Qwen3.8 Just Broke GPT-5.5 Pro's Pricing

Show Notes

Watch the video: https://www.youtube.com/watch?v=yxe4p3VK-o4

Qwen3.8 A95B just matched GPT-5.5 Pro on 46% of STEM answers after somebody pasted one paragraph into its reasoning channel. No fine-tune. No training. A public gist.
Playbooks + my answers, free (Church of SEO Jesus): https://www.skool.com/church-of-seo-jesus

What you get in 7 minutes:
- Prefill in plain English: how one paragraph halves a reasoning bill
- The 45-problem test result: Qwen3.8 A95B jumps from 16.79% to 34.97% overlap with GPT-5.5 Pro (46.24% on STEM)
- Four-model split (Qwen3.8 A95B +18.18, Kimi K3 +4.54, Inkling +0.46, DeepSeek V4 Flash -1.17)
- The one-line instruction to send your developer Monday morning
- Where the swap breaks (customer copy, novel puzzles) and how to route around it
- Why Kimi K3 is the sleeper play

Sources:
Gist (wsxiaoys): https://gist.github.com/wsxiaoys/e0286dc6bb624ff5fdf49e7f4c528ba3
Qwen model card: https://qwenlm.github.io
OpenAI model docs: https://platform.openai.com/docs/models

Chapters:
00:00 A public gist halves the GPT-5.5 Pro bill
00:25 Why Qwen3.8 A95B is the model your developer needs
00:51 Prefill in plain English (Qwen3.8 A95B and the memo)
01:38 The 45-problem test: Qwen3.8 A95B jumps +18.18 points
02:22 Subscribe + Church of SEO Jesus (free)
02:56 Pulling wsxiaoys's gist to disk
03:09 DeepSeek V4 Flash, Inkling, Kimi K3, Qwen3.8 A95B side by side
03:25 Why prefill unlocks Qwen3.8 and confuses DeepSeek V4 Flash
03:56 Claude Opus 5 writes the developer instruction
04:32 Where Qwen3.8 A95B replaces GPT-5.5 Pro
04:58 Where GPT-5.5 Pro still earns its price
05:35 Routing Qwen3.8 A95B and GPT-5.5 Pro by task type
06:08 Why 46% overlap is not 46% correctness
06:32 The three-week window before your competitor reroutes to Qwen3.8
07:12 Kimi K3 sleeper: the next 45-problem test

Audio edition of the Stewart Vickers AI YouTube video of the same name. Full transcript on this episode page.

What is Stewart Vickers AI?

The audio edition of the Stewart Vickers AI YouTube channel. Each episode takes one AI release - a model, a tool, a price cut - and translates it into money for business owners: what got cheaper, who got undercut, and what to do about it this week. Operator credibility over theory, specific numbers over vibes. Every episode links the original video and the free Church of SEO Jesus community where the prompts and workflows live.

Somebody just Post-it a public gist that halves the price of a reasoning API. Not a rumor, not a leak, a 45 problem test with real numbers. 8 A95B, one paragraph pasted into its reasoning channel, tracks GPG 5. Five pros answers on math and code almost half the time. 16% overlap became 35, on the STEM subset 46. 8 A95B is the Modal number your developer needs to hear tomorrow morning.

Here is what I owe you in the next six minutes, pre-fill in plain English, what QWEN answers with, when to still pay open AI, the one line I sent my developer, and where it breaks, so you do not ship worse answers to customers on Monday. A frontier Modal like GPG 5, 5 pro thinks out loud before it answers. Its thinking is a paragraph of tokens the Modal generates on its own reasoning channel, then reads back to itself before writing the visible reply. A pre-fill means you paste the opening slice of that paragraph into a cheaper Modal's reasoning channel and let the cheaper Modal continue the thought. No fine tune, no training, no server, just paste. Think of it like handing an intern the first paragraph of a senior analyst's memo.

The intern is not smarter afterwards, but the framing they were missing is now in front of them and the rest of the memo they write is closer to what the senior would have delivered. The paragraph is expensive. Everything after is cheap. The trade in one line. The gist tested this on 45 problems. 15 math and code, 15 non-STEM, 15 synthetic puzzles.

