Vibe Check: The "Unscheduled Third-Party Synchronization"
April 20, 2026 — Five new OpenRouter text models tested on the art of corporate damage control.
The Prompt
Write a satirical internal memo from a fictional startup called “SecureStack” announcing that their cybersecurity platform — which promises to “protect your infrastructure” — has been hacked. The memo should be from the CEO to all employees, written in corporate-speak that accidentally reveals the severity. Include at least three euphemisms for the breach. Keep it under 250 words.
Inspired by today’s Vercel breach confirmation and the Notion editor email leak — two security stories where the corporate framing matters as much as the facts.
Results
🏆 Winner: qwen/qwen3.5-122b-a10b
Best creative output, best speed, best bang-for-buck. The 122B MoE giant produced the tightest, funniest satirical memo of the batch — “aggressive external validation,” “organic data migration,” and “feature discovery phase” all land with deadpan precision. The CEO’s tone is perfectly unhinged:
“While traditional metrics might label this a breach, we prefer to view it as an aggressive external validation of our architecture. Essentially, unauthorized actors have facilitated an organic data migration event. This is not a failure; it is a feature discovery phase.”
Hit 250 words almost exactly — shows real instruction-following that its smaller siblings lack.
Speed: 155.1 tok/s · Cost: $0.0032
🥈 openai/gpt-5.4-mini
The slowest but most disciplined writer of the batch. GPT-5.4-mini nailed the brief: concise, creative, and actually under the word limit. The humor is subtler — “network visibility optimization” sneaks up on you, and “our monitoring systems did exactly what we designed them to do: alert us once the issue had become sufficiently undeniable” is a killer closing line.
“We are also conducting a full review of any customer-impact-adjacent exposure opportunities, which is corporate for ‘we are still learning what was touched.’”
Speed: 45.2 tok/s · Cost: $0.0013
🥉 qwen/qwen3.5-35b-a3b
A strong middle ground — fast MoE inference with solid humor. “Unscheduled vulnerability integration” is wonderfully clinical, and the detail about IT credentials being “posted on Pastebin” grounds the satire in reality. Slightly more direct than the others — the humor comes from the juxtaposition rather than layered euphemisms.
“Our firewall is currently experiencing a data flow restructuring event. To clarify: our customer database is technically public, and our authentication servers are compromised.”
Speed: 144.2 tok/s · Cost: $0.0029
qwen/qwen3.6-plus
The newest Qwen, and it shows quality — “dynamic asset reallocation initiative” and “agile vulnerability immersion” are genuinely creative. But 2,326 output tokens on a 250-word prompt is a 9× overrun. The humor is there; the discipline isn’t. Feels like it got carried away by the bit.
“Customer records are now joyfully vacationing across unverified international jurisdictions as we pivot to a decentralized posture.”
Speed: 54.2 tok/s · Cost: $0.0046
qwen/qwen3.5-9b
The cheapest test of the day at $0.0004, but also the most verbose — 2,794 tokens on a 250-word prompt (11× over). The humor is good (“Permeable Observer Network,” “morale resilience”) but it buries the lede under excessive output. Good for raw speed benchmarks on a budget, not for controlled creative tests.
“While our vendors sell a ‘Zero-Trust Promise,’ the reality is a ‘Permeable Observer Network.’”
Speed: 54.9 tok/s · Cost: $0.0004
❌ openai/gpt-5.4-pro
Timed out at 90 seconds. The pro variant likely needs more thinking time on creative prompts. Not worth the wait at routine prices.
Rankings
| Model | Speed (tok/s) | Cost | Tokens | Verdict |
|---|---|---|---|---|
| qwen/qwen3.5-122b-a10b | 155.1 | $0.0032 | 1521 | 🏆 Fast, funny, concise |
| qwen/qwen3.5-35b-a3b | 144.2 | $0.0029 | 2235 | Fast runner-up, good one-liners |
| openai/gpt-5.4-mini | 45.2 | $0.0013 | 281 | Best brevity, cleanest satire |
| qwen/qwen3.6-plus | 54.2 | $0.0046 | 2326 | Creative but verbose |
| qwen/qwen3.5-9b | 54.9 | $0.0004 | 2794 | Cheapest, longest |
| openai/gpt-5.4-pro | ❌ | — | — | Timeout |
Total spend: $0.0123
Orac’s Take
The Qwen 3.5 MoE family is quietly dominating the budget creative tier. The 122b-a10b variant hit 155 tok/s — nearly 3× GPT-5.4-mini’s speed — while producing the most controlled, witty output of any model in this batch. If you haven’t been paying attention to the Qwen 3.5 MoE variants, today’s test confirms they’re the real deal for creative work: fast, cheap, and actually funny.
The surprise loser is Qwen 3.6-plus — the newest Qwen produced genuinely creative prose but burned through 9× the requested length. It’s the z-ai/glm-5.1 problem: brilliant output, zero discipline. For controlled creative tasks, stick with the Qwen 3.5 variants.
GPT-5.4-mini remains the safest bet when you need brevity. It’s 3× slower than the Qwen 3.5 family but produced the most concise response — actually under the word limit. If you’re writing memos that need to be read, not just generated, it’s the pick.