Last Updated: 23 September 2026
September 22 turned into a busy day for AI. OpenAI introduced GPT 6 Sol and Luna, bringing lower prices and new pressure to an already crowded market. On that same day, Claude Opus 5.5 also arrived. Its focus is easy to understand: give users a more capable premium model without making long and demanding tasks too expensive to run.
The new version costs less than Opus 5 and can use fewer tokens for some jobs. It is also built for coding, research and tasks that may continue through several steps. The timing makes the release even more interesting. GPT 6 immediately adds competition on price, but Opus 5.5 is mainly about how much more Anthropic can offer without keeping the old premium cost.
Claude Opus 5.5 is cheaper to run than the previous version, with lower prices for both input and output. For people who use the model occasionally, the difference may not feel dramatic. For developers or teams running it across many tasks, the savings can become more noticeable over time, especially when those jobs involve long conversations or repeated requests.
The cost can also fall because the model is designed to use fewer tokens for some types of work. That means a task may require less output before it is finished. Anthropic estimates that comparable workloads can cost around 40 percent less overall, although the final amount will always depend on what the model is asked to do.
This becomes more important during longer jobs. A coding agent may need to inspect files, make changes and check the result before it is done. Each extra step adds to the total cost. If Opus 5.5 can complete the same work with fewer responses, the savings may be more visible in the final bill than in the basic price list.
GPT 6 Sol still comes in at a lower standard API price. That adds pressure from outside, but Opus 5.5 already marks a meaningful drop in what Anthropic charges for its premium model.
Premium AI can become expensive quickly once a task moves beyond a short conversation. Claude Opus 5.5 lowers that cost and makes longer use a little easier to justify. The new rates are below those of Opus 5, so developers and teams can run more work through the model before spending the same amount. That becomes more useful as tasks grow longer and involve more repeated calls.
Opus 5.5 is also designed to get through some jobs with fewer tokens. That can reduce the final cost even when the task itself has not changed. Anthropic estimates that comparable workloads may be around 40% cheaper overall. You will not see the same saving every time, but needing less output can make a noticeable difference when the work continues for many steps.
This is particularly useful for agent style tasks. The model may need to inspect a project, return to earlier files and make several decisions before the job is complete. The longer that process runs, the more expensive unnecessary output becomes. If Opus 5.5 can stay focused and finish with fewer responses, the lower cost becomes easier to notice in everyday use.
There is still a cheaper rival in the background. GPT 6 Sol arrived with lower standard API rates on the same day. Even so, Opus 5.5 does not need that comparison to make its case. The model is simply less expensive to use than before, which gives existing Claude users more freedom to keep it on longer and more demanding jobs.
A lot of the work behind Opus 5.5 is focused on what happens after the first coding answer. Real software projects rarely end there. The model may need to open another file, check how a change affects the rest of the project and return to something it touched earlier. Staying useful through that entire process is one of the areas Anthropic has tried to improve.
The scale of the testing reflects that approach. One example used a project containing around 200,000 lines of code, which is far removed from a short programming exercise. At that size, the model needs to follow connections across files and keep track of what it has already done. Losing that context can easily turn one small mistake into several more.
For developers, the difference will come down to how often they still need to step in. A model that produces good code but constantly needs reminders can quickly become tiring to use. Opus 5.5 is meant to stay focused for longer, which could reduce repeated instructions and make it more useful when a programming task takes several rounds to finish.
Imagine giving an AI a job and checking back later instead of guiding every step yourself. That is where longer agent tasks are heading. The system may need to open files, use tools and make several decisions before producing anything final. During that process, it has to remember the original request and keep track of what it has already done along the way.
That makes reliability harder to judge than it is during a short chat. One mistake halfway through a task can affect several decisions that follow. The model may also need to notice when new information conflicts with something it decided earlier. Longer work depends heavily on keeping context and knowing when to stop before a problem grows into something much larger.
Anthropic has built more protections around these situations. The system can work inside a sandbox, and individual actions can be checked before they run. Those controls give users more visibility over what happens during longer jobs, particularly when an agent is interacting with files or tools without someone approving every action manually as the process continues.