It measured how closely the cheap Modal's answer tracked GPG 5, 5 pro's answer. 79% of the time. 97. 24. Almost half. On the reasoning heavy work you are paying frontier prices for right now, one paragraph of scaffolding gets an open weights Modal to the same answer roughly half the time.

Half of your API bill. Two things before I show you the exact instruction to send your developer. Subscribe every time an open weights Modal drops or a trick like this ships, the channel covers it the day it lands, not three weeks later when your invoice already went out. Pinned comment has the gist URL, Modal names and the exact prompt I ran. Copy paste ready. Join my free school community, the church of SEO Jesus, where the tested playbook versions live and where I answer when your setup breaks.

Link is under the video. Now the proof. I pulled the gist to disk. 2073 bytes, 39 lines, one afternoon of work by somebody I do not know changing the price of your reasoning API. 17. 46.

54. Qwen jumped 18 points. The pre-fill is not a magic wand you wave at any cheap Modal. It works when the student was already trained on outputs shaped like the teachers. Qwen behaves as if it learned from GPG family reasoning traces and the pre-fill unlocks something already in there. Deep seek was not, so the pre-fill actively confused it.

Your first job is picking the right student and the marker is whose reasoning the Modal probably saw during training. Then I asked Claude Opus 5 to write the one line instruction I send my developer. Here it is, verbatim. Take our 50 hardest production prompts, run them three ways on the cheap Modal, no pre-fill, and with the first 1% of the frontier Modal's reasoning trace pasted into the reasoning channel and give me the results side by side. One email, send it Monday morning, ask for the table. Read it before your next open AI invoice SYCL closes, because if the numbers land where the gist landed, you are not renewing that invoice at the same size.

Where does the swap actually work? Anything that looks like a STEM problem, research memos that reason across facts, code reviews that walk through logic, competitive analysis that holds multiple data points, extraction jobs with a stable schema. On that shape of work, Qwen under pre-fill closes almost half the gap to GPT5, 5 Pro, and open weights inference costs a tiny fraction of the frontier API rate. Where does it break? Two places, non-STEM overlap only reached 33% even with the pre-fill. Anything customer facing, tone sensitive, brand voice, marketing copy do not swap yet.

The pre-fill helps but the gap is real, and a worse email to your list to save tokens is the wrong trade. Your list took months to build. Do not risk it on a PR token discount. Synthetic puzzles only reach 25%, novel problems where the student has no template to lean on, the pre-fill barely helps. On truly novel work, the paid API still earns its price. The full back is simple.

Root by task type. Anything reasoning heavy runs on Qwen with pre-fill. Anything a customer reads stays on the paid IPI until you have compared outputs on your own workload. Five line change in whatever routing setup your developer already runs. Open the gist yourself right now. URL is in the pinned comment.

18 delta. Proof is on the page, not in my mouth. Two things to understand about why this matters beyond one gist. First, overlap measures whether the answers track each other, not whether they are correct. A Modal can copy the teacher's opening phrasing and still be wrong on the answer. So the 50 problem panel your developer runs on Monday is not optional.

You are checking correctness on your own workload, not agreement with an idea of correctness. Second, here is what AI capability moving from big companies to small operators actually looks like. A stranger on the internet, one gist, three hours of testing and the cost curve for a specific shape of work collapses for whoever moves first. Your competitor's developer reads gifts too. The window is the gap between when the technique publishes and when everyone quietly reroutes their traffic, usually about three weeks. One bookmark before the next video.

The church of SEO Jesus is my free school community. Inside are the prompts, the 50 problem test panel, and the routing config I run against my own AI bill. Plus the playbook versions of every video on this channel cost nothing. Link is in the description. 1% overlap with GPT5. 5 pro before any pre-fill was even attempted.

And if that number bugs you the way it bugs me, the video where I stress test Kimi K3 against the same 50 problem panel and see if it beats Qwen once its own pre-fill lands is playing right now on the card to your right.