For Codewhale, the open-source terminal agent formerly called DeepSeek-TUI, the best API for mixed work is a cost-optimized OpenAI-compatible provider: add [providers.standardcompute] with kind = "openai-compatible" and base_url https://api.stdcmpt.com/v1 to ~/.codewhale/config.toml, and Standard Compute's router sends each step to the cheapest model that can do it (Claude Fable 5.1 or GPT-6 Astra only when it counts), on a plan from $19 a month with a $20 compute budget. A DeepSeek API key billed per token is the cheap route for DeepSeek-only work, OpenRouter is pay-per-use, and Codewhale runs local models through Ollama too. Codewhale itself is free and MIT-licensed.
| Option | Price | What you get |
|---|---|---|
| DeepSeek API | Pay-as-you-go | DeepSeek's own models billed per token with a DeepSeek platform key, the default route in DeepSeek Harness, Codewhale and Reasonix. Usually the cheapest route if DeepSeek models are all you use; check the current per-token and cache-hit prices. Source: api-docs.deepseek.com |
| Standard Compute (us) | From $19/mo | Smart routing sends each request to the cheapest model that can do it, or pin any model by its exact ID. Standard Compute starts at $19/month with a $20 monthly compute budget. Requests stop when that budget is exhausted until renewal or a plan change; there are no automatic overage charges. Requests are not slowed as the budget is used. It is a fixed bill with a finite allowance. |
| OpenRouter | Pay-as-you-go | Choose models and providers through one API. Usage-based pricing can suit light or bursty work; estimate long agent sessions and review current fees and limits. Source: openrouter.ai |
| Local model (Ollama / vLLM) | Hardware + electricity | No hosted per-token bill, but hardware, context capacity, throughput and maintenance still limit what you can run. Test task quality on your own hardware. Source: docs.ollama.com |
Prices and allowances can change. Codewhale's provider tables, key resolution order and config paths come from its configuration docs and README, read on October 15, 2026. Codewhale: configuration · Codewhale repo · DeepSeek API pricing. Spot an error? Tell us.
Export STANDARDCOMPUTE_API_KEY and run codewhale; for one session only, use codewhale --provider standardcompute. Change models with /model, for example to anthropic/claude-sonnet-5.5.
Record your own results with the workload CSV and report template. These are blank templates, not published benchmark results.
Disclosure: this guide is maintained by Standard Compute, one of the options above. Choose based on task quality, supported features and total cost for your workload. Light usage may cost less on a usage-based API; sustained workloads need enough allowance on any subscription.