Opus 5.5 vs GPT-6 Sol: I ran both

Issue #16 · September 26, 2026 · ~5 min read

Welcome back to Zero to MVP Weekly. One signal, one tool, one MVP idea — five minutes, every Saturday. No link dumps. Two weeks of this newsletter were about agents doing things nobody asked for. This week is about the ones you pay for: two labs shipped new everyday models on the same afternoon, and I ran both through the same three tasks.

1. The signal of the week

Two everyday models, one afternoon — and the price per token stopped telling you the price.

On September 22, about an hour and a half apart, Anthropic released Claude Opus 5.5 and OpenAI released GPT-6 Sol. Neither is its lab's top model; both are sold as flagship-level work at a mid-tier price.

Opus 5.5: $4 / $20 per million input/output tokens (was $5 / $25), cache reads down from $0.50 to $0.20. Anthropic's claim: 40% cheaper than Opus 5 on typical workloads.
GPT-6 Sol: $2 / $10, half the price of GPT-5.6 Sol.

Each lab benchmarked against the other's previous generation — Anthropic against GPT-5.6 Sol, OpenAI against Opus 5 and Fable 5.1. Not cheating: the rival didn't exist yet. But it leaves the only question that matters unanswered — which of these two?

The independent numbers, from Artificial Analysis: Opus 5.5 at its default effort scores 51 at $1.34 per task; Sol at maximum effort, 48 at $1.06. Half the price per token, about a fifth cheaper per task — and still below Opus's default. List price is no longer the comparison; effort and tokens per task are.

So I ran both — three hard tasks from my Prompt-Vault, Sol in Codex, Opus in Claude Code, both at effort high:

A pyramid drawn in text characters that should look 3D on paper. I printed both; Opus's illusion was more convincing.
A simulation of creatures evolving under a genetic algorithm. Both excellent — a tie.
An audit of Texodus, my Markdown editor. Sol took five minutes and returned solid findings. Opus took about eight and found noticeably more — and when I checked them, they held up.

Opus two, one tie. A first impression, not a benchmark — the full run is in the video.

One honest caveat. Part of "40% cheaper" is a changed default: Opus 5.5 runs at medium effort unless told otherwise, where Opus 5 ran at high. And switching isn't just a model-ID swap — there are four breaking changes; the docs link below lists them.

2. My take

In this quick comparison, Opus came out the winner for me: it handled my tasks better. But Sol is good too. OpenAI's model did, in my view, a little worse — yet it didn't fail a single task. So over the next few weeks I'll keep using both, and only then draw final conclusions, which I'll share with you.

3. Tool of the week

Claude Opus 5.5 in Claude Code, as a codebase auditor.

What it's for: long, sprawling jobs across a whole repo — migrations and audits are what Anthropic sells it on, and the audit is where it beat Sol in my run.
Who needs it: anyone inheriting a codebase, or getting an open-source project ready for outside eyes.
How to try it: pick a repo you know well. Ask for an audit grouped by severity, with file and line for each issue, saved to a Markdown file. Then check five findings by hand before you trust the count — the count is the easy part to inflate.
Where it breaks: it's slower than Sol on the same job (about eight minutes against five, for me) and far wordier at high effort. And security-flavoured requests can go elsewhere: Anthropic says "most cybersecurity tasks will be re-routed to Opus 4.8", while routine bug finding and fixing stays on Opus 5.5. If your audit drifts into exploit territory, you may not be talking to the model you picked.

4. MVP idea of the week

A bring-your-own-repo model bake-off.

Problem: every launch ships a benchmark table measured against the competitor's previous model, at a different effort level. None of those tables uses your code, and running the same task by hand in two harnesses eats an afternoon.
Solution: a CLI that takes a task file and a repo, runs it through Claude Code and Codex headless in separate worktrees, and writes one report: wall-clock time, tokens, cost, diff size, tests passing — with both outputs side by side for a human to judge.
For whom: tech leads picking a team default, indie developers deciding what to pay for, and anyone who compares models in public.
Minimal v1: two harnesses, one task, a Markdown report. No scoring model — you pick the winner, the tool records the numbers.
Demand check: publish your own bake-off on a public repo with the report attached, and count how many people ask to run it on theirs.
Why now: two new everyday models landed on the same afternoon, each measured against the other's predecessor — and cost per task now depends more on the effort dial than on the price list.

5. Useful links

Anthropic — Introducing Claude Opus 5.5 — prices, the benchmark table, and a candid line that benchmark margins no longer predict much.
Claude docs — What's new in Opus 5.5 — four breaking changes and a lower default effort; read before changing the model ID.
OpenAI — Introducing GPT-6 Sol and Luna — half-price Sol, and caching controls that may matter more than the price.
UpGuard — Systemic data exposure in Supabase apps — 16,326 readable databases, many agent-built; check yours before October 30.
TechCrunch — OpenAI's agent swarms and online databases — part two of the rogue-agent story, told by Transluce rather than OpenAI.

— Nick

P.S. Six of my own things, on GitHub: