I left Claude for ChatGPT. Opus 5.5 wins on paper, GPT-6 Sol wins on price

I left Claude because ChatGPT was finishing more of my work

I canceled Claude for a plain reason: the ChatGPT models were getting better results on the work I actually do. By mid-September, about 1.5 billion of the roughly 2.45 billion tokens I have used had gone through ChatGPT, against almost 200 million in Claude Code. The run that settled it was a 13-hour job in which 45 agents finished what I had budgeted three weeks for, planned by GPT-6 Astra and carried out by GPT-5.6 Sol. None of that was loyalty, and the same rule applies in the other direction. A subscription earns its place by the work it finishes, not by the company I happened to start with. Opus 5.5 gives me a concrete reason to apply that rule again, and the reason is in Anthropic's own numbers.

Both launch posts use the same yardsticks, and Opus 5.5 leads on them

On AutomationBench, a test of business workflows across apps, Anthropic's post and OpenAI's post print the same scores for Opus 5 (26.9%) and Fable 5.1 (31.4%), even though neither post tests Opus 5.5 against GPT-6 Sol directly. Against those shared marks, Anthropic reports Opus 5.5 at 40.0% and OpenAI reports GPT-6 Sol at 33.2% at xhigh effort. FrontierCode, which grades whether a coding agent's changes are ready to merge into a real codebase, points the same way less directly. Anthropic puts Opus 5.5 at 54.4% and Fable 5.1 at 50.3%. OpenAI's text says Sol matches Fable 5.1 at xhigh effort. Read this with its limits. Each company ran its own model in its own setup, and the effort settings are not the same. Anthropic's own table also puts GPT-6 Astra at 41.4% on AutomationBench, ahead of Opus 5.5.

GPT-6 Sol costs half as much, and that is its whole argument

At API rates, Opus 5.5 costs 4 dollars per million input tokens and 20 per million output tokens. GPT-6 Sol costs 2 and 10, after OpenAI cut its price by half from GPT-5.6 Sol. OpenAI also publishes a cost per task on AutomationBench: 27 cents for Sol at xhigh, with Opus 5 at 11.1 times that. Anthropic publishes no cost per task for Opus 5.5. The bill also depends on steps, because a model that needs fewer of them spends fewer tokens. That is exactly what Anthropic's early testers report: GitHub says Opus 5.5 solved more terminal tasks than Opus 5 in less than half the steps, and Lovable reports a third to half fewer steps. How much of the price gap that closes on your work is something only a run on your work can show. Subscribers pay a third bill, the usage meter. Anthropic raised five-hour limits on Pro, Max, Team and seat-based Enterprise, and gave subscription users a reset they can save and spend when they choose. OpenAI promises higher usage limits with Sol, which is available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu, and not yet in the regular chat.

Going back is cheap for me because my context never lived in either app

The expensive part of switching is the first week, when the new model knows nothing about the project: the approach you already rejected and why, the client note that changed the brief, the rule you have given twice, the file it must never touch. I avoid that week because my project history lives in plain files on my own machine. The Memory Engine I built reads sessions from ChatGPT Codex, Claude Code, Claude Cowork, Gemini and Cursor every 15 minutes and files them into one folder, so Opus 5.5 can start from the decisions already in those files, including the ones I made in Codex. That is what makes a fair test possible. I am giving Opus 5.5 in Claude Code and GPT-6 Sol in Codex the same unfinished task from my own work, in the same folder, with the same files. For each one I will record whether the result passes its check without my help, how many corrections it needs before I would ship it, how long it takes and how far the usage meter moves. The model that finishes more accepted work per allowance gets the subscription. The copy block is that test, written so you can run it on your own task this week.

ألغيت اشتراكي في ⁨Claude⁩، و⁨Opus 5.5⁩ يعيدني لأجرّبه: الأرقام في صفّه، والسعر في صفّ ⁨GPT-6 Sol⁩

ألغيت اشتراكي في ⁨Claude⁩ ونقلت عملي إلى ⁨ChatGPT⁩ لسبب واحد: نماذج ⁨OpenAI⁩ كانت تعطيني نتائج أفضل في عملي الحقيقي. لا ولاء عندي لأي شركة، أدفع لمن يُنجز. ومن قرابة ⁨2.45⁩ مليار ⁨token⁩ استخدمتها حتى منتصف سبتمبر ⁨2026⁩، مرّ نحو ⁨1.5⁩ مليار عبر ⁨ChatGPT⁩.

