by Upstage
Solar Pro 4 review, pricing and verdict
Upstage's agent-first flagship, built to finish multi-step work end-to-end rather than answer single-shot prompts.
- Best for
- Coding and terminal agents
- Not for
- Vision or document-scan workflows
Should you use Solar Pro 4?
Solar Pro 4 suits teams building coding agents, terminal automation or long-document review in English, Korean or Japanese, and Artificial Analysis clocks its output at 95.4 tokens per second. Skip it if you need image input or open weights, or want agent scores run by a third party, since that independent evaluation is still pending.
Best for
- Coding and terminal agents Scores 57.0 on Terminal-Bench v2.1, 13.8 points above Solar Open 2 in Upstage's launch table, the largest gap Upstage reports.
- Long-document and multi-file review 524,288-token context window plus a 71.0 AA-LCR score indicate recall holds up deep into large inputs, not just near the top.
- Teams already on OpenAI-compatible agent tooling The Console API is OpenAI-compatible, so switching means changing the endpoint and the model name solar-pro4.
Not for
- Vision or document-scan workflows No image input; scanned documents need a separate OCR step before reaching the model.
- Procurement that requires third-party-run agent benchmarks SWE-bench Verified, MCP-Atlas and BrowseComp figures are Upstage in-house runs, and Artificial Analysis' own evaluation is still forthcoming as of 13 October 2026.
Solar Pro 4 pricing
- Input, per 1M tokens
- $0.30/1M
- Output, per 1M tokens
- $1.20/1M
- Cached input, per 1M tokens
- $0.06/1M
- Pricing model
- per-token
- Free tier
- No
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.0090 | $0.0012 | $0.010 |
| Support reply | $0.0006 | $0.0004 | $0.0010 |
| One coding agent run | $0.060 | $0.024 | $0.084 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Cost examples
- Summarize a 100,000-token document (list price)
- $0.03
- Fill the full 524K-token context window once (list price)
- $0.16
- Coding agent loop, 2M tokens in / 300K out (list price)
- $0.96
Upstage lists Solar Pro 4 at $0.30 per million input tokens, $1.20 per million output tokens and $0.06 per million cached-input tokens on the Upstage Console API, the same rates Artificial Analysis lists. The 90% launch discount on the Console and OpenRouter ended on September 10, 2026 at 23:59 UTC with no phased step-up, so every request now bills at those three rates. Upstage sells dedicated and on-premises deployments on quote only.
What Solar Pro 4 can do
Solar Pro 4 is Upstage's agent-first reasoning LLM, released in August 2026. It accepts 524,288 tokens of text input, returns text only, and is sold through an API rather than as open weights. Upstage built it for multi-step agent work such as terminal tasks, tool calling and long-document review.
What goes in and what comes out
Reads
- text
Writes
- text
Capabilities
- Function callingSupported
- Structured outputSupported
- Tool useSupported
- Audio inputNot supported
- Audio outputNot supported
- Code executionNot supported
- Video inputNot supported
- VisionNot supported
- Web browsingNot supported
- Reasoning modes
- agentic-tool-use
- Long context recall
- high
Key features
524K Context Window
Loads several contracts, reports and data files into one session without splitting them, with output running up to 128K tokens.
Agentic Tool Use
Trained on OfficeVerse tasks spanning 12 task types and graded on the finished deliverable, then measured on multi-turn tool-use and work-deliverable evals rather than single questions.
Long-Context Recall
Built for questions whose answers are scattered across roughly 100,000 tokens of documents, the shape Artificial Analysis' AA-LCR eval tests.
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.
Show all 6 features
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.
Reasoning-Effort Control
Reasoning is on by default with the trace returned in each response; a single setting trades depth for speed, from high-effort analysis to low-latency turns.
Known limitations
- Text-only: no image, audio or video input, so scanned documents need a separate OCR step.
- Artificial Analysis lists only an estimated Intelligence Index score of 28 (v4.3.2) as of 13 October 2026; its independent evaluation is still forthcoming.
- Upstage has not disclosed parameter count or dense-vs-MoE architecture for this model, unlike its open-weight Solar Open 2 sibling.
