Claude Opus 5.5 vs GPT-6 Sol vs GPT-6 Luna: which model to choose after September 22, 2026

On September 22, 2026 Anthropic and OpenAI launched their new models less than two hours apart. Claude Opus 5.5, the first model in the new Claude 5.5 family, arrived at 16:31 UTC; shortly after 18:00 OpenAI announced GPT-6 Sol and GPT-6 Luna, which bring the advances of its flagship GPT-6 Astra to much lower price points. Here we compare pricing, claimed performance and cost per task, with one caveat: all the numbers come from the labs themselves and should be verified on your own use cases.

The three models at a glance

Claude Opus 5.5 is the new flagship of Anthropic’s main line. According to the company, it performs at the level of Claude Fable 5.1 on most work, costs about 40% less than Opus 5 on typical workloads and generates output more than 30% faster. It has a 1-million-token context window and up to 128,000 output tokens, and it uses the same safeguards as Fable 5.1 in cybersecurity, biology and distillation: requests that trigger them are passed to an earlier Claude model.

GPT-6 Sol is OpenAI’s model for coding, agents and business process automation. It sits below Astra and inherits its training methods. It has a 1.05-million-token context window, 128,000 output tokens and a reasoning effort setting that ranges from no reasoning up to maximum.

GPT-6 Luna is the small, fast model in the family, designed for tasks repeated thousands of times a day, such as extracting fields from documents or summarizing tickets, with the same context window as Sol.

Pricing compared

Opus 5.5 cut list prices by 20% compared with Opus 5, and cache reads by 60%. OpenAI halved input prices compared with the 5.6 generation. For prompts over 272,000 tokens, OpenAI charges twice the input rate and 1.5 times the output rate, while Anthropic keeps the same price across the full million-token window. Both offer faster modes at a premium: Opus 5.5 in fast mode costs $8 and $40, while OpenAI’s fast mode costs twice the standard rate.

Performance: where Opus 5.5 leads

In its launch charts OpenAI compared Sol and Luna with Claude Opus 5 and Fable 5.1, since Opus 5.5 came out the same day. On two benchmarks reported the same way by both labs, a direct comparison is still possible.

On AutomationBench, which measures complete business workflows with dozens of tools across sales, marketing, support and finance, Opus 5.5 scores 40.0% against 33.2% for Sol and 20.7% for Luna, close to GPT-6 Astra’s 41.4%. On FrontierCode 1.1, which grades agent-written code on quality and ease of merging as well as correctness, Opus 5.5 reaches 54.4% against 49.3% for Sol and 42.4% for Luna, ahead of Astra too (53.3%). Anthropic also reports 66.4% on Terminal-Bench 4.0, above Astra’s 57.9%; OpenAI has not published this figure for Sol and Luna.

On peak scores, then, Opus 5.5 is the strongest of the three, with a five-to-seven-point margin over Sol. Anthropic itself warns, however, that at these capability levels benchmark margins reflect real-world differences less and less.

OpenAI’s argument: cost per task

OpenAI presented every score against the average cost per task, at each reasoning effort level. Read that way, Sol is very competitive: on AutomationBench it reaches 33.2% at about $0.27 per task, beating Opus 5 at maximum effort, which costs more than eleven times as much, and on Agents’ Last Exam it reaches 56.4%, above Opus 5’s best result.

There are also some steps back. On DeepSWE v1.1 and OSWorld 2.0, GPT-6 Sol’s peak score (68.8% and 64.4%) is lower than GPT-5.6 Sol’s (72.7% and 66.2%): on these tests the new generation’s advantage lies in cost per task, which drops by about 58%. Progress on reliability, on the other hand, is clear: in OpenAI’s internal factuality evaluation, Sol makes about half as many errors as its predecessor.

Luna, the surprise at the low end

The most interesting result of September 22 may concern the smallest model. At maximum effort GPT-6 Luna scores 66.6% on DeepSWE at about $0.22 per task and 50.9% on Agents’ Last Exam at about $0.15, not far from models that cost dozens of times more. At low reasoning effort, however, Luna is almost unusable for agentic work (2.4% on DeepSWE) and only becomes competitive from medium effort upward. The effort level should therefore be tested on your own workload.

What an agentic session really costs

To get an order of magnitude, we calculated the list-price cost of a typical agent session: 2 million input tokens, 90% of them served from cache, and 150,000 output tokens. With GPT-6 Luna the session costs about $0.11, with GPT-6 Sol $2.26, with Claude Opus 5.5 $4.16 and with GPT-6 Astra $11.30. The calculation only considers per-token pricing: a more capable model can finish the same job with fewer tokens and fewer steps.

Alternatives worth watching

Claude Sonnet 5 has the same list price as GPT-6 Sol ($2 and $10), and Anthropic has announced Sonnet 5.5 and Haiku 5.5 for the coming weeks. Gemini 3.8 Flash costs $0.75 and $3.75 until December 31, after which rates double. xAI’s Grok 4.7 costs $2 input and $6 output, with a 500,000-token context window and a claimed 71.0% on DeepSWE.

Which one to choose

For high-volume tasks with a clear goal, such as classification, data extraction, routing and summarization, GPT-6 Luna at medium or high effort is hard to beat on price today. As the default model for agents and process automation, GPT-6 Sol offers the best balance between results and cost on the published benchmarks. If you need the highest scores in agentic coding and knowledge work, and the budget allows it, Claude Opus 5.5 is the model to try first, especially with very long prompts, where Anthropic’s flat pricing narrows OpenAI’s price advantage. GPT-6 Astra remains OpenAI’s option for the hardest tasks, particularly computer use, at a per-token cost up to five times higher.

In many cases the right answer includes more than one model. Most real products combine simple, repetitive requests with complex steps, and using a single model for everything means overpaying for the former or getting too little on the latter. Routing each request to the right model, and being able to switch providers at the next launch without rewriting the application, is exactly what we want to make simple with ModelHive.