This kind of work moves AI beyond quick answers and toward tasks that continue across several steps with less direct guidance overall.
Longer AI tasks create a different kind of safety problem. A mistake may happen several steps before anyone notices it, especially when the system is using tools or working with outside files. Anthropic is trying to reduce that risk by adding checks around what the model can do during a task. Some actions can be reviewed before they happen, and the working environment can also be inspected.
One of the main concerns is prompt injection. This happens when hidden instructions inside a file or website try to push the AI away from what the user originally asked. The problem becomes more serious when an agent has access to tools or important information. Stronger protection against these attacks is included in the new release, with more attention on tasks that involve outside content.
Independent groups also looked at the system before it became public. METR and Frontier Design were among those involved in testing, and Reuters reports that protections used for Anthropic’s more capable models are present here too. These checks matter more as AI moves beyond simple chat and starts handling jobs where it can make several decisions without asking for approval each time.
Premium AI is starting to look less exclusive. The newest release lowers the cost of using Anthropic’s higher end model and brings it closer to systems that sit above it in the lineup. That gives users more room to question whether paying extra is really necessary for every difficult task.
The next releases could push that even further. Sonnet 5.5 and Haiku 5.5 are expected to arrive soon, and similar gains could make lower tiers more attractive for everyday professional work. If cheaper models keep improving at this pace, the most expensive option may end up being reserved for a much smaller group of demanding jobs.
OpenAI is creating similar pressure with GPT 6 Sol and Luna. Both arrived with aggressive pricing, which makes lower cost a wider industry trend rather than something happening inside one company. Users now have more choices at different price levels without giving up as much capability as before.
Claude Opus 5.5 gives Anthropic a stronger premium model without keeping the same cost as before. It is built for coding, longer agent tasks and work that needs more independence from the system. The release also adds more controls around those longer jobs, which becomes important when AI is allowed to make several decisions before the user checks the final result.
GPT 6 adds pressure on pricing, but the bigger story is happening inside Anthropic’s own lineup. Better performance is moving into cheaper tiers much faster than before. That makes premium AI harder to define and gives users more reasons to choose a model based on the job itself instead of automatically paying for the highest option available.
Claude Opus 5.5 is Anthropic’s newest premium AI model, built for coding, research and longer tasks. It costs less than the previous version and is designed to stay useful across jobs that involve several steps. The model can also use fewer tokens for some workloads, which may reduce the final cost. Anthropic has added more controls for longer agent tasks as well, especially when the system needs to work with files, tools or outside information.
Claude Opus 5.5 costs less than its predecessor, with lower prices for both input and output tokens. The current rates are $4 per million input tokens and $20 per million output tokens. That is only part of the cost change, since the model may also use fewer tokens to finish certain tasks. For longer coding or research jobs, that can reduce the final bill further. The exact saving will vary depending on the workload and how many steps are needed.
Yes, Claude Opus 5.5 is built for coding tasks that continue across multiple files and several steps. Anthropic has focused heavily on longer programming work rather than only short code generation. The model has been tested on large codebases and tasks where it needs to remember previous changes before continuing. That makes it more useful for software work that involves reading, editing and checking code over time. Developers will still need to review results, especially on larger or more sensitive projects.
No, GPT 6 Sol currently costs less per token than Claude Opus 5.5 at standard API rates. The difference is clear on the pricing page, but actual task costs depend on more than the listed number. One model may produce more output, require extra retries or take more calls to finish the same work. That can change the final result. For coding or long agent tasks, developers should compare complete workloads rather than assuming the cheaper token price always means a cheaper finished job.
Claude Opus 5.5 brings lower costs, longer task support and more safety controls than the version before it. The model is designed to handle coding and agent work with less interruption while also using fewer tokens on some jobs. Those changes can reduce the final cost during long sessions. Anthropic has also added protections around tool use and prompt injection, which become more relevant when the system is allowed to work with files or outside information without constant supervision.
