Teams building coding agents, terminal automation, or long-document review should evaluate Solar Pro 4 before defaulting to a bigger-name model: it pushes max output to 131,072 tokens, runs at roughly 37.7 tokens per second, and was trained on synthetic tasks spanning 12 work types. It is not the right pick if you need vision input or a disclosed parameter count for capacity planning.
Solar Pro 4 is Upstage's agent-first flagship LLM, released in August 2026 with a GPQA Diamond score of 89.0. It targets multi-step agent work like terminal tasks and long-document review rather than single-shot chat, competing on price against GPT and Claude-class models.
Provider: Upstage · Family: Solar Pro
Context window: 524,288 tokens · Max output: 131,072
Input modalities: text · Output: text
About Solar Pro 4
Solar Pro 4 is Upstage's flagship large language model, released August 10, 2026 as the direct successor to January 2026's Solar Pro 3. Upstage has not disclosed whether the model uses a dense or mixture-of-experts architecture, nor its parameter count, a departure from the fully documented 102-billion-parameter Solar Pro 3 and the open-weight Solar Open 2 sibling. It is positioned as an agent-first model: built, in Upstage's own framing, to finish multi-step work end-to-end rather than answer single-shot prompts. Independent scores from Artificial Analysis, published alongside the August 2026 launch, put Solar Pro 4 at 89.0 on GPQA Diamond graduate-level reasoning, 71.0 on the AA-LCR long-context reasoning eval, and 57.0 on Terminal-Bench v2.1, a benchmark of real agent execution in a terminal environment. It also scores 38.8 on GDPval-AA v2 (work-deliverable quality), 23.0 on the tau-cubed-Banking multi-turn tool-use eval, and 81.9 on the Japanese-language Arena-Hard v2. Upstage has not published SWE-bench, MMLU, GSM8K or HumanEval scores for this specific release, so direct comparison to competitors on those standard benchmarks is not yet possible from public data. Solar Pro 4 ships with a 524,288-token context window (524K) and a 131,072-token maximum output (128K), both a large step up from Solar Pro 3's working range. Upstage markets the larger window specifically for long-document and multi-file agent tasks, and the AA-LCR score suggests the model retains usable recall deep into that window rather than only near the top. The model is text-only: it accepts and returns text, with no confirmed vision, audio, or video input despite several competing 2026 flagships shipping multimodal input by default. It supports tool calling, function calling, structured and JSON outputs, and streaming, and Upstage's benchmark selection (Terminal-Bench, tau-cubed-Banking) is built specifically around multi-turn tool use rather than chat quality alone. No computer-use or web-browsing capability has been confirmed for this model specifically. Output throughput measures roughly 37.7 tokens per second on third-party trackers, in the lower-middle of tracked models. List pricing is $0.30 per 1 million input tokens and $1.20 per 1 million output tokens, with cached input at $0.06 per 1 million tokens, billed through the Upstage Console API. A launch promotion cuts that by 90 percent, to $0.03 input and $0.12 output, on the Upstage Console and OpenRouter through September 10, 2026 at 23:59 UTC, after which pricing reverts to the full list rate with a hard cutoff rather than a gradual step-down. Solar Pro 4 is available through the Upstage Console API, the consumer-facing SolarChat interface, OpenRouter, Upstage's own Hermes Agent and Upstage Studio products, and as a dedicated or on-premises deployment for enterprise buyers. Upstage has not confirmed listings on AWS Bedrock, Google Vertex AI, or Azure for this specific model, unlike some smaller Solar models that ship through Amazon SageMaker. There are no open weights: unlike the Solar Open 2 sibling released three weeks earlier, Solar Pro 4 has no Hugging Face release and cannot be self-hosted. Upstage has not published a formal system card or safety report for Solar Pro 4. Its one publicly stated safety design choice is to have the model say it cannot verify a claim rather than fill the gap with a plausible-sounding guess, a deliberate reliability-over-completeness tradeoff built into its OfficeVerse training pipeline. Training data cutoff is reported as February 2026 by third-party model trackers; Upstage's own launch materials do not state a cutoff date directly. Upstage built Solar Pro 4's training data through an internal pipeline it calls OfficeVerse, which synthesizes tasks across 11 industry domains and 12 task types rather than relying primarily on scraped open-web text. Upstage states company-wide that it does not use customer API data to train its models by default, and enterprise Document Intelligence workloads can run entirely inside a customer's own AWS account for data residency; no Solar-Pro-4-specific retention policy beyond that company-wide statement has been published. Solar Pro 4 is built for long-document review, terminal and coding agents, and multi-turn tool-use workflows where finishing the task matters more than chat polish, and its list pricing undercuts most frontier-adjacent competitors on a per-token basis. Teams needing vision or audio input, a disclosed parameter count for capacity planning, or verified SWE-bench and MMLU scores for procurement comparisons should look at a more fully benchmarked and documented competitor instead. Compared with Solar Pro 3, Upstage reports a 13.8-point gain on Terminal-Bench, an 11.9-point gain on the BrowseComp web-research eval, and a 7.4-point gain on GDPval work-deliverable quality, with general knowledge performance roughly flat. That pattern, concentrated gains on agentic execution rather than raw knowledge, matches Upstage's own framing of Solar Pro 4 as an agent-first release rather than a general capability upgrade.
