Upstage Solar Pro 3 is a reasoning model whose thinking tokens cost the same as its context tokens. On the Upstage first-party API it is $0.25/$0.25 per Mtok, symmetric, across 131,072 tokens of context. Nobody else in the registry charges the same for a thinking token as for a context token. OpenRouter lists the same model at $0.15/$0.60 per Mtok with a 128,000 token context, a cheaper input and a 4x output premium. Inside Atlas, Solar Pro 3 is the pick when reasoning traces are long and you want the bill to stay flat. The first-party API caps max output at 8,192 tokens.
Why is Upstage Solar Pro 3 priced symmetrically at $0.25 per Mtok?
Upstage Solar Pro 3 charges $0.25 per Mtok in and $0.25 per Mtok out on the first-party API, and that symmetry is the single most interesting fact about the model in 2026. Reasoning models normally punish output at 4x, so a thinking trace normally costs 4x a context token.
Symmetric $0.25/$0.25 per Mtok pricing on the first-party API means a reasoning model whose thinking traces cost exactly what its input costs, which is the opposite of the usual 4x output penalty. For an agent like Atlas, that reshapes where you spend effort. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, and a long plan on Solar Pro 3 is billed at the same rate as the repo context that fed it. Atlas also fans out work to subagents that can run in the foreground or in parallel background sessions, and each of those subagent reasoning traces lands at the same $0.25 per Mtok. The result is that budgeting for Solar Pro 3 is close to linear in total tokens moved, not skewed toward generation.
Upstage API versus OpenRouter for Solar Pro 3: which listing should Atlas use?
Solar Pro 3 has two listings in 2026 that genuinely suit different workloads. The Upstage first-party API is $0.25/$0.25 per Mtok across 131,072 tokens. OpenRouter is $0.15/$0.60 per Mtok across 128,000 tokens, a cheaper input with a 4x output premium.
The Solar Pro 3 routing decision comes down to your output-to-input ratio. The OpenRouter route trades a smaller context (128,000) and a 4x output premium ($0.60) for a cheaper $0.15 input. If your Atlas sessions are read-heavy, pulling a large repo slice and emitting a small diff, the $0.15 input on OpenRouter wins. If reasoning traces are long, which is the normal case for a reasoning model driving a build agent, the symmetric $0.25/$0.25 first-party route wins and it hands you 131,072 tokens instead of 128,000. Atlas lets you switch the active model and provider on the fly with favorites and recents, so you can carry both listings and pick per task rather than committing once in atlas.json.
What is the context window on Upstage Solar Pro 3?
Upstage Solar Pro 3 holds 131,072 tokens on the Upstage API and 128,000 tokens on OpenRouter. Solar Pro 3 doubles the context of Solar Pro 2, 131,072 versus 65,536 tokens, at an identical price per Mtok, so there is no cost reason to stay on the older model.
A 131,072 token window on Solar Pro 3 is enough for a substantial repo slice, and Atlas is built to fill it well rather than fill it fully. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, and Atlas indexes code by AST declarations using tree-sitter, not blind line windows. What Solar Pro 3 receives inside those 131,072 tokens is therefore whole declarations, not truncated line ranges. Doubling from Solar Pro 2's 65,536 tokens at the same $0.25/$0.25 per Mtok is a straight upgrade. The 128,000 token OpenRouter listing is close enough that context is rarely the deciding factor between the two routes; the $0.60 output price is.
What are the tradeoffs of Upstage Solar Pro 3 in a coding agent?
Upstage Solar Pro 3's first-party API caps max output at 8,192 tokens, which for a reasoning model means thinking plus answer have to fit in 8K together. That is the hard constraint, and it is the one that decides whether Solar Pro 3 fits your Atlas workflow.
Two limits are worth stating plainly about Solar Pro 3. The 8,192 token output cap on the Upstage API forces thinking and answer to share one budget, so a deep reasoning trace eats the room the diff needs. The practical fix inside Atlas is to keep tasks scoped: Atlas computes a unified diff for every file edit and surfaces it for approval before writing, and small diffs are what that review loop is built for. The second limit is reputational. Upstage is a small Korean lab. Solar Pro 3 does not appear on the large-repo agentic coding leaderboards that Claude, GPT-5, and Kimi K2 compete on. That is not evidence it is bad, but it is an absence of evidence, and you should treat it as one.
When should you switch off Upstage Solar Pro 3 to another model?
Switch off Upstage Solar Pro 3 when a task needs more than 8,192 output tokens in a single turn, or when you need a model with published large-repo agentic coding results. Solar Pro 3 does not appear on the leaderboards Claude, GPT-5, and Kimi K2 compete on.
