# Custom Engine > Pin one exact model for an agent or a chat, and see what happens when a model retires. - Source: https://www.taskade.com/docs/models/custom-engine - Section: Models - Updated: 2026-10-17 - Index of all pages: https://www.taskade.com/docs/llms.txt Custom Engine pins **one exact model**. Taskade stops choosing for you on that surface. Use it when you are comparing models on the same prompt, or when one model suits a task. For everything else, a [rung](https://www.taskade.com/docs/models#tsk-rungs) ages better. ## Where to find it Open the model menu in the builder chat or on an agent. - In the builder chat, open **More models**. The list starts with **Taskade TSK models**, then shows one group for each provider. While a pinned model runs, that row reads **Current Model**. - In the agent editor, the same list appears under **All Models**. An AI step in an automation does not offer this list. Its model menu shows TSK-Auto and the [rungs](https://www.taskade.com/docs/models#tsk-rungs) under a **Taskade credits** heading. A step that already stores a pinned model keeps it. If that model is retired, or no longer runs on Taskade AI credits, the step runs on TSK-Auto. ## The catalog Taskade offers frontier models from top AI labs: OpenAI GPT, xAI Grok, and open-weight models such as DeepSeek, Qwen, Kimi, and GLM. This table lists the models that run on Taskade AI credits. Each model card in the picker shows a speed bar, an intelligence bar, and a token-usage bar. The table turns the first two bars and the credit rate of each model into bands: - **Speed and Intelligence** follow the bars on the model card. - **Cost** groups the credit rate of each model into three bands. Low is the cheapest band and High is the most expensive band. The picker does not show the credit rate. | Model | Provider | Good at | Speed | Intelligence | Cost | Availability | | --- | --- | --- | --- | --- | --- | --- | | GPT-6.1 Sol | OpenAI | Frontier reasoning for hard problems | Slow | Frontier | Medium | Every paid plan | | GPT-6 Luna | OpenAI | Everyday building | Fast | High | Low | Every paid plan | | GPT-5.6 Sol | OpenAI | Deepest reasoning for hard problems | Slow | Frontier | High | Max and Enterprise | | GPT-5.6 Terra | OpenAI | Fast frontier all-rounder | Fast | Frontier | Medium | Max and Enterprise | | GPT-5.6 Luna | OpenAI | Fast, economical quick tasks | Fast | High | Low | Every paid plan | | Grok 4.7 | xAI | Frontier coding and knowledge work | Medium | Frontier | High | Max and Enterprise | | Grok 4.6 | xAI | Frontier coding and STEM reasoning | Medium | Frontier | High | Max and Enterprise | | Grok Build 0.1 | xAI | Fast agentic app building | Fast | High | Medium | Max and Enterprise | | Kimi K3 | Moonshot | Long-context agentic reasoning | Slow | Frontier | Medium | Max and Enterprise | | Kimi K2.6 | Moonshot | Open-weight generalist | Medium | High | Medium | Max and Enterprise | | Kimi K2.7 Code | Moonshot | Agentic coding | Medium | High | Medium | Max and Enterprise | | Qwen 3.7 Max | Alibaba | Agent foundation flagship | Medium | Frontier | High | Max and Enterprise | | Qwen 3.7 Plus | Alibaba | Cost-effective agentic coding | Fast | High | Medium | Every paid plan | | MiniMax M3 | MiniMax | Fast open-source generalist | Fast | High | Medium | Every paid plan | | GLM 5.3 | Z.AI | Latest GLM for autonomous coding | Slow | Frontier | Medium | Max and Enterprise | | GLM 5.2 | Z.AI | Autonomous coding | Slow | Frontier | Medium | Max and Enterprise | | Nemotron 3 Super | NVIDIA | Efficient mixture-of-experts for agents | Fast | High | Medium | Every paid plan | | DeepSeek V4 Pro | DeepSeek | Frontier reasoning and coding | Fast | High | Low | Every paid plan | Claude models run on your own Anthropic key. See [Claude in Taskade](https://www.taskade.com/docs/models/claude) for the Claude models and the setup. > ℹ️ > **This list moves.** Treat it as accurate on the date in this page's metadata, not as a contract. > > Taskade adds frontier models as they ship, and rotates which ones the picker offers. If you want a model you cannot see, ask us. Enterprise agreements can also open models that are not in the default picker, and you can always run a model on your own provider key. > > A Custom Engine pick **is** stored. Your agent or chat keeps the exact model id you chose, which is why a pin ages and a rung does not. Nothing breaks when a model leaves this list. Taskade keeps a retired model on file so your saved pick still reads back correctly, and it runs that turn on [TSK-Auto](https://www.taskade.com/docs/models#tsk-auto) instead. When a newer version replaces a model, a saved pick runs on the newer version. For example, a saved GPT-6 Sol pick runs on GPT-6.1 Sol, and a saved DeepSeek V4 Flash pick runs on [TSK-Instant](https://www.taskade.com/docs/models#tsk-rungs). ## How cost works The **Cost** column is a band, not a price. Three things make a single per-model number misleading: - **Request size changes the total.** Credits scale with the tokens a step carries. The same model therefore costs more on a large request than on a small one. A few models add a further multiplier above a very large input. - **Rates get re-anchored.** Taskade measures real usage after a model launches and adjusts its rate to match. - **Some rates are introductory.** A model can cost fewer credits at launch than it does later. The model menu shows no credit figure for each model. A rung that costs more shows **Uses more credits** on its card. The usage chip on each message shows the credits that the turn used, and that figure is the one to trust. For how metering works, see [Model credits](https://www.taskade.com/learn/genesis/model-credits). ## Context windows These docs do not list a context-window figure for each model. Some model cards name the figure that the provider publishes. The number a provider advertises describes the raw model, not the budget your step receives. A running step also carries workspace memory, agent knowledge, attachments, and conversation history. The usable remainder therefore differs by surface and by request, and the provider figure overstates what you can pass in. ## If a model is not in your picker The picker rotates, so a model you used last week can be absent today. Nothing breaks when that happens. Taskade routes the turn through [TSK-Auto](https://www.taskade.com/docs/models#tsk-auto) instead, and the work completes on a current model. If you need a specific model that is not listed, ask us and we can enable it. Enterprise agreements open models outside the default picker. A provider step runs any model that your own key can access. This is also why a rung is the better default. A rung follows the catalog forward, and a pin depends on what the picker offers that day. ## Run on your own account To bill a provider account you control, connect it and add a provider step. Taskade supports OpenAI, Anthropic, Google Gemini, Groq, and OpenRouter accounts for this. The step runs on your key, and every step around it keeps working normally. See [Multi-model AI access](https://www.taskade.com/learn/genesis/ai-models) for the setup. ## Next steps - Read [how Taskade picks a model](https://www.taskade.com/docs/models) for Auto and the rungs. - See [Model credits](https://www.taskade.com/learn/genesis/model-credits) for metering.