Opus 5.5, GPT-6 Astra or GPT-6.1 Sol: Which Model to Build Your Company's Software With

What is on the table

First, a note on names: “GPT-6.1” on its own does not exist. OpenAI launched GPT-6.1 Sol on September 29, an upgrade to its mid-range lineup, and withdrew GPT-6.1 Astra, which it had ready, before launch over safety concerns. GPT-6 Astra is still its top model.

Prices and context, from Anthropic and OpenAI; the index, from Artificial Analysis with each model at its maximum effort.

Opus 5.5 tops the Artificial Analysis index, but GPT-6.1 Sol costs half as much per token, and Astra two and a half times what Opus costs. That alone suggests there is no single winner for everything.

The short answer

Cost per task matters more than a benchmark

Price per token is misleading, because each model spends a very different number of tokens on the same work. What matters is how much a finished task costs. Artificial Analysis measures this with FrontierCode 1.1, a test that checks whether a project's maintainers would accept the code the model writes:

Cost per task on FrontierCode 1.1, with the percentage of tasks accepted

• Claude Opus 5.5 (52%): $10.15
• GPT-6 Astra (47%): $4.22
• GPT-6.1 Sol (49%): $0.81

Artificial Analysis, FrontierCode 1.1.

Opus 5.5 gets three more tasks right out of every hundred than GPT-6.1 Sol, and costs about twelve times as much per task. Part of the gap is how much it writes: to complete the full Artificial Analysis index, Opus generated about 260 million output tokens, compared with about 60 million for Astra and 67 million for Sol. And it thinks longer before answering: at maximum effort, its first token took 652 seconds on average, versus 278 seconds for Astra.

For a team, that means two things. In an agent that makes hundreds of small changes a day, GPT-6.1 Sol gives almost the same quality for a fraction of the price. In the change that touches twenty files and cannot go wrong, three points of accuracy are worth more than ten dollars.

By type of work

All the figures in the table, from Vals AI.

The table does not say one is a better programmer than another. It says each has its own strengths, and that splitting the work between them works out better and cheaper than committing to just one. That is why, in the agent projects we build, the model is a part chosen per task, not a brand decision.

The surprise: Claude Sonnet 5.5

Anthropic released Sonnet 5.5 on September 28, six days after Opus 5.5, at $2 and $10 per million tokens, the same price as GPT-6.1 Sol. On Vals AI's overall index it ranks first, with 70.85% versus 70.18% for Opus 5.5, and on its Terminal-Bench 4.0 it trails Opus by one point (64.14% versus 65.15%).

If we started a project from scratch today, Sonnet 5.5 would be our first candidate for daily work, we would reserve Opus 5.5 for the hardest tasks, and GPT-6.1 Sol would take the volume. That is still before measuring it on each client's code, which is what really decides.

Your company's data

For a European company, where data is processed and how long it is kept matter as much as code quality.

• Zero retention: Opus 5.5 can be used without Anthropic storing the data. With OpenAI, zero retention has to be requested and they approve it, and by default they keep logs for up to 30 days to monitor abuse.
• EU residency: OpenAI's own API offers processing GPT-6.1 Sol in the EU. Anthropic's API only offers the US or global; for Claude in Europe you have to go to the regional endpoints of its partner clouds, with a 10% surcharge.
• Clouds: Opus 5.5 is on AWS, Google Cloud and Microsoft. The GPT-6 models are on Amazon Bedrock and Microsoft Foundry, and we have found no evidence that they are on Google Cloud.
• Security: OpenAI classifies Astra as critical risk in cybersecurity and refuses advanced cybersecurity tasks, which matters if your team builds security tools.

None of these things show up in a benchmark, and any of them can rule out a model before you test it. It is the first thing we check in AI consulting work.

What the tables don't tell you

• Each lab tests its models in its own environment (Claude Code, Codex) and at the effort that suits it: Anthropic compares Opus at very high effort with Astra at high. That is why here we only use figures from third parties that test everyone the same way.
• SWE-bench Verified, the classic programming test, is no longer useful for comparison: OpenAI stopped publishing it, and Vals AI stopped running it on new models in September.
• In some cybersecurity and biology tasks, Anthropic has an earlier model answer, for safety reasons, and warns that this will probably lower Opus 5.5's scores.
• Opus 5.5 brings changes that break integrations: its reasoning cannot be turned off, and asking it to use a specific tool with tool_choice returns a 400 error. If you are coming from Opus 5, check your code before changing the model name.

How we decide it ourselves

1. We gather between twenty and thirty real tasks from the project: closed tickets, bugs already fixed and a few large changes, each with its good solution next to it.
2. We run the candidates through those tasks with the same environment and the same effort, and count how many solutions a team member would accept without changing them.
3. We divide what each model has cost by the accepted tasks: the cost per good task, not per token.
4. We split the work: one model by default, another for the hard work and another for volume, behind a layer that lets us switch providers without touching the product.
5. We repeat this when a new model comes out. In October, several have.

That is how we put together the AI agents and the custom software we build, with the model as a piece that can be changed.

The other models from October

The three in this article are not alone. These are the ones that come up in the conversation, with their place on the Artificial Analysis index (v4.3.2) so you can compare them with the ones above:

• Claude Sonnet 5.5 (Anthropic, Sep 28): 56. At $2 and $10, Anthropic's best quality for the price.
• Kimi K3.1 Pro (Moonshot AI, Oct 23): 55, fourth. At $2 and $10, with the weights promised for early November.
• Gemini 4 Argon (Google, available since Oct 13): 53. At $2 and $10 until November 30, and at $4 and $20 after that.
• Claude Fable 5.1 (Anthropic, Sep 1): 53. At $10 and $50, and without zero retention.
• Grok 4.7 (xAI, Sep 21): 46, with a 500,000-token context.
• MiMo-V2.6-Pro (Xiaomi, Sep 21): 46, the best open-weights model, with an MIT license.
• DeepSeek V4.1 Flash (Oct 17) and GLM-5.3 (Z AI, Aug 18): 45, both open-weights.

A warning about the vendors' figures: on Terminal-Bench, some are well above the independent ones. Xiaomi publishes 89.9% for MiMo-V2.6-Pro on version 2.1, and Vals AI measures it at 67.79%. To choose one, trust the independent figures or, better, your own test.