (21 September 2026); Artificial Analysis, Benchmarking Grok 4.7; Google, Gemini 3.8 Flash (2 September 2026) and its statement on Gemini 4 pre-training (21 July 2026); Z.ai, GLM-5.3 and GLM-5.3-Flash; Reuters/CGTN citing JPMorgan (30 June 2026) on GLM-5.5; OpenAI, Previewing GPT-5.6 Sol; Anthropic, Claude Opus 5.5 and the models overview (22 September 2026); Artificial Analysis, Claude Opus 5.5; OpenAI, Introducing GPT-6 Sol and Luna and API pricing (22 September 2026); OpenAI, Introducing GPT-6.1 Sol, Introducing dots and the DevDay 2026 recap (29 September 2026); Google, Gemini 4 Argon (30 September 2026); Artificial Analysis, GPT-6.1 Sol and Gemini 4 Argon. Checked 1 October 2026.