Pricing
List price is $0.30 per 1M input tokens and $1.20 per 1M output tokens, with cached input at $0.06 per 1M, billed via the Upstage Console API. A launch promotion cuts all three by 90% (to $0.03 / $0.12 / roughly $0.006) on the Upstage Console and OpenRouter through September 10, 2026 at 23:59 UTC, then reverts to list price with a hard cutoff rather than a gradual step-down.
Key Features
- 524K Context Window: 524,288 tokens of input context, aimed at long-document and multi-file agent tasks rather than short chat turns, with output length scaling to match sustained agent responses.
- Agentic Tool Use: Scores 38.8 on GDPval-AA v2 for work-deliverable quality and 23.0 on the tau-cubed-Banking multi-turn tool-use eval, benchmarks built around finishing real tasks rather than answering questions.
- Long-Context Recall: Long documents stay coherent deep into the context window, per Artificial Analysis' long-context reasoning evaluation, not just near the start.
- Verify-or-Decline Behavior: Trained via Upstage's OfficeVerse pipeline to state when it cannot verify a claim rather than produce a plausible-sounding guess, a deliberate reliability tradeoff.
- Prompt Caching: Cached input tokens carry an 80 percent discount versus uncached input on the Upstage Console API, cutting the cost of repeated system prompts or reused documents.
Pros
- Concentrates its generation-over-generation gains on real agent work over raw knowledge: Terminal-Bench and BrowseComp both improved by double digits versus Solar Pro 3, while general knowledge stayed roughly flat.
- Its long-context reasoning score on Artificial Analysis' AA-LCR evaluation suggests the model's context window holds up on real documents rather than only in needle-in-haystack tests.
- Standard per-token rates beat most frontier-adjacent competitors even before Upstage's steep temporary launch discount, which runs into mid-September 2026.
- Supports tool calling, structured and JSON outputs, and multi-turn agent loops natively, reflected in dedicated benchmark coverage that most general chat models are not evaluated on.
Cons
- Text-only: no vision, audio, or video input, unlike several competing 2026 flagships that ship multimodal by default.
- Upstage has not disclosed Solar Pro 4's parameter count or dense-vs-MoE architecture, making it harder to reason about cost or latency the way you can with the fully documented Solar Open 2.
- No published SWE-bench, MMLU, GSM8K, or HumanEval scores yet, so head-to-head comparison against most competitors relies on a different, Artificial-Analysis-sourced benchmark set.
- Output throughput lands in the lower-middle of tracked models on third-party speed rankings, noticeably slower than models optimized specifically for chat latency.
Benchmarks
- aa lcr: 71
- gdpval aa v2: 38.8
- gpqa diamond: 89
- tau cubed banking: 23
- terminal bench v2 1: 57
- arena hard v2 japanese: 81.9
- artificial analysis price blended per m: 0.05
- artificial analysis speed tokens per sec: 37.7
Frequently Asked Questions
How much does Solar Pro 4 cost per 1M tokens?
Solar Pro 4 bills $0.30 per million input tokens and $1.20 per million output on the standard Upstage Console rate, with a discounted $0.06 per million for cached input. Upstage cut all three by 90 percent at launch, so the promotional rate through September 10, 2026 is $0.03 / $0.12 / roughly half a cent, then jumps straight back to the standard rate with no gradual phase-out. That standard rate still undercuts most frontier-adjacent competitors on a blended per-token basis.
How does Solar Pro 4 compare on benchmarks vs Solar Pro 3?
Upstage has not published SWE-bench, MMLU, or GSM8K scores for Solar Pro 4, so like-for-like comparison against GPT or Claude-class models is limited. On the benchmarks Upstage did report, the model's 71.0 score on Artificial Analysis' long-context reasoning eval stands out, alongside solid results on Terminal-Bench and GDPval-AA v2, though independent verification against SWE-bench-style coding benchmarks is not yet available.
Is Solar Pro 4 open source or proprietary?
Solar Pro 4 is proprietary and API-only: Upstage has not released its weights on Hugging Face or elsewhere, unlike the open-weight Solar Open 2 model it shipped three weeks earlier. Access is through the Upstage Console API, OpenRouter, the SolarChat consumer app, or a dedicated on-premises deployment for enterprise buyers, not through self-hosting.
Does Solar Pro 4 train on user data?
Solar Pro 4's training data comes from an internal Upstage pipeline called OfficeVerse, which generates synthetic agent tasks across 11 industry domains rather than relying primarily on scraped web text. Company-wide, Upstage states that customer API inputs are not used to train its models by default; the February 2026 training cutoff comes from third-party trackers rather than an Upstage-published figure.
Who is Solar Pro 4 best for and who should avoid it?
Solar Pro 4 suits coding agents, terminal automation, and long-document review workflows where finishing a multi-step task matters more than chat polish or multimodal input. Buyers who need image or audio support, want a published parameter count for capacity planning, or require verified SWE-bench and MMLU scores for procurement should look at a more fully benchmarked competitor instead.