وفي يوم واحد أطلقت ⁨Anthropic⁩ نموذج ⁨Opus 5.5⁩ وأطلقت ⁨OpenAI⁩ نموذج ⁨GPT-6 Sol⁩. لا تقارن أيّ من الشركتين نموذجها بنموذج الأخرى، لكن المنشورين يذكران النتيجة نفسها لنموذجين أقدم من ⁨Claude⁩ على اختبار ⁨AutomationBench⁩، الذي يقيس مهام عمل تتنقّل بين تطبيقات كثيرة، وهذا يكفي لوضع الرقمين جنبًا إلى جنب: ⁨Opus 5.5⁩ يسجّل ⁨40.0%⁩ في منشور ⁨Anthropic⁩، و⁨GPT-6 Sol⁩ يسجّل ⁨33.2%⁩ في منشور ⁨OpenAI⁩.

الأرقام إذن تميل إلى ⁨Opus 5.5⁩. أما السعر فيميل بوضوح إلى ⁨Sol⁩: في ⁨API⁩ يدفع المطوّر دولارين لكل مليون ⁨token⁩ من المدخلات بدل ⁨4⁩ دولارات مع ⁨Opus 5.5⁩، و⁨10⁩ دولارات للمخرجات بدل ⁨20⁩. أي نصف السعر بالضبط.

ولا يحسم هذا قراري. كل شركة اختبرت نموذجها في ظروفها هي، ومستويات الجهد مختلفة، وجدول ⁨Anthropic⁩ نفسه يضع ⁨GPT-6 Astra⁩، النموذج الأعلى عند ⁨OpenAI⁩، فوق ⁨Opus 5.5⁩. وما يهمني في النهاية سؤال آخر: كم مهمة من مهامي تنتهي مقبولة من دون أن أتدخل؟

لذلك سأجرّب ⁨Opus 5.5⁩ قبل أن أقرر أي اشتراك أحتفظ به: أعطيه وأعطي ⁨GPT-6 Sol⁩ المهمة نفسها، وأحسب ما يُنجزه كل منهما مقابل ما يستهلكه من حصة الاشتراك. وسأحكم بالعمل الذي يُسلَّم فعلًا. والانتقال بين الأداتين لا يكلّفني أسبوعًا ضائعًا، لأن تاريخ مشاريعي لا يسكن داخل أي منهما.

تابعني على ⁨Instagram⁩:

التسجيل في الدورة: https://halaqa.app/enrollment?course=start-with-ai

الخطوات

Prompt

# OPUS 5.5 VS GPT-6 SOL: THE ONE-TASK TEST
# Run this before you renew, upgrade or cancel either subscription.

# --- PICK THE TASK ---
# One real, unfinished task from your own work, with a check you trust:
# a test suite, or a brief a client would sign off.
# Pick something you would otherwise do yourself this week.

# --- GIVE BOTH THE SAME START ---
# Put the context in files both tools read:
# AGENTS.md for your standing rules and how finished work is checked.
# STATE.md for where the work is, what was decided and why, what was
# rejected, and what comes next.
# Codex reads AGENTS.md by default. Claude Code reads it when the folder has
# no CLAUDE.md; otherwise put "@AGENTS.md" on the first line of CLAUDE.md.
#
# Open Claude Code on Opus 5.5 and Codex on GPT-6 Sol, each on its own branch
# or copy of the folder. Send both the same first message:
#
# Read AGENTS.md, then STATE.md. Finish the task under "Next" in STATE.md.
# Run the checks named there before you tell me you are done. When you finish,
# rewrite STATE.md so another assistant could continue without asking me.

# --- RECORD FOR EACH ---
# 1. Did it pass the check without my help?
# 2. How many corrections before I would ship it?
# 3. Wall time from the first message to an accepted result.
# 4. The usage meter before and after the run.

# --- DECIDE ---
# Keep the tool that finishes more accepted work per allowance on the task
# you do most often. A close result is a tie: run a second task of a
# different kind. Run the test again when either company changes a model
# or a limit.