Upstage Solar Pro 3 is a strong default for cost-stable reasoning at $0.25/$0.25 per Mtok, and a poor default for high-stakes, high-volume refactors. Atlas makes the handoff cheap: Atlas lets you switch the active model and provider on the fly with favorites and recents, so the pattern is to keep `upstage/solar-pro3` on plan-heavy work and flip the build agent to a frontier model when the diff gets large. Atlas snapshots file changes as git patches so edits can be diffed and rolled back, and every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, so experimenting with Solar Pro 3 on a real branch is a bounded risk rather than an open one.
Setup
- 01Export UPSTAGE_API_KEY for the symmetric $0.25/$0.25 first-party route at 131,072 tokens.
- 02Run `atlas models upstage` and confirm the `solar-pro3` row.
- 03Pin `"model": "upstage/solar-pro3"` in atlas.json, with `"small_model": "upstage/solar-mini"` at $0.15/$0.15 per Mtok.
- 04If output-heavy turns are rare and you want cheaper input, use `openrouter/upstage/solar-pro-3` at $0.15/$0.60 per Mtok instead and accept the 128,000 token window.
- 05Keep tasks scoped, because the first-party API caps max output at 8,192 tokens for thinking plus answer combined.
Frequently asked questions
- how much does upstage solar pro 3 cost
- Solar Pro 3 is $0.25/$0.25 per Mtok on the Upstage first-party API, symmetric input and output. OpenRouter lists it at $0.15/$0.60 per Mtok with a 128,000 token context instead of 131,072.
- what is solar pro 3 context window
- Solar Pro 3 holds 131,072 tokens on the Upstage API and 128,000 tokens on OpenRouter. That is double the 65,536 token window of Solar Pro 2 at an identical price per Mtok.
- should I use the upstage api or openrouter for solar pro 3
- Use the Upstage API at $0.25/$0.25 per Mtok if reasoning traces are long, since output is not penalized. Use OpenRouter at $0.15/$0.60 per Mtok if turns are read-heavy and you accept a 128,000 token window.
- how do I set solar pro 3 as the model in atlas
- Export UPSTAGE_API_KEY, run `atlas models upstage` and confirm the `solar-pro3` row, then pin `"model": "upstage/solar-pro3"` in atlas.json with `"small_model": "upstage/solar-mini"`.
- what is the max output of solar pro 3
- The Upstage first-party API caps Solar Pro 3 max output at 8,192 tokens. For a reasoning model that means the thinking trace and the answer have to fit inside 8K together.
- is solar pro 3 good for agentic coding
- Solar Pro 3 is cheap and reasoning-capable at $0.25/$0.25 per Mtok, but Upstage is a small Korean lab and Solar Pro 3 does not appear on the large-repo agentic coding leaderboards that Claude, GPT-5, and Kimi K2 compete on.
- why is solar pro 3 output the same price as input
- Upstage prices Solar Pro 3 symmetrically at $0.25 per Mtok in both directions on the first-party API. For a reasoning model, that means thinking traces cost exactly what context costs, the opposite of the usual 4x output penalty.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Symfony in 2026
Atlas is a terminal-native AI coding agent for Symfony in 2026. It reads autowired services and Doctrine mappings, and shows migration SQL before anything runs.
Atlas for Pandas: Terminal-Native AI Coding in 2026
Atlas is a terminal-native AI coding agent for Pandas. Vectorize df.apply, fix chained assignment under Copy-on-Write, and pin DataFrames with assert_frame_equal.
Atlas vs PearAI: Choosing Your AI Coding Agent in 2026
Compare Atlas, the terminal-native AI coding agent, with PearAI, a VS Code fork. Evaluate features, pricing, and community support for developers in 2026.
Atlas vs Windsurf: Terminal AI Coding Agents in 2026
Atlas and Windsurf offer distinct AI coding experiences in 2026. Atlas provides a terminal-native TUI with permission-gated tools, while Windsurf features an in-editor live preview and a dedicated GUI.
Atlas for Crystal: A Terminal-Native AI Coding Agent for shard.yml Projects in 2026
Atlas is a terminal-native AI coding agent for Crystal in 2026. It reads shard.yml and union types, narrows Nil unions, runs crystal spec behind a prompt, and formats the diff.
Atlas for .NET: Your AI Coding Agent in 2026
Atlas empowers .NET developers in 2026 with a terminal-native AI coding agent. Securely integrate Atlas with ASP.NET Core for web APIs and services, leveraging local embeddings and robust safety features for efficient
Trace a Runtime Bug from a Stack Trace with Atlas in 2026
How to trace a runtime bug from a stack trace with Atlas in 2026: read each frame at its offset, grep for the error string, and use the lsp tool to find callers.
Atlas vs Zed: Terminal AI Coding Agents in 2026
Comparing Atlas, the terminal-native AI coding agent, with Zed, the GPU-accelerated collaborative editor, for developers in 2026. Evaluate their AI autonomy, privacy, and workflow.