Every AI Model Coming in 2026–2027
Every AI prediction is wrong. The only question is whether it is wrong by months or decades. But predictions are the only map we have — not because they are accurate, but because they reveal what the people building these systems actually believe about what they are building.
Here is every major AI model release that is confirmed, credibly rumored, or predicted by prediction markets for the rest of 2026 and into 2027. Dates come from official announcements, primary-source reporting, and Polymarket/Metaculus prediction markets. Where no official date exists, the source is stated.
Confirmed: what has already shipped in September 2026
| Model | Lab | Date | Price (input/output per M) | Notes |
|---|---|---|---|---|
| GPT-6 Astra | OpenAI | Sep 3 | Not published | First “Critical” cybersecurity model. Limited release; broader availability expected at DevDay Sep 29 |
| Claude Fable 5.1 | Anthropic | Sep 1 | $10 / $50 | Generally available. Same weights as Mythos 5.1 but with standard safeguards |
| Claude Mythos 5.1 | Anthropic | Sep 1 | Restricted | Identical to Fable 5.1 but with reduced safeguards for vetted cyber defenders and life scientists |
| Gemini 3.8 Flash | Sep 2 | Not published | Flash-tier successor to Gemini 3.7 Flash | |
| Gemini 3.8 Flash Cyber | Sep 2 | Restricted | Restricted-access cybersecurity variant through Fairwind program | |
| Muse Spark 1.3 | Meta | Sep 2 | Not published | Latest in Meta’s closed-weight Muse line |
| K2 Horizon | IFM | Sep 3 | Free | Open weights, training code, data, checkpoints, and logs |
| Qwen3.8-Flash-Next | Alibaba | Sep 4 | Free | Architecture preview for Qwen4 |
Coming in the next 30 days (September–October 2026)
These dates come from prediction markets (Polymarket, Metaculus) and credible reporting. Treat them as educated estimates, not confirmed announcements.
| Model | Lab | Median date | Source | Confidence |
|---|---|---|---|---|
| Next GPT Luna (5.7+) | OpenAI | Sep 21 | Polymarket | 98% by Dec 31 |
| Next GPT Sol (5.7+) | OpenAI | Sep 22 | Polymarket | 98% by Dec 31 |
| Next Claude Opus | Anthropic | Sep 24 | Polymarket | 99% by Dec 31 |
| Next GPT Terra (5.7+) | OpenAI | Sep 24 | Polymarket | 87% by Nov 30 |
| Next Claude Sonnet | Anthropic | Oct 13 | Polymarket | 96% by Dec 31 |
| Gemini 4.0 | Oct 21 | Polymarket | 88% by Nov 30 | |
| Next Google Gemini Pro | Oct 23 | Polymarket | 64% by Oct 31 | |
| Next Claude Haiku | Anthropic | Oct 25 | Polymarket | 90% by Dec 31 |
| Next Fable (5.2+) | Anthropic | Oct 31 | Polymarket | 88% by Dec 31 |
| GPT Astra 6.1+ | OpenAI | Oct 31 | Polymarket | 88% by Dec 31 |
OpenAI DevDay — September 29. OpenAI’s annual developer conference. GPT-6 Astra broader availability is expected. No official agenda published yet, but historical pattern suggests major API announcements and new model previews.
OpenAI IPO. Altman told Fortune on September 12: “I would say not 2026.” The CFO told staff in August that OpenAI will be public in 2027, or sooner if the business continues to inflect. A confidential S-1 was filed June 8 targeting September, but the public prospectus has not appeared on EDGAR.