Who it is for
- AI agent engineers building coding or terminal automation
- Teams doing long-document or multi-file review
- Cost-sensitive teams running high-volume agent workloads
- Korean- and Japanese-language enterprise document workflows
Supported languages
- English
- Korean
- Japanese
Solar Pro 4 benchmarks
Solar Pro 4 against its nearest rivals
Output speed tokens per second, higher is better
Standout result
GPQA Diamond score, per Artificial Analysis
Source: Upstage Solar Pro 4 launch post / Artificial Analysis, Aug 10 2026Where it sits
Cheaper than 70% of the 103 GA models with a published price, mid-pack on SWE-bench Verified (rank 26 of 39), and one of 89 whose vendor states it does not train on customer data.
| Measure | This model | Rank | Median |
|---|---|---|---|
| Blended price (3:1)$ per 1M tokens | $0.525/M | #29 / 103 | peer median $1.71/M |
| Output speedtokens/s | 95 tok/s | #26 / 57 | peer median 94 tok/s |
| SWE-bench Verified% solved | 70.6% | #26 / 39 | peer median 76.5% |
| GPQA Diamond% correct | 89% | #28 / 63 | peer median 87.8% |
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
Measured figures
- GDPval-AA v2
- 38.8
- Output speed
- 95 tok/s
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Strengths and limits
Strengths
- Its gains sit where agent builders need them: Upstage's own table shows the lead over Solar Open 2 concentrated on terminal, web-research and deliverable tasks while knowledge and maths stay level.
- Long-context reasoning holds up on Artificial Analysis' AA-LCR eval, which scatters its answers across roughly 100,000 tokens of reports and filings.
- It says when a document lacks the evidence instead of inventing a clause or figure, which suits contract and filing review where a fabricated answer costs more than a gap.
- Scores 89.0 on GPQA Diamond and 71.0 on AA-LCR long-context reasoning (Artificial Analysis figures in Upstage's launch post), evidence the 524,288-token window holds up on real documents.
Show 2 more
- Upstage's launch table puts its lead over the open-weight Solar Open 2 on agent work: +13.8 on Terminal-Bench v2.1 and +11.9 on BrowseComp, with knowledge scores level.
- Reasons by default with an adjustable reasoning-effort setting, so one model covers deep document analysis and low-latency interactive turns.
Limits
- Text-only: no vision, audio or video input, unlike several 2026 models at a similar per-token price.
- Upstage has not disclosed the parameter count or whether the model is dense or mixture-of-experts, so capacity planning has to start from measured throughput.
- Most of the headline coding and agent figures, including SWE-bench Verified, MCP-Atlas and BrowseComp, come from Upstage's in-house runs rather than a third party.
- Reasoning runs by default and Artificial Analysis measured about 23 seconds to the first answer token, so real-time chat needs the low reasoning-effort setting.
- No system card or safety report has been published for this model.
Solar Pro 4 alternatives
The nearest AI models on Hokai for the same job.
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| Solar Pro 4 This page | Upstage | 524K | $0.30 / $1.20 | text → text |
 Claude Fable 5Superseded by Fable 5.1 and Fable 5.2 but still sold on the Claude API: Anthropic's first generally available Mythos-class model, released June 2026, with a 1M-token context. | Anthropic | 1M | $10 / $50 | text, image, pdf → text |
| Claude Fable 5.1Superseded by Claude Fable 5.2 in November 2026 but still sold on the Claude API: Anthropic's second Mythos-class model, released Sept 1, 2026, for long-horizon coding and research. | Anthropic | 1M | $10 / $50 | text, image, pdf → text |
 Gemini Decision 1Google DeepMind's decision model scores many questions over one multimodal input of up to 262,144 tokens and bills zero output tokens. | 262K | $0.10 / $0 | text, image, audio, video, pdf → text |
Solar Pro 4 specs
- Context window
- 524K tokens
- Max output
- 131K tokens
- Latency (P50)
- 1600 ms
- Tokens per second
- 95.4 tok/s
- Mixture of experts
- No
- Parameters
- Not disclosed
Training data
Upstage has not published a full training-data breakdown for Solar Pro 4. The model was built via an internal 'OfficeVerse' synthetic-data pipeline that generates tasks from 11 industry domains and 12 task types, targeting agentic, document-heavy work rather than open-web chat.
API endpoints
| Name | Endpoint | Auth |
|---|---|---|
| Upstage Console API | https://api.upstage.ai/v1/chat/completions | API Key |
| OpenRouter | https://openrouter.ai/upstage/solar-pro4 | API Key |
Prompting patterns
Show 2 prompting patterns
Ask for best-effort estimates explicitly
Add "Give your best estimate even if you're not fully certain, and flag your confidence" to the system prompt.
The model defaults to declining rather than guessing, which is safer but can under-deliver on brainstorming-style tasks.
Front-load long documents, then ask
Paste the full document or codebase first, then ask the question in a separate final message.
Matches how the AA-LCR long-context eval structures its prompts, the scenario Upstage optimized for.
Known quirks
Show 4 known quirks
Declines rather than guesses
Trained via Upstage's OfficeVerse pipeline to state it cannot verify a claim rather than produce a plausible-sounding answer when evidence is thin.
