Claude Opus 5.5 vs GPT-6 Astra: the fight is no longer about who answers better

September 2026 left us with a pretty good comparison for understanding where artificial intelligence is heading.

OpenAI unveiled GPT-6 Astra on September 3. Nineteen days later, Anthropic responded with Claude Opus 5.5.

After reviewing both launches, what interests me most isn't deciding which one answers a question better. That already feels like too basic a comparison.

The battle is now somewhere else: how much work you can delegate to a model, how long it can keep doing it, and how much it costs you to get the task finished.

Two models built to work

GPT-6 Astra is the most capable model OpenAI currently offers for complex end-to-end work.

It has a context window of 1.05 million tokens and allows up to 128,000 output tokens. OpenAI positions it for complex reasoning, programming, research, computer use, and document creation.

Claude Opus 5.5 arrives with a similar philosophy, though Anthropic is putting a lot of emphasis on efficiency.

The company says the new model needs fewer steps and fewer tokens to solve many tasks than Opus 5.

That seems like a technical detail until we leave an agent running for hours.

Then every tool call, every reasoning step, and every token starts showing up on the bill.

The interesting numbers

Comparing benchmarks across vendors always requires caution because configurations, tools, and reasoning levels can change.

Even so, there are tests where we have comparable results.

Source: Anthropic, Claude Opus 5.5 and OpenAI, GPT-6 Astra.

I don't see one model crushing the other in absolutely everything here.

I see specialization.

Claude has especially interesting results in agentic programming and professional work. Astra shows a lot of strength in scientific research, automation, and other tasks where the agent needs to interact with tools and environments.

And that seems more interesting to me than declaring a universal winner.

Claude did a migration of 680,000 lines

Anthropic showed a case that really caught my attention.

One of its evaluators used Opus 5.5 to complete a migration across 680,000 lines of code in less than a day.

Another evaluator left it working for more than 18 hours unsupervised on a task involving six repositories.

They also ran a test on a codebase of roughly 200,000 lines. Opus 5.5 completed the audit and the fixes in under three hours; Opus 5 needed more than 20.

These are tests provided by Anthropic and its evaluators, so I'd take them as examples of capability and not as a guarantee that we'll get those same results on any project.

But they show perfectly where this is heading.

Not long ago we evaluated an AI by asking:

Can it write this function?

Now we're starting to ask:

Can it handle this entire migration while I do something else?

That shift is huge.

Price may end up mattering more than the benchmark

Here Anthropic made an aggressive move.

Claude Opus 5.5 costs:

• $4 USD per million input tokens.
• $20 per million output tokens.
• $0.20 per million tokens read from cache.

GPT-6 Astra costs:

• $10 USD per million input tokens.
• $50 per million output tokens.
• $1 per million input tokens stored in cache.

On top of that, OpenAI charges higher rates when a request exceeds 272,000 input tokens.

On paper, Astra costs 2.5 times more per input and output token.

But we also can't just multiply tokens and declare a winner.

A more expensive model that finishes a task using fewer attempts could end up costing less.

That's why I much prefer a metric we'll see more and more of: cost per successfully completed task.

Anthropic is already using exactly that argument. According to its own measurements, Opus 5.5 can match or beat Astra's results on certain coding tests using a fraction of the cost per task.

That will need to be verified with real workloads.

OpenAI has a fairly serious card: Astra

It would be a mistake to read Anthropic's launch as if Astra had become outdated three weeks later.

There are capabilities in OpenAI's model that really stand out.

One of them is scientific research.

On Terminal-Bench Science 0.1, Astra reaches 64.6%, compared to the 58.7% reported for Opus 5.5.

OpenAI also reports strong advances in computer use, browsing, math, and problem-solving in unfamiliar environments.

But there's an even more striking data point.

GPT-6 Astra is the first model that OpenAI classifies at the Critical level for cybersecurity capabilities within its Preparedness Framework.

That means the company believes that, with the right tools and access, the model can discover previously unknown vulnerabilities and develop ways to exploit them in protected systems without needing human instructions for every step.

That's impressive.

It also explains why OpenAI is putting so many restrictions around the model.

Anthropic is also tightening security

And here an interesting coincidence appears between the two companies.

Anthropic claims that Opus 5.5 is less prone to executing actions that are hard to reverse or going beyond the limits set by the user.

It also improved its resistance against prompt injection attacks.

There's a fairly obvious reason.

Previous models mainly talked.

These models act.

They can use browsers, terminals, applications, files, and external services.

When a model has that level of access, a mistake stops being just an incorrect answer.

It can become an incorrect action.

That's why I think we'll see more and more work around permissions, isolation, monitoring, and limits for agents.

Something we usually ignore also changed: communication

Anthropic acknowledges that one of the criticisms of Opus 5 was its way of communicating.

Opus 5.5 tries to put the important information first, use less jargon, and produce responses that are easier to review during long sessions.

It may seem cosmetic.

To me it isn't.

If I'm going to leave an agent working for four, ten, or eighteen hours, I need to quickly understand what it did, what it changed, and where it ran into problems.

Technical capability is of little use if I need half an hour to decipher the agent's report.

This changes how I would choose a model

A year ago it made sense to compare models mainly by intelligence.

Now I would add several questions.

How much does it cost to finish my task?

Not how much a million tokens costs.

How much it costs to solve the whole problem.

How long can it work without intervention?

An excellent model that needs confirmation every three minutes can end up being less useful than another slightly less capable one that correctly completes a two-hour task.

What tools can it handle?

Browser, terminal, documents, APIs, and applications are becoming part of the practical model we use.

What happens when something goes wrong?

For me this will be one of the most important questions of 2027.

An intelligent agent also needs to know when to stop.

And Sonnet 5.5 Is Still to Come

Anthropic wasn't done with this launch either.

The company confirmed that Claude Sonnet 5.5 and Claude Haiku 5.5 will arrive in the coming weeks.

That could be even more important for many developers than Opus.

Mid-tier models often end up being the ones we use in production because they offer a more reasonable combination of speed, capability, and cost.

If Sonnet 5.5 inherits a good portion of Opus 5.5's efficiency improvements, we could have an even more interesting comparison against OpenAI's more economical models.

My Takeaways

This generation leaves me with four clear things.

• AI programming is moving from code generation to project delegation.
• Cost per completed task is starting to matter more than the isolated price per token.
• Agents can work for increasingly longer periods without supervision.
• Security and permissions are starting to become a fundamental part of the architecture.

For those of us who develop software, the last one seems especially important.

We're starting to give AI access to the same tools we use ourselves.

Conclusion

GPT-6 Astra and Claude Opus 5.5 show two interesting approaches to the same direction.

OpenAI is pushing frontier capabilities very hard: science, computer use, research, cybersecurity, and complete complex tasks.

Anthropic is attacking capability and efficiency simultaneously, trying to make an agent able to work longer using fewer steps, tokens, and money.

Benchmarks will keep changing every few weeks.

What does seem quite a bit more stable is the direction.

We're no longer comparing assistants that write code. We're starting to compare digital workers capable of receiving a task, using tools, and spending hours trying to finish it.

For me, that's the important shift in these launches.

And before choosing one for production, I'd test both with a real project task and measure three things: correct result, total time, and final cost.