Q4 2026 (October–December)
| Model | Lab | Expected window | Source |
|---|---|---|---|
| Gemini 4.0 | Oct–Nov 2026 | Polymarket median Oct 21. Yearly cadence + Hassabis Jan 2026 statement | |
| Next GPT Image Model (2.1+) | OpenAI | By Dec 31 | Polymarket. 50% by Jan 31, 2027 |
| Grok 5 | xAI | Q3–Q4 2026 | Musk indicated early-to-mid September. Polymarket ~33% by Jun 30 (missed). Pre-training reportedly complete with SpaceX engineering data |
| GPT-6 Astra broader GA | OpenAI | Sep–Oct 2026 | Limited release Sep 3; broader availability expected at DevDay |
| DeepSeek V4 GA | DeepSeek | Q3 2026 | Preview released; general availability expected based on typical preview-to-GA cadence |
2027 (what prediction markets say)
| Model | Lab | Median date | Source |
|---|---|---|---|
| Claude 6 | Anthropic | May 24, 2027 | Polymarket. 86% by Dec 31, 2027 |
| GPT-7 | OpenAI | Jul 23, 2027 | Polymarket. 78% by Dec 31, 2027 |
These are prediction-market medians, not official announcements. Neither Anthropic nor OpenAI has confirmed these model names or dates. The estimates are based on historical release cadence and should be treated as rough guides.
The pricing landscape
The summer of 2026 saw a dramatic price war. Here is where things stand:
| Model | Input per M | Output per M | Open weights? |
|---|---|---|---|
| GPT-6 Astra | Not published | Not published | No |
| Claude Fable 5.1 | $10 | $50 | No |
| Claude Opus 5 | $5 | $25 | No |
| Claude Sonnet 5 | $3 | $15 | No |
| GPT-5.6 Sol | $5 | $30 | No |
| GPT-5.6 Terra | $2.50 | $15 | No |
| GPT-5.6 Luna | $0.20 | $1.20 | No |
| Gemini 3.7 Flash | $0.75 | $3.75 | No |
| DeepSeek V4 Flash | $0.14 | $0.28 | Yes (MIT) |
| DeepSeek V4 Pro | $0.435 | $0.87 | Yes (MIT) |
| GLM-5.3-Flash | $0.15 | $0.50 | Yes (MIT) |
| K2 Horizon | Free | Free | Yes (full) |
| Qwen3.8-Flash-Next | Free | Free | Yes |
The open-weight models are not at frontier capability, but they are closing the gap. DeepSeek V4 Flash at $0.14 per million input tokens is 70x cheaper than Claude Fable 5.1. For many production use cases, the price-performance ratio of open-weight models now exceeds that of closed frontier models.
The open-weight landscape
Four open-weight releases this summer, four different licenses:
| Model | License | Parameters | Min hardware |
|---|---|---|---|
| K2 Horizon | Full open (weights, code, data, checkpoints) | 0.9B–375B | Varies |
| GLM-5.3-Flash | MIT | 320B/18B | 4x H200 or 8x H100 |
| DeepSeek V4 Flash | MIT | Not disclosed | Not disclosed |
| Kimi K3 | Modified MIT | 2.8T/104B MoE | 8x H100 |
| Qwen3.8-Max | Custom | Not disclosed | Not disclosed |
| Gemma 4 | Apache 2.0 | Not disclosed | Runs on everyday devices |
Before building on any open-weight model, check the license file. “Modified MIT” does not necessarily carry the same semantics as standard MIT. Qwen3.8-Max uses custom terms despite being described as “open.”
What this means for builders
Build for model portability. The model you use today will not be the best model in six months. Use abstraction layers — Anthropic-compatible gateways, model-agnostic APIs, runtime model discovery — so you can swap models without rewriting your application.
Do not wait for the next model. GPT-6 Astra, Claude Fable 5.1, and Gemini 3.8 Flash are all production-ready today. The binding constraint on your product is almost never the model — it is the evaluation harness, the prompt engineering, and the integration work. That work transfers across model generations.
Watch the pricing, not the benchmarks. GPT-5.6 Luna at $0.20 per million input tokens changed what is economically feasible. DeepSeek V4 Flash at $0.14 made self-hosting competitive with API pricing for many workloads. The cost curve is falling faster than the capability curve is rising.
Expect tokenizer changes. Claude Sonnet 5’s tokenizer change added 10–35% more tokens on code. When a new model ships, recompute any context budgets pinned to exact token counts. Do not assume token counts are stable across model versions.
