Is Claude Opus 5.5 or GPT-6.1 Sol better for coding?

Claude Opus 5.5 and GPT-6.1 Sol arrived a week apart, on September 22 and September 29, 2026, and they are the models most people meet first in Claude Code and Codex. They are not quite rivals at the same price. Opus 5.5 is Anthropic's main high-end model, the one it calls a daily driver for agentic coding. GPT-6.1 Sol is OpenAI's middle model, which OpenAI pitches as close to its top model, GPT-6 Astra, at a fifth of Astra's price.

So the honest answer has two parts. On capability, the evidence leans toward Opus 5.5. On value, Sol often comes out ahead.

Prices as of October 2026

A token is a small piece of text, about three-quarters of an English word. Coding agents resend your project context again and again, so the cached price matters more than it looks: Anthropic notes that cache reads make up most of the cost of agentic coding.

What the tests say

Each company publishes its own results, and each picks tests that suit its model. Anthropic reports Opus 5.5 at 66.4% on Terminal-Bench 4.0, a test of multi-step command-line work, against 57.9% for GPT-6 Astra; Sol is not in that table. OpenAI, for its part, says Sol matches Astra on DeepSWE 1.1, a test of long software-engineering tasks in real codebases, at about a fifth of the cost, and that it beats Opus 5.5 by 2.2 points on AutomationBench, a business-workflow test, at medium effort.

An outside view helps. Google's Android Bench runs each model in its usual coding agent on 30 long Android development tasks. As of October 26, it shows:

Opus solves more tasks, but the two ranges overlap widely, because 30 tasks is a small sample. The cost gap is not small: Sol finished the whole run for less than a tenth of Opus's bill.

A worked cost example

Prices per token only tell part of the story, since models also use different numbers of tokens for the same job. Still, a simple calculation shows the scale. Suppose one agent task reads 2 million tokens, of which 1.8 million are cached project context, and writes 60,000 tokens:

• Opus 5.5: 0.2 million × $4 = $0.80, plus 1.8 million × $0.20 = $0.36, plus 0.06 million × $20 = $1.20. About $2.36.
• GPT-6.1 Sol: 0.2 million × $2 = $0.40, plus 1.8 million × $0.10 = $0.18, plus 0.06 million × $10 = $0.60. About $1.18.

These figures are illustrative and leave out cache-write fees. Run fifty such tasks a week and the gap becomes about $59 a week. If Opus needs fewer retries, some of that comes back; if Sol finishes the job first time, none of it does.

When is the cheaper model the smarter choice?

Sol is the smarter pick when the work is well defined and repeated. Fixing a failing test, adding an endpoint that follows an existing pattern, writing documentation from code, updating dependencies, or reviewing small pull requests: here the gap in capability rarely shows, and paying half as much per token, or a tenth as much per long run, adds up fast. The same holds when an agent loops over many files or runs on a schedule, since every extra step multiplies the price.

Opus 5.5 earns its price on long, tangled work: migrations across a whole codebase, bugs that span several services, or changes where a wrong guess costs you hours of review. Anthropic's examples are of this kind, such as a 200,000-line codebase audited and fixed in under three hours. When the cost of your own time fixing a weaker result is higher than the token bill, the stronger model is the cheaper one.

A practical rule: start every task on the cheaper model at default effort. Move up to Opus only when the first attempt fails your tests or misses the point twice. Keep a short log of which tasks needed the upgrade; after a few weeks it tells you exactly where your money is best spent.