Claude Opus 5.5 and GPT-6 Sol launched within 90 minutes of each other on September 22, 2026. Opus 5.5 costs 4 dollars per million input tokens and 20 per million output. GPT-6 Sol costs 2 and 10, exactly half. Neither is straightforwardly better: Opus 5.5 aims at the top of the capability range, Sol aims at the best result per dollar. Which one fits depends almost entirely on whether your bottleneck is capability or cost.
How do the prices compare?
- Claude Opus 5.5: 4 dollars input / 20 dollars output per million tokens. Cache write 5 dollars, cache read 0.20 dollars.
- GPT-6 Sol: 2 dollars input / 10 dollars output per million tokens. Cached input 0.20 dollars.
- GPT-6 Luna: 0.10 dollars input / 0.50 dollars output, for high-volume work that does not need either model’s reasoning.
Raw rates understate the picture on both sides. Anthropic claims typical workloads cost about 40 percent less than Opus 5 overall, because Opus 5.5 also uses fewer tokens to finish a task, so the effective gap is narrower than 2x. OpenAI’s cost-per-task charts similarly show Sol finishing work far cheaper than its own predecessor. Per-task cost, not per-token price, is the number that actually decides your bill.
Which one performs better?
On the vendors’ own numbers, Opus 5.5 sits higher. Anthropic reports it surpassing Claude Fable 5.1 on agentic coding, knowledge work, computer use, visual chart recognition, and multidisciplinary reasoning, and calls it the strongest model it has tested. OpenAI, by contrast, states plainly on its own launch page that GPT-6 Astra remains its best model across the board, positioning Sol as a cost-efficiency release rather than a capability ceiling.
The detail worth knowing: on two of six headline evaluations, GPT-6 Sol’s best score is actually lower than GPT-5.6 Sol’s best. On DeepSWE it scores 68.8 percent against 72.7 percent, and on OSWorld 2.0 64.4 percent against 66.2 percent. It is cheaper per task, not uniformly stronger. Where Sol does win is cost-adjusted: on AutomationBench it scores 33.2 percent at 0.27 dollars per task, beating Claude Opus 5 at max effort (26.9 percent) at roughly one eleventh the cost.
What about context and speed?
GPT-6 Sol carries a 1.05 million-token context window with input capped at 922K and up to 128K output tokens. Anthropic reports Opus 5.5 generating output more than 30 percent faster than Opus 5. Both support the standard agent tooling: function calling, web search, file search, computer use, and MCP connections on OpenAI’s side.
So which should you actually pick?
Choose Opus 5.5 if your work is long-running coding agents, complex refactors, or knowledge tasks where a wrong answer costs more than the tokens do. The 680,000-line code migration Anthropic cites as an early tester result is the shape of problem it is built for.
Choose GPT-6 Sol if you are running high volumes of moderately hard work and cost per task is what determines whether the product is viable. Choose GPT-6 Luna for classification, extraction, and routing, where 0.10 dollars per million input tokens changes what is economically possible at all.
Consider neither if your workload is mostly routine. The enterprise pattern documented in our report on AT&T routing 40 percent of its AI traffic to open models found a 56 percent cost cut for a 2 percent quality drop, and DeepSeek V4.1 Flash undercuts both on cached input. Routing by task rather than standardising on one model is usually the cheaper answer.
The caveat that applies to both
Every number above comes from the vendor that sells the model. Independent benchmarking has not caught up with either release, and as our guide to reading AI model rankings explains, vendor-selected comparisons routinely flatter the vendor. One example from this launch: OpenAI’s claim that Sol matches Claude Fable 5.1 at much lower cost uses Fable’s weakest setting on that chart. Run both against your real workload for a week before committing. Anthropic’s Opus 5.5 announcement and VentureBeat’s GPT-6 coverage hold the source figures. For the full picture see our coverage of the Opus 5.5 launch and the GPT-6 Sol and Luna release.
