Solar Pro 4: a 512K-context agent model at $0.30/1M input tokens
Upstage's drop-in, OpenAI-compatible model scores 57 on Terminal-Bench, flags gaps instead of bluffing, and runs 90% off to Sep 10 — plus Meta's 30B open agent.
Copy markdown
It finishes multi-step jobs, not just chats
Solar Pro 4 is tuned for agent loops rather than single answers: it scores 57 on Terminal-Bench v2.1 (live, multi-step shell tasks) and handles multi-turn tool use across knowledge bases. Upstage's whole pitch is end-to-end completion — the model runs the job to the finish instead of handing back a plan.
$0.30 in, $1.20 out — and 90% off until Sep 10
Pricing is $0.30 per 1M input tokens and $1.20 per 1M output, with cached input at $0.06 — already cheap for a long-context agent model. A launch promo cuts all of it by 90% (roughly $0.03/$0.12) through Sep 10, 2026 on the Upstage Console and OpenRouter.
512K context, and it says “I can’t verify” instead of bluffing
The window is 512K tokens with up to 128K output, and it scores 71 on AA-LCR long-document reasoning. It's built to flag missing evidence rather than fill the gap — a hallucination guardrail that matters when you point an agent at contracts, codebases, or research.
Migration is a one-line endpoint swap
The API is OpenAI-compatible: point your client at Upstage's endpoint and set the model to solar-pro4 — that's the whole change. It's also live on OpenRouter, SolarChat, Hermes Agent, and Upstage Studio, so you can A/B it against your current backend today.
Need the weights? Solar Open 2 self-hosts
If you want on-prem control or data residency, Upstage's open-weights Solar Open 2 is the companion release. It matches Pro 4 on knowledge, math, and coding but trails on agent work (about 14 points lower on terminal tasks), so reach for Pro 4 when you need autonomy.
Elsewhere: Meta open-sources Muse Glimmer, a 30B agent for one GPU
Meta returned to open weights with Muse Glimmer, an Apache-2.0 30B model tuned for agents that runs on a single consumer GPU. It's a commercial-friendly local alternative if you'd rather host than pay per token — no caps, no metered bill.