Dex answers questions about Irish government, politics and law, and a good answer here is mostly a reading job. The model has to hold a long record, such as a bill, a committee transcript or a run of votes, and say only what that record supports. Different models are good at different parts of that, so Dex lets you choose.
This week OpenAI’s GPT 6.1 Sol replaces GPT 6 Sol in the model menu. Here is what it is, where we think it earns its place, and the rest of the line-up around it.
GPT 6.1 Sol
OpenAI released GPT 6.1 Sol on 29 September 2026. It has a 1.05 million token context window and a 128,000 token maximum output, takes text and images, and is priced at $2 per million input tokens and $10 per million output tokens. OpenAI positions it between GPT 6 Astra and GPT 6 Luna, and says it reaches Astra-level capability at about a fifth of the task cost. That is OpenAI’s claim; we haven’t run our own benchmarks.
What matters for Dex is the combination of a very large window and clear, structured writing. Sol is a good pick when a question needs several sources side by side, for example how a bill changed between stages, or what a TD said in a debate compared with how they voted. It is open to anyone who agrees to cloud training, and to Pro.
The specs, side by side
- Claude Opus 5.5 (Anthropic), 1M tokens context. 128K max output, text and image input, June 2026 knowledge cutoff. Released 22 September 2026. Anthropic says it is over 30% faster and about 40% cheaper on typical workloads than Opus 5.
- Claude Sonnet 5.5 (Anthropic), 1M tokens context. 128K max output, text and image input. Released 28 September 2026 at the same price as Sonnet 5, $2 / $10 per million tokens, and over 30% faster.
- GPT 6.1 Sol (OpenAI), 1.05M tokens context. 128K max output, text and image input. $2 / $10 per million tokens in and out.
- Kimi K3 (Moonshot), 1M tokens context. 2.8 trillion parameters, open weights, native vision.
- Qwen 3.8 Max Prime (Alibaba), 1M tokens context. 2.4 trillion parameters, 95 billion active per token, sparse mixture of experts. Prime is the higher-throughput tier.
- Gemini 3.8 Flash (Google), 1M tokens context. 64K max output, adjustable thinking levels. Free for everyone in Dex.
A million tokens is roughly the length of several long novels. In practice it means Dex can read a full bill and its amendments in one go instead of working from excerpts.
Open models, and the Prime tier
Two of the strongest models in Dex are open. Kimi K3 has 2.8 trillion parameters and open weights. Qwen 3.8 Max Prime comes from Alibaba’s 2.4 trillion parameter Qwen 3.8 Max, with 95 billion parameters active for each token. Prime versions are tuned for throughput, so they answer faster than their general counterparts. Qwen 3.8 Max Prime is Pro-only in Dex.
We care about open models because anyone can inspect and host them, which fits a service built on public records. It is also why we are training Irish-language open models of our own.
Claude Opus 5.5 and Sonnet 5.5
Claude Opus 5.5, released on 22 September 2026, is the most capable model in Dex and is Pro-only. It has a 1 million token context window and up to 128,000 tokens of output, and Anthropic says it runs over 30% faster than Opus 5. Claude Sonnet 5.5 followed on 28 September with the same 1 million token window, the same price as Sonnet 5, and published scores within a few points of Opus 5.5 on most benchmarks.
Sonnet 5.5 and Sol are our two favourites right now. Both tend to be careful about what a source does and doesn’t say, which is what we want from a civic assistant.
Shout-out: Gemini 3.8 Flash
Gemini 3.8 Flash is one of the fastest models in Dex, and possibly the fastest on the free plan, with a 1 million token context window. Anyone can use it without signing in, including for long documents.
Why offer a choice
No single model wins every question. A fast model is right for “who is my TD?”, and a larger one is right for comparing two versions of a bill. Whichever you pick, Dex stays civic-only, looks up the official record and links to its sources, so you can check the answer yourself. Compare models by asking the same question twice.
Open Dex or see Dex Pro for the Pro-only models.
— The DáilDex team