Astra, Opus 5.5, and what comes next

In my fall coding-model rankings, Astra took intelligence and wide problem handling. Grok kept usage. I wanted Astra for the hard jobs, then wanted off it when I looked at the bill.

Opus 5.5 makes that arrangement worth another look. Anthropic is promising work around Fable 5.1’s level for less money than Opus 5. If that holds up in my repos, I can spend longer on the difficult jobs before the cost makes me switch models.

September’s releases

OpenAI’s API changelog dates GPT-6 Astra to September 3, 2026. It covers difficult work across coding, reasoning, computer use, research, and documents. The release also added asynchronous tool calls and the ability to steer a response while it is running. You can change reasoning effort during a conversation too.

That steering is useful. On a long job, I want to correct the agent when I spot a bad assumption. Waiting for it to finish working from that assumption wastes time and tokens.

Anthropic announced Opus 5.5 on September 22. It says the model performs at Fable 5.1’s level on most work and costs 40% less than Opus 5 on typical workloads in its tests. Launch rates are $4 per million input tokens, $20 per million output tokens, and $0.20 per million cache-read tokens. Anthropic attributes the savings to lower rates and fewer tokens per task.

Those results give me a reason to try Opus on the work where I reach for Astra. I want to run that comparison before changing the rankings.

Astra holds a wider job together

A change to one function is usually easy enough to check. Once a change crosses the client and server, you can get a patch that looks reasonable in every file and still breaks how they work together. Shared state makes that particularly annoying.

That is where Astra earned its place in September. It held more of the problem together, and I spent less time stitching its changes back into the codebase. I became willing to hand it larger jobs.

I still have to check the result. And I still have to pay for it. Astra’s current model page lists standard rates of $10 per million input tokens and $50 per million output tokens, with higher rates for prompts above 272,000 input tokens.

So I use it selectively. There are plenty of jobs where I would like that level of help for longer than I am willing to pay for it.

What does a finished change cost?

This is where I want to test Opus 5.5. A lower token price is easy to compare. The cost of getting a change finished takes a whole session to find out.

An agent can generate a fix in ten minutes and leave me with an afternoon of repairs. Those repairs belong in the cost. So does the time I spend explaining the same constraint again after it loses the thread.

If Opus needs fewer corrections, the savings could be useful beyond what the token rates suggest. I could afford another pass on a feature after it first works, or check an interaction I would otherwise leave until later. That is what I want to measure against Anthropic’s launch claims.

Try it on the games

I have two new releases up: Acting, a multiplayer RPG-style test with dialogue, and Cars, a complete car fabricator and multiplayer racing simulator with Box3D physics. There is a short release post for both.

Cars gives me an obvious check: build a vehicle, bring another player into the session, and race it. The car the fabricator produces has to work in the simulation. If it doesn’t, the fact that the code compiled is not much comfort.

With Acting, I would want to see what happens to a conversation when another player joins or leaves. Getting a dialogue window on screen leaves a lot of behavior to check.

These are examples of the work I want an agent to verify. A benchmark score does not tell me whether either model can do it in my project. Giving it access to the running game and checking what it does would tell me more.

Then come back and fix it

I expect better agents to make more small projects worth building. A game can have a small audience and still be worth the effort if getting it working takes less time.

I am less sure how that plays out once there are ten of them to maintain. A physics fix can change how an existing car drives. A player can find a state I never tested. Shipping another game adds to the work waiting next week.

I want an agent that can come back to an existing project, reproduce a bug, and fix it without turning the job into a rewrite. Then it needs to run the thing and show what happened. That is why I am still interested in operating harnesses: keeping enough context to resume useful work is a problem of its own.

For the next comparison, I want to give Opus the same sort of wide jobs that made me choose Astra. I will count the repairs as well as the initial patch. If the lower cost survives that, it changes how often I can afford to use a model at this level.

Acting and Cars give me somewhere to start.