GPT-6 Sol beats Claude Opus 5.5 on speed and cost, but falters on consistency

OpenAI unveiled GPT-6 Sol in late September, positioning it as a cost-efficient alternative to Anthropic's top-tier models. The company cited benchmark data suggesting Sol outperforms Claude Opus 5 on business automation tasks while operating at roughly 9% of the per-token cost. However, when tested against the newer Opus 5.5—released the same day—the story becomes more nuanced.

Pricing and token economics#

Sol carries a price tag of $2 per million input tokens and $10 per million output tokens, positioning it significantly below Opus 5.5's $4 and $20 structure. This advantage compounds when models generate verbose responses: in real-world testing, Opus 5.5 consumed roughly four times as many output tokens as Sol on identical tasks, making Sol's cost advantage even more pronounced.

Head-to-head testing methodology#

To evaluate the models beyond single-run performance, independent testing ran each through five identical iterations at maximum effort settings. Three categories received scrutiny:

• CI triage: Reading 40 failed build logs and applying conditional rules to decide whether to retry, block, or escalate each failure
• Incident analysis: Parsing 3,664 lines of outage logs to answer seven postmortem questions requiring precision counting and timezone conversion
• Code generation: Writing a complete dependency resolver implementation from specification text, then grading against 120 hidden test cases

Consistency reveals the gap#

On routine CI triage, both models achieved perfect scores across all runs. Sol clocked in at just $0.02 per attempt (18 seconds, 1,143 output tokens) versus Opus 5.5's $0.24 (87 seconds, 11,127 tokens)—aligning with OpenAI's 9% cost claim.

Performance diverged sharply on harder problems. Opus 5.5 maintained flawless accuracy across all 15 tests. Sol, meanwhile, succeeded on 12 of 15, with failures occurring in the two most complex scenarios. On incident logs, Sol twice missed the same customer during refund reconciliation due to a 52-second timing edge case, and miscounted failed transactions in another run. On the resolver spec test, one of five attempts produced syntactically invalid code (a stray parenthesis) that failed all 120 validation tests.

The speed advantage held throughout: Sol ran incident analysis in 1 minute 41 seconds and cost $0.30, compared to Opus 5.5's 6 minutes 44 seconds and $1.68.