AI MODEL RELEASE TRACKER — LAST UPDATED 5 SEPTEMBER 2026
● Most recent release: GPT-6 Astra (3 September, OpenAI) — no API pricing published at launch
● Only confirmed date this month: 14 September — Claude Code weekly limits settle 17% below current levels
● Targeted, not dated: OpenAI listing (no public S-1 on EDGAR) · Grok 4.7 (no model ID published)
● Fully open in 2026: K2 Horizon is the only family released with weights, training code, training data, checkpoints and logs
● Licence warning: four open-weight releases this summer, four different licences, none Apache
Released — Most Recent First
| Model | Lab | Date | Price input/M | Open weights |
|---|
| K2 Horizon (6 models) | Institute of Foundation Models | Sep 3 | Free | ✓ Plus training data |
| Qwen3.8-Flash-Next | Alibaba | Sep 4 | Free | ✓ Check licence |
| GPT-6 Astra | OpenAI | Sep 3 | Not published | ✗ Closed |
| Claude Fable 5.1 / Mythos 5.1 | Anthropic | Sep 1 | $10 (cache reads $0.25) | ✗ Closed |
| GLM-5.3 | Z.ai | Aug 14 (weights Aug 28) | See Z.ai | ✓ Licence unstated |
| GLM-5.3-Flash | Z.ai | Aug 26 | $0.15 | ✓ MIT |
| Qwen3.8-Max | Alibaba | Aug 2026 | See Alibaba | ✓ Custom licence |
| Claude Sonnet 5 (repriced) | Anthropic | Aug 31 | $3 (was $2) | ✗ Closed |
| Grok 4.6 | xAI | Aug 2026 | $2.00 | ✗ Closed |
| DeepSeek V4 Flash 0731 | DeepSeek | Jul 31 | $0.14 | ✓ MIT |
| Claude Opus 5 | Anthropic | Jul 24 | $5.00 | ✗ Closed |
| Gemini 3.6 Flash | Google | Jul 21 | $1.50 | ✗ Closed |
| Kimi K3 | Moonshot AI | Jul 16 (weights Jul 27) | $3.00 | ✓ Modified MIT |
| GPT-5.6 Luna / Terra / Sol | OpenAI | Jul 9 | $0.20 / $2.50 / $4 promo | ✗ Closed |
| Laguna S 2.1 | Poolside | Jul 2 | $0.10 | ✓ OpenMDW |
| Claude Fable 5 / Sonnet 5 | Anthropic | Jun 9 / Jun 30 | $10 / $3 | ✗ Closed |
| Gemma 4 | Google | Jun 2026 | Free | ✓ Apache 2.0 |
| Llama 4 Scout / Maverick | Meta | Apr 2026 | Free | ✓ Llama 4 Licence |
THE ASTRA STORY RESOLVED
OpenAI paused Astra on 8 August after evaluations could not rule out a Critical cybersecurity capability. The finding held. On 3 September it shipped anyway — with the cyber capability gated behind a vetted application programme called Daybreak while the rest of the model reaches ordinary subscribers.
What actually shipped →
Confirmed Dates Ahead
14 September 2026 — Claude Code weekly limits. Settle at 25% above the pre-May baseline, which Anthropic itself states is a 17% reduction against current levels. Indexed: 100 before May, 150 today, 125 from the 14th. The arithmetic →
1 October 2026 — OpenAI vs Apple hearing. Apple filed a trade secrets suit; OpenAI responded with a motion to dismiss.
24 October 2026 — deepseek-chat and deepseek-reasoner deprecated. Migrate to deepseek-v4-flash.
12 November 2026 — OpenAI models leave Cursor. Following the SpaceX acquisition. Anthropic has not said whether Claude follows. Full story →
Around 21 November 2026 — GPT-5.6 Sol promotional pricing ends. $4/$20 returns to $5/$30. Price any migration at the standard rate.
Targeted, But Not Dated
Widely reported as September items. Neither has a date, and the distinction matters if you are planning around them.
OpenAI public listing
Ruled out for 2026 on 12 September — Altman told Fortune "I would say not 2026" and declined to confirm 2027. Full story → Confidential S-1 filed 8 June targeting September. The public prospectus has still not appeared on EDGAR, and registration must be public at least 15 days before a roadshow — so each day without a filing narrows the window. The CFO separately told staff in August that OpenAI will be public in 2027, or sooner if the business continues to inflect.
Grok 4.7
Musk indicated an early-to-mid September window. Pre-training reported complete with supplemental training on SpaceX engineering data. No model ID at docs.x.ai, no published price, no benchmark card. Do not plan a migration around it.
DeepSeek V4-Pro general availability
1.6T flagship, confirmed as coming soon in the July changelog. No date published. Preview available via API.
Qwen4 — architecture previewed, model not announced
Alibaba released Qwen3.8-Flash-Next on 4 September specifically to preview the Qwen4 architecture. That usually means a flagship follows within months rather than weeks. What the preview shows →
Anthropic public listing
Reported as targeting October after a confidential filing on 1 June. Worth tracking, since preparation is visible before a date is.
Free and Open-Weight Models — Read the Licence First
OPEN WEIGHTS AND OPEN SOURCE HAVE DRIFTED APART
Four open-weight releases this summer, four different licences, and not one was Apache. Qwen 3.8-Max under custom terms, Kimi K3 under a modified MIT, GLM-5.3 flagship terms still unstated.
What to check before you build on any of them →
K2 Horizon (3 September): Six models, 0.9B to 375B. The only 2026 release published with weights, training code, training data, checkpoints and logs. Not a frontier capability claim — a verifiability one. Why the training data matters →
GLM-5.3-Flash (MIT, 26 August): 320B/18B, 1M context, $0.15/$0.50. Genuinely permissive, unlike the flagship. FP8 checkpoint around 306 GiB — needs 4x H200 or 8x H100 minimum.
DeepSeek V4 Flash 0731 (MIT, 31 July): $0.14/M or self-hosted. Terminal-Bench 82.7%. Weights at deepseek-ai/DeepSeek-V4-Flash-0731. Full review →
Kimi K3 (Modified MIT, 27 July): 2.8T/104B MoE, SWE Marathon #1. 8x H100 minimum. Modified means the standard MIT semantics do not carry over — read the file.
Laguna S 2.1 (OpenMDW, 2 July): 118B/8B MoE. $0.10/M or self-hosted on a desktop DGX Spark. Terminal-Bench 70.2%. Full review →
Qwen3.8-Flash-Next (4 September): Architecture preview for Qwen4. Reported 125B total, ~6B active, plus a 51B component designed for system RAM rather than GPU memory. Check the licence file — Qwen 3.8-Max used custom terms.
Already Happened — 31 August
Claude Sonnet 5 repriced from $2/$10 to $3/$15 per million, with a tokenizer change adding 10 to 35% more tokens on code. A coding workload can sit 65 to 100% above where it did in August.
GPT-5.4 and GPT-5.4 mini left Codex for ChatGPT sign-in users. Both remain available via API key — same tool, different behaviour depending on how you authenticated.
kimi-k2.5 and moonshot-v1 sunset. Anything still pointing at those model IDs is failing now. Migrate to kimi-k3.
How We Track Releases
Updated manually as part of our daily AI news coverage. We source from official lab announcements, API changelogs and verified third-party trackers. A model does not enter the released table until pricing, model string and availability are confirmed from the official source. Anything without a date sits in Targeted, But Not Dated — because confirmed and reported are different things, and a tracker that blurs them is not worth citing.
Sources: LLM Stats · LLM Gateway timeline · AI Release Tracker · Hugging Face model cards · Related: Every AI deadline in September →