Sources
Prediction markets
- Polymarket AI model release timelines (accessed September 16, 2026)
- Metaculus AI forecasting (accessed September 16, 2026)
Official announcements
- OpenAI, “GPT-6 Astra,” September 3, 2026
- Anthropic, “Claude Fable 5.1 and Mythos 5.1,” September 1, 2026
- Google, “Gemini 3.8 Live,” September 15, 2026
- Meta, “Muse Spark 1.3,” September 2, 2026
Reporting
- AIToolsRecap, “AI Model Release Tracker,” September 5, 2026
- Capital & Compute, “New & Upcoming AI Models (2026)”
- Fazm, “Upcoming LLM releases for the rest of 2026,” May 26, 2026
- Fortune, Altman IPO statement, September 12, 2026
Previously on Father of AI
- Why Every AI CEO Is Saying ‘Slow Down’
- The September 2026 AI Model War
- AI in 2027: What Every Prediction Gets Wrong
- Father of AI
Frequently asked questions
What AI models are coming in the next 30 days?
Based on prediction markets and credible reporting, the next 30 days include: Next Claude Opus (median September 24), Next GPT Luna 5.7+ (median September 21), Next GPT Sol 5.7+ (median September 22), Next GPT Terra 5.7+ (median September 24), Next Claude Sonnet (median October 13), and Gemini 4.0 (median October 21). OpenAI’s DevDay on September 29 is expected to feature GPT-6 Astra broader availability. These dates come from prediction markets (Polymarket, Metaculus) and should be treated as educated estimates, not confirmed announcements.
When will GPT-7 be released?
Prediction markets currently place the median release date for GPT-7 at July 23, 2027. OpenAI has not officially announced GPT-7. The estimate is based on OpenAI’s historical release cadence: GPT-5.5 shipped April 2026, GPT-5.6 shipped July 2026, and GPT-6 Astra shipped September 2026. If the cadence continues, GPT-6.x point releases through early 2027 are likely before a GPT-7 jump. The combined distribution reaches 78% by December 31, 2027.
When will Claude 6 be released?
Prediction markets currently place the median release date for Claude 6 at May 24, 2027. Anthropic has not officially announced Claude 6. The estimate is based on Anthropic’s rapid release cadence: Claude Fable 5 shipped June 2026, Claude Opus 5 shipped July 2026, Claude Fable 5.1 shipped September 2026. Anthropic does not pre-announce models, so treat this date with skepticism until a model ID appears in the API.
When will Gemini 4.0 be released?
Prediction markets currently place the median release date for Gemini 4.0 at October 21, 2026. Google’s yearly cadence (Gemini 1.0 late 2023, 2.0 late 2024, 3.0 late 2025) supports a Q4 2026 window. Demis Hassabis stated in January 2026 that Gemini 4 is the team’s focus for this year. The combined distribution reaches 88% by November 30, 2026.
What is the difference between GPT-6 Astra and GPT-7?
GPT-6 Astra, released September 3, 2026, is OpenAI’s current flagship model. It is the first model to trigger the “Critical” cybersecurity threshold in OpenAI’s Preparedness Framework, meaning it can independently discover zero-day vulnerabilities. GPT-7 has not been officially announced. Based on historical patterns (GPT-5 to GPT-6 took approximately 18 months), GPT-7 would represent a major generational jump with significantly improved reasoning, multimodal capabilities, and potentially recursive self-improvement abilities.
Should I wait for the next model before building?
No. The best model to build with is the one available today. Models ship on unpredictable schedules, and waiting for the “next one” means waiting forever. GPT-6 Astra, Claude Fable 5.1, and Gemini 3.8 Flash are all production-ready today. Build for model portability by using abstraction layers (Anthropic-compatible gateways, model-agnostic APIs) so you can swap models when new ones ship. The binding constraint on your product is almost never the model — it is the evaluation harness, the prompt engineering, and the integration work.