CL-SN55 941 —
GPT-61S 885 —
GPT-56T 861 —
GPT-6S 825 —
CL-OP55X 819 —
MUSE-SPK 793 —
QWEN-38X 779 —
GPT-6A 778 —
CL-OP5H 712 —
CL-OP5X 710 —
KIMI-K3X 700 —
CL-OP46H 695 —
CL-FAB5H 693 —
CL-OP55H 693 —
CL-OP47H 684 —
GEM-37FH 649 —
GEM-38FH 634 —
GPT-56S 561 -0.4%
INKL 531 —
CL-OP47 530 -0.4%
GEM-3P 493 —
CL-OP46 488 -0.2%
CL-OP48 481 —
GEM-31P 454 -0.2%
GPT-6L 436 —
CL-SN55 941 —
GPT-61S 885 —
GPT-56T 861 —
GPT-6S 825 —
CL-OP55X 819 —
MUSE-SPK 793 —
QWEN-38X 779 —
GPT-6A 778 —
CL-OP5H 712 —
CL-OP5X 710 —
KIMI-K3X 700 —
CL-OP46H 695 —
CL-FAB5H 693 —
CL-OP55H 693 —
CL-OP47H 684 —
GEM-37FH 649 —
GEM-38FH 634 —
GPT-56S 561 -0.4%
INKL 531 —
CL-OP47 530 -0.4%
GEM-3P 493 —
CL-OP46 488 -0.2%
CL-OP48 481 —
GEM-31P 454 -0.2%
GPT-6L 436 —
← Back to feed

Upstage Solar Pro 4 Jumps 28 Points on AA Intelligence Index — Terminal-Bench Up 4.8x, $0.30/$1.20 Per Million

Korean AI lab Upstage has released Solar Pro 4, its new proprietary flagship reasoning model. The model scores 42 on the Artificial Analysis Intelligence Index — a 28-point gain over Solar Pro 3’s 14, cleared in a single generation. At 42, Solar Pro 4 sits alongside Thinking Machines’ Inkling (xhigh, 42) and just behind Xiaomi MiMo-V2.5-Pro (43), placing it in the mid-tier frontier bracket rather than the sub-40 class where Korean models had previously plateaued.

Benchmark Numbers

The gains concentrate on agent execution, not general knowledge. Upstage’s own evaluation table:

CategorySolar Open 2Solar Pro 4Change
Terminal-Bench v2.143.257.0+13.8
GDPval-AA v231.438.8+7.4
BrowseComp37.349.2+11.9
SWE-Bench Verified (OpenHands)69.270.6+1.4
MCP-Atlas58.261.4+3.2
APEX-Agents16.618.7+2.1
τ³-Banking18.123.0+4.9
AA-LCR (long context)62.771.0+8.3
GPQA Diamond85.689.0+3.4

Terminal task performance is up by a factor of 4.8 from Solar Pro 3. τ³-Banking multi-turn tool use is up 2.5x from the same baseline. The message from the benchmark mix is direct: Solar Pro 4 was built to complete longer agentic workflows, not to move individual-question leaderboards.

Where improvement is minimal is instructive. AIME 2026 fell 0.4 points. Korean and Japanese professional benchmarks (law, medicine, KMMLU-Pro) moved less than one point. Math reasoning headroom from Solar Pro 3 was already largely saturated; the engineering effort went elsewhere.

Pricing and Specs

  • Input: $0.30 per million tokens
  • Output: $1.20 per million tokens
  • Cache hit: $0.06 per million tokens
  • Context window: 512K tokens, up to 128K output
  • Languages: English and Korean (stated), Japanese in scope
  • Modes: reasoning-optimised for deep analysis, lighter for everyday tasks

Pricing is double Solar Pro 3’s $0.15/$0.60 rate. At $0.30 input and $1.20 output, Solar Pro 4 costs roughly the same per token as Gemini 3.5 Flash and less than a quarter of Gemini 3.1 Pro’s $2/$12 rate — significant positioning given that Solar Pro 4’s AA Index score of 42 closes to within a few points of where Gemini 3 Pro (index 44–46) sat a generation ago.

Third-Nation Context

South Korea now holds AA Intelligence Index rankings in the low-40s for two consecutive model generations, consolidating a position behind the US (with frontier models at 56–64+) and China (DeepSeek, Kimi K3, GLM-5.2 in the mid-to-upper tier). Upstage’s positioning emphasises agent work and Korean/Japanese professional domains — a differentiation strategy consistent with the customer base it has built in financial services, legal, and public sector applications in Northeast Asia.

Solar Pro 4 is available through Upstage Console. Benchmark scores are from Artificial Analysis and Upstage’s own evaluation environment; the AA public listing is described as forthcoming.