The Short Version
Anthropic and OpenAI shipped within hours of each other on 22 September 2026. Sol is half the price. Opus 5.5 is meaningfully better at long-running agentic work. Those two sentences cover most decisions.
Price
| Claude Opus 5.5 | GPT-6 Sol |
|---|
| Input / 1M | $4.00 | $2.00 |
| Output / 1M | $20.00 | $10.00 |
| Cached input / 1M | $0.20 | $0.20 |
| Context window | 1M | 1.05M |
| Max output | 128K | 128K |
Note the cached input row. Both land at $0.20 per million, so on a cache-heavy pipeline the 2x gap narrows considerably. If most of your input is a reused system prompt and a stable document set, the real spread is much closer than the headline suggests.
Where Opus 5.5 Wins
AutomationBench: 40.0% against 33.2%. Nearly seven points on the benchmark that measures how far an agent gets on a task without human intervention. On a 100-task queue that is seven more finished and seven fewer escalations.
Terminal-Bench 4.0: 66.4%. Opus 5.5 beats even GPT-6 Astra here, which costs $10 in and $50 out. If your work is shell-driven, this is the widest margin in the comparison and it runs the opposite way to price.
FrontierCode v1.1: 54.4% against 49.3%. Five points on harder code tasks.
Where Sol Wins
Price, by exactly half. That is not a small thing at volume, and it is the only argument Sol needs for a large class of workloads.
Published cost per task. OpenAI stated 33.2% AutomationBench completion at $0.27 per task. Anthropic published no equivalent figure for Opus 5.5, so you can compare Sol's agentic economics directly and you cannot do the same for Opus without running it yourself.
DeepSWE v1.1: 68.8%. Anthropic did not report Opus 5.5 on this benchmark, so it is a Sol data point with no counterpart rather than a win on a contested field.
The Comparison Nobody Runs
Both vendors publish benchmark scores. Neither publishes tokens consumed per completed task, and that is the number that decides your bill.
A model at half the per-token price that needs 2.5x the tokens to finish is more expensive. A model at twice the price that finishes in one pass where the cheaper one needs three attempts plus a human is cheaper.
Take ten real tickets from your backlog. Run both end to end. Compare total spend per merged result, not per million tokens. It is an afternoon of work and it will contradict at least one assumption you currently hold.
Decision Framework
- Long agentic runs, few human checkpoints - Opus 5.5. The seven-point AutomationBench gap compounds over a long session.
- Terminal and shell-heavy work - Opus 5.5, decisively. 66.4% beats Astra at five times Sol's price.
- High volume, bounded tasks, human review in the loop - Sol. You are paying for capability you will not use.
- Cache-heavy pipeline - run the numbers again. Both cache at $0.20 and the 2x gap may be a 1.3x gap on your actual traffic.
- Cost is the binding constraint - neither. Look at GPT-6 Luna at $0.10 / $0.50 first and find out where it breaks.
- Already integrated with one - the four Opus 5.5 breaking changes make a migration to Anthropic more expensive than the price comparison implies. Factor that in.
Verdict
GPT-6 Sol for volume and for anything with a human in the loop. Claude Opus 5.5 where the agent has to keep going on its own - long sessions, terminal work, tasks where the cost of an incomplete run is higher than the cost of the tokens. The 2x price gap is real, but so is the seven-point gap in how often the work actually finishes.