The Frontier

Upstage ships Solar Pro 4, an agentic model with 512K context

Editorial

Editorial

— 3 min read
Share
Upstage ships Solar Pro 4, an agentic model with 512K context

Korean AI lab Upstage rounded out its 2026 lineup this week with Solar Pro 4, a closed-API flagship aimed squarely at document-heavy agent work, roughly a week after the open-weights Solar Open 2. It is the company's clearest bet yet that the next AI battleground is finishing office jobs, not passing chat tests.

Upstage released Solar Pro 4, a proprietary agentic model for multi-document, multi-tool workloads, with a 512K-token context window and up to 128K output tokens. It is the commercial sibling to Solar Open 2, which ships open weights for self-hosting; Solar Pro 4 is API-only, targets jobs that chain tool calls across many steps, and handles English, Korean, and Japanese. Reasoning is on by default with an effort dial for latency, and the API is OpenAI-compatible, so Upstage's migration pitch is that swapping the endpoint and model name is the whole integration. The model is live on Upstage's console, SolarChat, OpenRouter, and Upstage Studio, with dedicated or on-premises deployments available through sales.

Pricing is built for agent workloads that burn many calls per task: $0.30 per million input tokens, $0.06 cached, and $1.20 for output — with a 90% launch discount through September 10 that cuts the OpenRouter price to $0.03 in and $0.12 out. Upstage's benchmark claims center on agentic evals, where the gap over Solar Open 2 is largest: 57.0 vs 43.2 on Terminal-Bench v2.1, 71.0 vs 62.7 on the long-document AA-LCR, 49.2 vs 37.3 on BrowseComp, and 23.0 vs 18.1 on multi-turn banking (τ³-Banking). Knowledge and coding scores are roughly level (GPQA Diamond 89.0). Several of those numbers are vendor-cited or in-house, so independent verification is still pending; the model has already appeared on the Agent Arena leaderboard.

The product story is as deliberate as the benchmark story. Upstage says Solar Pro 4 was trained on finished work via OfficeVerse, a pipeline that synthesizes office tasks across 11 industry domains and grades them pass or fail on the final deliverable, and the launch demo shows the model turning three prompts into an Excel workbook, a review report, and a slide deck in a single session. A companion agent cookbook ships seven work agents, and the model is explicitly designed to label answers as grounded, not-in-document, or mismatch rather than fabricate when the evidence runs out — the "say it can't verify" behavior enterprise buyers keep asking for.

What makes this worth watching isn't just another long-context API. A mid-tier lab is betting its flagship on finish-the-job reliability at aggressive prices, in a market where OpenAI and Anthropic still charge a premium for agentic access. If independent evals hold up, Solar Pro 4 becomes a credible budget option for the office-automation workloads that are quietly the biggest real-world agent market — and another sign Korea's AI push is producing exportable products, not just chip rallies.

What to watch: independent benchmark runs before the September 10 promo ends.

Do you trust vendor benchmarks for agentic models, or do you wait for independent runs? Tell us in the comments.

Read more

Brain organoids classified Braille at 83% — in a trio, not alone

Brain organoids classified Braille at 83% — in a trio, not alone

Two stories today: human brain tissue on a chip decoded a tactile signal, and the data-centre company behind some of the largest AI contracts in Europe turns out to be six months behind on its flagship site. Human forebrain organoids grown on a microelectrode array classified Braille letters with 61 percent accuracy on their own and 83 percent when the responses of three organoids were combined, in a peer-reviewed study from the University of Bristol. The team took event-based tactile readings

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute

Microsoft finished pre-training its biggest in-house model yet — then used it to make a point about the company it bankrolls. Microsoft AI completed pre-training of MAI-Base-2, a mixture-of-experts foundation model with 80 billion active and 2.8 trillion total parameters, and says the recipe is 4.2–12.3x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. The run itself is the number worth keeping: 16,000 Nvidia GB200 GPUs for 38 days across August and September, at 21% average model

Amazon's new Bee devices listen all day — for $11.99 a month

Amazon's new Bee devices listen all day — for $11.99 a month

Amazon turned its always-listening Bee line into a two-device family on Thursday, and NVIDIA opened a 2.8-trillion-token Indic corpus recipe. The connecting thread: both are bets that the next AI money is in memory — yours, or your model's. Amazon put two new always-listening devices on sale, and both now lead to a subscription. Bee Frontier, the next-generation wristband that also clips to clothing, ships now at $74.99 in Onyx, Lavender and Sage — up from the $50 predecessor — with seven-day

Perplexity open-sources its agent sandbox after 118 failed escapes

Perplexity open-sources its agent sandbox after 118 failed escapes

Perplexity handed its agent containment layer to the open-source world, and Washington's AI task force stopped reading briefings and started working. Perplexity open-sourced SPACE on Thursday — the sandbox it built to keep its own AI agents from getting out — and says nine frontier models failed to break out of it across 118 attempts. Each agent runs inside a small virtual machine with its own operating system, with credentials and network access managed outside the sandbox so an agent can re