Workaround: Explicitly prompt for a best-effort estimate ("answer even if uncertain, and flag your confidence") when a tentative answer is more useful than a refusal.
Reasoning on by default
Every response carries a reasoning trace, and Artificial Analysis measured about 23 seconds to the first answer token on its 13 October 2026 page.
Workaround: Set reasoning effort low for interactive turns and keep high effort for document analysis and multi-step agent runs.
No disclosed architecture
Upstage has not stated whether Solar Pro 4 is dense or mixture-of-experts, or its parameter count, unlike the fully documented Solar Open 2.
Workaround: Benchmark your own workload directly for cost and latency planning instead of estimating from parameter count; Artificial Analysis' measured 95.4 output tokens per second is a usable capacity baseline.
Text-only despite agent focus
Solar Pro 4 has no confirmed vision, audio, or video input, so document-heavy agent workflows that start from scans or photos need a separate step.
Workaround: Route scanned documents or images through an OCR step, such as Upstage's own Document Parse API, before passing text to Solar Pro 4.
Solar Pro 4 facts
- Deployment
- API · On-premises
- Trains on your data
- No
- Licence
- Proprietary (Upstage Commercial API Terms)
- Safety posture
- Balanced
- Released
- 10 Aug 2026
- Current version
- solar-pro4
- Knowledge cutoff
- 01 Feb 2026
- Family
- Solar Pro
- Variant
- Chat
- Data retention
- Upstage states company-wide that customer API inputs are not used to train its models by default; no Solar-Pro-4-specific retention window has been separately published.
Safety controls
- Red teaming
- not publicly disclosed
- System prompt
- Yes
- Constitutional AI
- No
- Content filter API
- No
- Moderation endpoint
- No
Data governance
- Iso 27001
- Yes
- Soc2 type2
- No
- HIPAA eligible
- Yes
- Trains on inputs
- No
Version history
solar-pro4 · 10 Aug 2026
Initial Solar Pro 4 release. GPQA Diamond 89.0 and Terminal-Bench v2.1 57.0 (Artificial Analysis figures, +13.8 vs Solar Open 2 in Upstage's launch table); 524,288-token context.
solar-pro-3 (2x agentic update) · 23 Mar 2026
Solar Pro 3 update roughly doubling agentic benchmark scores.
solar-pro-3 · 26 Jan 2026
Solar Pro 3 initial release: 102B MoE (12B active), Arena-Hard v2 62.5, IFBench 55.78.
What changed
Questions about Solar Pro 4
What does Solar Pro 4 cost now that the launch discount has ended?
Each million tokens costs $0.30 going in and $1.20 coming out on the Upstage Console API, with cached input at $0.06, per Upstage's launch post read on 13 October 2026. The 90 percent launch discount stopped on September 10, 2026, so OpenRouter and the Console now charge these list rates. Cached input is 80 percent cheaper than fresh input, which lowers the bill for agents that resend the same system prompt or document.
How does Solar Pro 4 compare on benchmarks vs Solar Pro 3?
Upstage's launch post quotes Artificial Analysis figures of 71.0 on AA-LCR long-context reasoning, 38.8 on GDPval-AA v2 and 23.0 on tau-cubed-Banking, and calls the agent results a several-fold step over Solar Pro 3 without printing Solar Pro 3's own numbers. Its side-by-side table is against Solar Open 2 instead: Solar Pro 4 leads by 13.8 points on terminal tasks and by 0.1 on MMLU-Pro.
Can you download Solar Pro 4 weights and self-host it?
No. Solar Pro 4 is proprietary and Upstage has released no weights on Hugging Face or anywhere else; the open-weight option in the family is Solar Open 2. You reach Solar Pro 4 through the Upstage Console API, OpenRouter, SolarChat, Hermes Agent or Upstage Studio, and enterprises can contract a dedicated or on-premises deployment instead.
Will Upstage use my API prompts to train Solar Pro 4?
Not by default: Upstage states company-wide that customer API inputs stay out of model training, though it has published no retention window specific to Solar Pro 4. The model was trained on OfficeVerse, an internal pipeline that synthesizes office tasks from public data across 11 industry domains. The February 2026 training cutoff comes from third-party trackers, not from Upstage.
Who is Solar Pro 4 best for and who should avoid it?
Solar Pro 4 fits coding agents, terminal automation and long-document review in English, Korean or Japanese, where a finished multi-step task matters more than chat speed. Avoid it for image or audio input, for self-hosting, or when procurement requires third-party-run agent scores: as of 13 October 2026, Artificial Analysis still lists its own evaluation as forthcoming.
