Claude Fable 5.1 and Opus 5
Fable 5.1 ($10/$50) targets demanding long-horizon agentic work; Opus 5 ($5/$25), Sonnet 5 ($2/$10), and Haiku 4.5 ($1/$5) cover the rest - with 1M-token contexts everywhere except Haiku's 200K.
Model wave report
The September 2026 Model Wave Rewrote the Budget Math for Builders.
OpenAI shipped the GPT-6 Astra generation with 1.05M-token context, Google stabilized Gemini 3.8 Flash for long-horizon agents, DeepSeek cut long-context costs with the 552B V4.1 Flash MoE, and OpenRouter added genuinely free frontier endpoints like Union Alpha and Nex-N2.5-Mini. Here is what changed, what it costs, and what to pick for your project.
Last updated September 17, 2026. Prices and endpoints move weekly; verify official pages before launch.
Releases
Every major September 2026 release, in one table.
| Model | Provider | Released | Context | Pricing | Best for |
|---|---|---|---|---|---|
| GPT-6 Astra (Latest) | OpenAI | Sep 11, 2026 | 1.05M | $10/M in, $50/M out | Long-document reasoning, deep research agents, premium features |
| GPT Sol (Latest) | OpenAI | Sep 11, 2026 | 1.05M | $2/M in, $10/M out (50% off on OpenRouter) | Budget frontier quality for chat and agents |
| Gemini 3.8 Flash | Sep 2026 | 1M-class | Free tier + paid | Long-horizon software engineering, autonomous agents, enterprise workflows | |
| Gemini 3.8 Live | Sep 2026 | Live API | Paid | Low-latency voice agents without reasoning delays | |
| DeepSeek V4.1 Flash | DeepSeek | Sep 10, 2026 | 1.05M | $0.15/M in, $0.60/M out | Cheap long-context agents, coding, terminal and computer-use automation |
| Union Alpha | Stealth (OpenRouter) | Sep 16, 2026 | 262K | $0 | Frontier-level multimodal work with zero budget |
| Nex-N2.5-Mini | Nex AGI | Sep 8, 2026 | 262K | $0 (free variant) | Agentic coding with self-verification, GUI QA, computer use |
| Mercury 2.5 | Inception | Sep 8, 2026 | 260K | $0.04/M in, $0.15/M out | 1,100+ tokens/sec diffusion-LLM inference for voice and search agents |
| GPT Image 2.5 (Flare/Sunburst) | OpenAI | Sep 9, 2026 | 400K | $8/M in, $30/M out | Speed- and precision-oriented image generation and editing |
| Muse Voice Transcribe 1.0 | Meta | Sep 11, 2026 | 10-min audio | $0.18/hour | Push-to-talk, endpointing, speaker-aware transcription |
| Qwen3.8 family | Qwen (Alibaba) | Aug-Sep 2026 | Model varies | Open weights | Local multimodal chat, GGUF builds, fine-tuning, local coding agents |
| GLM-5.3-Flash | Z.ai | Sep 2026 | Long | Open weights | Multilingual open-model deployments at 321B scale |
Other lineups
Provider lineups that matter this month.
Grok 4.6
500K context, configurable reasoning at $2/$6 per M tokens, plus Imagine image/video APIs and a realtime Voice API - strong for search-grounded agents.
GPT-OSS and Compound
GPT-OSS 120B at ~500 tokens/sec and GPT-OSS 20B at ~1,000 tokens/sec on the free developer plan, with Compound adding built-in web search and code execution.
Wafer-scale speed
$5 free credit and OpenAI-compatible APIs for Qwen 3.8 27B, GPT-5.6 Sol, Kimi K2.6, and Codex-Spark - inference claimed up to 30x faster than GPU clouds.
Medium 3.5 and OCR 4.1
Frontier multimodal Medium 3.5, Apache-2.0 Small 4 and Large 3, compact Ministral 3 models, OCR 4.1 document AI, and Voxtral transcription.
Enterprise retrieval stack
Command A models plus Embed, Rerank, Parse, and Transcribe APIs - a one-vendor RAG stack for regulated teams.
The headline: for the first time, zero-budget builders can reach frontier-level quality. Union Alpha (multimodal, 262K context) and Nex-N2.5-Mini (agentic coding with self-verification) cost exactly $0 on OpenRouter, and Google keeps a real free tier on Gemini 3.8 Flash.
Free paths
How to run a September 2026 model for free.
Union Alpha
A frontier-level multimodal stealth model at $0. Prompts may be retained by the anonymous provider, so never send sensitive data.
Nex-N2.5-Mini
Explores codebases, edits multiple files, runs commands, and re-tests its own work - ideal for autonomous fixes at no cost.
Gemini 3.8 Flash
The strongest free official tier: agent-friendly speed, vision, structured output, and 1M-class context through Google AI Studio.
DeepSeek V4.1 Flash
When free limits run out, $0.15/M input with a 1.05M window is the practical next step for document-heavy projects.
Selection method
Match the model to the workload, not to the hype.
The September 2026 wave is not one model to rule them all. Frontier models like GPT-6 Astra justify their price only on long documents and hard reasoning. Cheap MoE models like DeepSeek V4.1 Flash win on cost-per-agent-loop. Diffusion LLMs like Mercury 2.5 win when latency compounds. Free endpoints win when the budget is zero and the data is not sensitive.
Pick your tier
- $0 experiments: Union Alpha, Nex-N2.5-Mini, Gemini 3.8 Flash free tier
- Cheap long context: DeepSeek V4.1 Flash at $0.15/M input
- Low latency: Mercury 2.5 or Groq-hosted models
- Premium quality: GPT-6 Astra or GPT Sol
- Local and private: Qwen3.8-27B GGUF via Ollama
People also ask
September 2026 model questions.
What is GPT-6 Astra?
GPT-6 Astra is OpenAI's newest frontier generation. The GPT-6 Astra Latest alias always routes to the newest family model with a 1.05M-token context window at $10/M input and $50/M output, while GPT Sol covers the $2/M budget tier.
Which September 2026 models are actually free?
On OpenRouter, Union Alpha and the Nex-N2.5-Mini free variant cost $0. Google also keeps a free tier for Gemini 3.8 Flash in Google AI Studio. Check availability daily because free and stealth endpoints change without notice.
Is DeepSeek V4.1 Flash good enough for production?
For coding agents, terminal automation, and long-context analysis, DeepSeek reports it exceeds V4 Pro on performance, speed, and task completion. Test it on your own tasks, verify the data-retention policy, and keep a fallback provider documented.
What is a diffusion LLM like Mercury 2.5?
A diffusion LLM (dLLM) generates and refines multiple tokens in parallel instead of sequentially. Mercury 2.5 reaches 1,100+ tokens/sec on standard GPUs with tunable reasoning levels and schema-aligned JSON output.