LLM Digest
Subscribe

Model Release Radar

Price/capability frontier position, benchmark scores, and community signal

View as JSON

Model Release Radar

GPT-5.6 Luna

openai·closed weights·Proprietary·released Thursday, Jul 9, 2026

Frontier position
  • On frontier DeepSWE pass@1 (agentic coding) - no tracked model currently beats it on both price and capability.
Cost vs. capability

Loading chart…

Benchmark scores

Measured cost per task (DeepSWE)

BenchmarkPass@1Measured cost / taskFrontier
DeepSWE (agentic coding) (max effort)67.2%$0.606frontier

Artificial Analysis scores (no matched cost)

Artificial Analysis scores below have no measured per-task cost, so no frontier claim is made for them here.

BenchmarkScore
AA intelligence index50.1
AA coding index68.6
GPQA Diamond (science)89.5%
Humanity's Last Exam37.0%
LCR73.3%
SciCode (coding)50.0%
Tau2 (banking)28.7%
Terminal-Bench 2.1 (agentic)77.9%
Variants

This page covers 6 reasoning-effort variants of the same model.

VariantAA coding indexAA intelligence indexPrice /1M
Extra-high effort (shown above)68.650.1$0.45/1M
High effort63.347.0$0.45/1M
Low effort44.233.9$0.45/1M
Max effort71.452.3$0.45/1M
Medium effort50.738.9$0.45/1M
Non-reasoning39.326.8$0.45/1M
Community signal
VariantCoding EloOverall EloVotes
Extra-high effort1,4621,42922,852