Opus vs Sol
Anthropic flagship architect vs OpenAI max-effort generalist — hard repos and judgment. Public stats are peer-relative ranks (40–95 band among scored Builds), not absolute “% intelligence.”
Claude Opus 4.8 · API · High effort (default)
The expensive architect — pays for hard repos, long agents, and fewer retries.
GPT-5.6 Sol · API · Max effort
The ecosystem operator — max-effort Sol for hard agents, not cheap chat.
Six attributes (peer public)
Same-harness & loadout
| Signal | Opus | Sol |
|---|---|---|
| Vals SWE Verified · mini-swe-agent | 88.6% | 96.2% |
| AA GPQA Diamond | 92.0% | 94.1% |
| AA HLE (closed-book) | 45.7% | 47.2% |
| AA-Omniscience | index ~27 · lower hall | acc ~59% · high hall pattern |
| List price (in/out) | $5 / $25 | $5 / $30 |
Large Existing Codebase
Decisive on Execution: Sol — public ranges do not overlap (Opus 73.5–81.5 vs Sol 90.9–98.9).
- Large Existing CodebaseSol · conf A
Same mini-swe harness: Sol 96.2% vs Opus 88.6%; peer STR also Sol.
Quest Fit → - Deep ResearchOpus · conf B
Peer WIS max; better calibration narrative despite Sol INT raw edge.
Quest Fit → - Autonomous Coding AgentSol · conf B
Coding ceiling; Opus still preferred if you need Claude stack + judgment loops.
Quest Fit → - Fix a Difficult BugTie → Sol slight · conf B
Both elite; Sol STR lead, Opus WIS for careful isolation.
Quest Fit →
Opus: Calibration and WIS lead; still elite coding when you want Claude tooling and fewer wild guesses.
Sol: Independent agentic coding peak of the peer set; weaker Omniscience-style judgment.