Claude Opus 5.5 and GPT-6 Sol/Luna Arrive: How Should You Use Them in Claude Code and Codex?

September 23, 2026Updated: September 23, 2026Articles

On September 22, 2026, Anthropic and OpenAI announced Claude Opus 5.5 and GPT-6 Sol and GPT-6 Luna, respectively. As someone who uses both Claude Code and Codex, I am less interested in which model tops a benchmark than in which parts of my everyday coding work I should give to each one.

All three models arrived on the same day, but they are designed for different roles. Opus 5.5 and GPT-6 Sol are intended for complex coding and longer-running work. GPT-6 Luna is better suited to repeatedly handling clearly defined tasks at a lower price.

Claude Opus 5.5 brings Fable 5.1-level performance to the Opus price tier

Claude Opus 5.5 is the first model in the Claude 5.5 family. Anthropic says it performs at roughly the same level as Claude Fable 5.1 on many tasks, and highlights long-running coding agents and knowledge work as its main uses.

API pricing is $4 per million input tokens and $20 per million output tokens. That is a 20% reduction from Opus 5’s $5 and $25 rates. Cache reads have also fallen from $0.50 to $0.20 per million tokens.

Anthropic says that, after accounting for the model completing work with fewer tokens, a typical task costs about 40% less than with Opus 5 and output is more than 30% faster. These figures come from Anthropic’s own evaluation, so the actual cost and completion time still need to be checked with your own workload.

The context window is 1 million tokens, and the maximum output is 128,000 tokens. Thinking is always enabled, and the default effort setting is now medium. If you use the Opus 5 API, there are breaking changes involving settings that disable thinking and the way tools are specified. Rather than changing only the model name, it is worth checking the migration guide as well.

Anthropic also announced higher five-hour usage limits for Claude Pro, Max, and Team. However, the announcement does not specify the exact increase.

For more about Anthropic’s earlier high-end model, see our introduction to Claude Fable 5.1.

GPT-6 Sol and Luna give Codex users more choices

OpenAI released GPT-6 Sol and GPT-6 Luna as additions to the lineup that already includes GPT-6 Astra. Both are reasoning models that accept text and images as input and produce text as output. They support the Responses API and Chat Completions API.

GPT-6 Sol is intended for complex coding and work involving AI agents. OpenAI positions it as an everyday model for “writing, coding, and work that needs judgment.” The company says it improves on GPT-5.6 Sol in coding, factual accuracy, and communication.

GPT-6 Luna is designed for clearly scoped tasks and high-volume, repetitive processing. Examples include summarization, information extraction, and small code fixes—jobs where the instructions and finish line are easy to define.

Both models are rolling out in Codex and ChatGPT Work. They are available to Plus, Pro, Business, Enterprise, and Edu users, while Free and Go users can access Luna from the desktop app. They cannot be selected in ordinary ChatGPT chats; they are models for Work or Codex. In the Codex CLI, choose one with /model, or specify it with a command such as codex --model gpt-6-sol or codex --model gpt-6-luna.

For how these models differ from the top-tier option, see our introduction to GPT-6 Astra.

Comparing the specifications of Opus 5.5, Sol, and Luna

Model	Main use	Context window	Maximum output	API input/output
Claude Opus 5.5	Long-running coding and knowledge work	1 million	128,000	$4/$20
GPT-6 Sol	Complex coding and agentic work	1.05 million	128,000	$2/$10
GPT-6 Luna	Clearly scoped tasks and high-volume processing	1.05 million	128,000	$0.10/$0.50

API prices are standard rates per million tokens. For GPT-6 Sol and Luna, when input exceeds 272,000 tokens, input and cached-input pricing for the entire request doubles, while output pricing increases by 1.5 times. Claude Opus 5.5 keeps its standard pricing up to 1 million tokens. Subscription users are subject to their plan’s usage limits rather than API pricing.

Looking only at token prices, Luna is far less expensive, while Opus 5.5 costs twice as much as Sol. However, models can require different amounts of output and different numbers of revisions to finish the same job. The API price table alone cannot tell you the total cost of completing a task.

Fable 5.1 and Astra have not been replaced

At Anthropic, Fable 5.1 remains the high-end model for difficult, long-running work. At OpenAI, GPT-6 Astra remains the top-tier model for complex analysis and large deliverables.

What this release changes is that you no longer need to choose the most expensive model every time to get strong performance. In Claude, Opus 5.5 can handle everyday complex work; in Codex, Sol can fill that role. You can then move only the hardest requests to Fable 5.1 or Astra. Simple extraction and small fixes can go to Luna, helping you keep working while using less of your allowance.

There is no official head-to-head comparison of the three models yet

As of September 23, there is no official benchmark comparing Opus 5.5 with GPT-6 Sol and Luna under the same conditions. Anthropic’s announcement includes comparisons with GPT-6 Astra and GPT-5.6 Sol, but not with GPT-6 Sol and Luna, which launched on the same day.

Benchmarks from different companies may use different tools, reasoning settings, and numbers of trials. Rather than placing the scores side by side, I think it is more useful to give the models the same request in the same repository and compare whether they finish, how many corrections they need, and how much time and usage the work consumes.

How I want to use the three models

First, I want to give both Opus 5.5 and GPT-6 Sol tasks such as adding a WordPress feature across multiple files or producing an article that requires lengthy research. I will compare not only the finished result, but also how often each model needs clarification and how much I have to correct myself.

For GPT-6 Luna, I want to assign work whose scope can be made explicit: finding related files, converting content into a fixed format, or classifying test results. It may also be useful to let Sol handle the larger job while Luna performs smaller checks in parallel.

There is no need to commit to either Opus 5.5 or Sol exclusively. I plan to try the same work in Claude Code and Codex, then choose between them based on how few corrections they need, how long they take to finish, and how quickly they consume my usage limits.

Summary

Claude Opus 5.5 arrives as a model that offers performance close to Fable 5.1 at a lower price than Opus 5. GPT-6 Sol is aimed at complex implementation work in Codex, while GPT-6 Luna is for low-cost, repetitive tasks.

There is no direct comparison of all three models yet. I will start by giving Opus 5.5 and Sol the same complex job, while assigning Luna smaller tasks with a clearly defined scope.