About Solar Pro 4
Solar Pro 4 is the flagship large language model from Upstage, released in August 2026 as the successor to January 2026's Solar Pro 3. Upstage has not disclosed its parameter count or whether it is dense or mixture-of-experts. It reasons by default, returns the reasoning trace with each response, and exposes a reasoning-effort setting: high for deep analysis, low for interactive turns.
Read the full background
The scores Upstage attributes to Artificial Analysis (as of August 2026) are 89.0 on GPQA Diamond, 71.0 on AA-LCR long-context reasoning, 57.0 on Terminal-Bench v2.1, 38.8 on GDPval-AA v2 and 23.0 on the tau-cubed-Banking multi-turn tool-use eval. Upstage's own in-house runs, per the same launch post, add 70.6 on SWE-bench Verified with the OpenHands scaffold, 87.8 on LiveCodeBench, 86.3 on MMLU-Pro, 95.3 on AIME 2026, 61.4 on MCP-Atlas and 49.2 on BrowseComp. Artificial Analysis' own page for the model, read on 13 October 2026, shows only an estimated 28 on its Intelligence Index v4.3.2 and marks its independent evaluation as forthcoming.
Upstage's side-by-side launch table compares Solar Pro 4 with its open-weight sibling Solar Open 2, not with Solar Pro 3. The gap is widest on agent execution (Terminal-Bench v2.1 +13.8, BrowseComp +11.9, GDPval-AA v2 +7.4) and close to zero on knowledge and maths (MMLU-Pro +0.1, AIME 2026 -0.4). For Solar Pro 3, Upstage describes the Terminal-Bench, tau-cubed-Banking and AA-LCR results as a several-fold step but publishes no per-benchmark figures.
The model takes 524,288 tokens of text input (Upstage rounds it to 512K) and writes up to 131,072 output tokens, in English, Korean and Japanese. It is text-only: no image, audio or video input. It supports tool calling, function calling, structured JSON output and streaming, and the Artificial Analysis page puts its output speed at 95.4 tokens per second. Scanned documents need an OCR pass, such as Upstage's Document Parse, before they reach the model.
Access runs through the OpenAI-compatible Upstage Console API (model name solar-pro4), the SolarChat browser app, OpenRouter, Nous Research's Hermes Agent, the no-code Upstage Studio document agents, and dedicated or on-premises deployments for enterprises. There are no open weights. On October 1, 2026 Upstage added Solar Mini 4, a 3B-active-parameter sibling aimed at high-volume agent calls. Per-token rates and the ended launch discount are covered under pricing below.
Upstage has published no system card or safety report for Solar Pro 4. Its stated design choice is that the model says it cannot verify a claim instead of filling the gap, trained through the OfficeVerse pipeline, which synthesizes office tasks across 11 industry domains and 12 task types and grades each on the final deliverable. Upstage says it does not train on customer API data by default, and says its service runs under SOC 2 and ISO 27001 certification.
Who should look elsewhere: teams that need image, audio or PDF input, open weights to self-host, or agent scores run by a third party before a procurement sign-off. Gemini 3.1 Flash-Lite accepts images, audio and video, while MiniMax M3, DeepSeek-V4.1-Flash, Qwen3.8-27B and GLM-5.3-Flash ship downloadable weights; Hokai lists the rest in its model directory.
Sources and guides
Official sources
- Solar Pro 4 model page Vendor
- Licence text Licence
Specifications on this page (context window, token prices, benchmark scores, licence) are read from Upstage's own model card and documentation. The verdict is Hokai's reading of them, not the provider's claim.
Hokai guides that mention Solar Pro 4
Lists that include Solar Pro 4
The provider
More AI Models on Hokai
Shieldstral 1.0Mistral's open-weight multimodal safety classifier, released 4 August 2026, that screens text and images against policies you write.
- Claude Mythos 5.2Anthropic's November 2026 Mythos model: Fable 5.2's weights with relaxed cyber and biology safeguards for verified teams, with a 1M-token context.
ERNIE-5.5-Flash-PreviewBaidu's text-only preview mixture-of-experts model, served through Qianfan on OpenRouter with a 1M-token context window and tool calling.
Gemini 4 CarbonGoogle DeepMind's Gemini 4 Carbon, open to paying Gemini API developers since 10 November: a reasoning model for agentic coding, research and cyber defense.
JevTypeSafe's first System One Model, launched in 2026: skips text generation for typed, calibrated decisions software can act on directly.Nex-N2.6Nex AGI's hosted agentic model for computer use and coding, with image input, tool calling and about 321B parameters, served first-party on OpenRouter.