Solar Pro 4 is the generation where the focus moves from answering to finishing. The API surface stays familiar, so the migration is mostly about what you can now hand over in one task — and about re-checking the assumptions you built around a shorter context window.
What changed
- Agentic completion. Solar Pro 4 is tuned for multi-step jobs that end in a deliverable: reading documents, running terminal tasks, and chaining tool calls. The published scores — Terminal-Bench v2.1 57, τ³-Banking 23, AA-LCR 71 — are each a several-fold step over Solar Pro 3, and the benchmark guide explains what each one measures.
- Reasoning by default. Responses carry the reasoning trace, and reasoning effort is adjustable from real-time interaction to deep analysis.
- Longer work in one session. A 512K context with up to 128K output tokens means several contracts, reports, and data files can stay in a single conversation.
- Evidence discipline. Solar Pro 4 is built to say it cannot verify rather than fill the gap — useful when the deliverable has to survive review.
What stays the same
- The endpoint is OpenAI-compatible: change the endpoint and the model name (
solar-pro4) and your existing code runs. - Pricing stays token-based: $0.30 per 1M input tokens, $0.06 per 1M cached input tokens, $1.20 per 1M output tokens.
- English, Korean, and Japanese are supported for both input and output.
Migration checklist
- Swap the model name. Set
solar-pro4and keep the rest of your request shape; see the API guide for the setup. - Re-check your context budget. Prompts that were trimmed for a smaller window can now carry whole document sets — and output limits up to 128K change how much a single call can return.
- Re-test your tool schemas. Multi-turn tool use is where the generation gap shows; run the same bank of tool-calling conversations you use today.
- Handle reasoning traces. If you parse responses strictly, account for the reasoning trace the response carries by default.
- Re-run your evaluations. Keep the tasks, change the model, and compare completion — not tone.
When Solar Pro 3 is still fine
If your workload is single-turn generation with a short prompt, the newer agentic behavior buys you little. The migration pays off where tasks are long, tools are involved, and completion is the metric.
FAQ
Is the migration a rewrite?
No. The endpoint is OpenAI-compatible, so the change is the endpoint, the model name, and any context or output assumptions in your code.
Does the pricing change between generations?
The published token rates for Solar Pro 4 are on the pricing page, alongside the plan options.
How do I know whether completion actually improved?
Run a fixed set of multi-step tasks and grade the final deliverable, the way the benchmarks do. The benchmark guide describes the task shapes to copy.
