On a difficult new SWE benchmark, ProgramBench, GPT5.5 high/xhigh solves a task for first time, significantly outperforms Opus 4.7
Summary
GPT5.5 achieved the first solve on the difficult ProgramBench SWE benchmark, significantly outperforming Opus 4.7.
Similar Articles
@sashimikun_void: GPT-5.5 outperformed Claude Opus 4.8 on the DEEPSWE benchmark. Opus 4.8 takes twice as long, generates three times the …
GPT-5.5 outperforms Claude Opus 4.8 on the DEEPSWE benchmark, achieving higher scores with lower cost and less token bloat.
ProgramBench result for Fable 5 is in, doubling Opus 4.8 even with 4.8 fallback "99% of the runs"
ProgramBench results show Fable 5 achieving double the performance of Opus 4.8, even with fallback to 4.8 in 99% of runs.
@browser_use: Opus 5 and GPT-5.6 Sol are neck-and-neck on this!
Alexander Yue introduces a new browser-use benchmark where Opus 5 and GPT-5.6 Sol show similar performance, emphasizing the benchmark's robust design with verified rubrics for LLM judges.
@BenjaminDEKR: So it's just over? GPT5.6 Sol Ultra scores 91.9% on TerminalBench Coding is approaching solved, the same way arithmetic…
GPT5.6 Sol Ultra achieves 91.9% on TerminalBench coding benchmark, suggesting coding tasks are approaching solved.
I asked Sol Max to compare the output of Claude Opus 5 High and GPT 5.6 Sol Max on a specific puzzle on ARC-AGI-3 where Opus 5 had a 98.81% score and GPT 5.6 Sol Max had a 21.42% score
An analysis comparing Claude Opus 5 High and GPT 5.6 Sol Max on an ARC-AGI-3 puzzle shows Opus winning by preserving detailed state in visible output, while Sol relies on discarded hidden reasoning.