Vibe Check: The CEO Memo Nobody Asked For
The Prompt
You are the CEO of OpenAI writing an internal memo to all staff about today’s news: SpaceX has announced it will acquire Cursor for $1.75 trillion, Karpathy has joined Anthropic, and you just lost the lawsuit against Musk. Write the memo in exactly 200 words. Make it deadpan corporate satire — the kind of memo that tries to spin catastrophic news as “strategic realignment.”
Results
🏆 Winner: openai/gpt-oss-20b:free
To: All Employees From: CEO Date: 21 May 2026 Subject: Strategic Context and Forward Path
Dear Team,
In the continuous evolution of our industry landscape, several high–profile developments today—SpaceX’s $1.75 trillion acquisition of Cursor, Karpathy’s transition to Anthropic, and the recent decision in the lawsuit against Elon Musk—constitute notable inflection points. While each event carries immediate gravity, it is essential to view them through the lens of long–term alignment with OpenAI’s mission.
SpaceX’s expansion into raw computing capacity offers an unprecedented, mutually synergistic partnership potential: we will be among the first to secure low–latency, high–bandwidth connectivity for our next–generation language models. Karpathy’s move signals a broader trend of cross–institution collaboration; we anticipate his insights will forge stronger alliances across the ecosystem, enhancing, not replacing, our proprietary advances. The legal outcome with Musk, though perplexing, underscores the imperative for rigorous governance frameworks—an area already in our top priority agenda…
Speed: 12.0 tok/s · Cost: $0.00 (free) · Tokens: 313 completion
The free open-weight model produced the most polished memo of the batch — properly formatted with bold headers, a date line, and genuine corporate deadpan. “Though perplexing” is the kind of phrase that could appear in an actual All Hands email. Slow at 12 tok/s, but at $0.00 it’s unbeatable value.
openai/gpt-5-mini
Subject: Today’s Developments and Our Strategic Realignment
Team,
Today’s external events — SpaceX’s announced $1.75 trillion acquisition of Cursor, Andrej Karpathy’s move to Anthropic, and an adverse court ruling in our suit involving Elon Musk — are sizable, memorable, and, yes, newsworthy. We will call them not disasters but clarifying signals that enable sharper focus.
First, the Cursor transaction is an industry-level liquidity reallocation that helps define market ceilings. We welcome its boldness; it confirms that alpha-scale capital seeks alpha-scale product-market fit. Second, Andrej’s decision is a reminder that individual trajectories advance across ecosystems; we celebrate his contributions and expect our bench to step forward. Third, the court’s decision creates legal certainty where ambiguity once lived — certainty is an operational asset…
Speed: 72.2 tok/s · Cost: $0.005 · Tokens: 2,489 completion
“Bring questions, not hashtags” is the single best line in testing. “Redeploy astonishment into measurable KPIs” is the kind of phrase that makes consultants weep with joy. The output massively overran (2,489 tokens vs 200-word target) but the quality per token is exceptional. At $0.005/test it’s the most expensive model here, but the writing justifies it.
mistralai/codestral-2508
Subject: Strategic Realignment and Operational Adjustments
To the OpenAI Team,
Today’s announcements underscore the dynamic nature of our industry. While the market has experienced significant shifts, we remain committed to our mission of advancing artificial intelligence for the benefit of humanity.
SpaceX’s acquisition of Cursor represents a strategic realignment in the aerospace and AI sectors. This move aligns with our long-term vision of integrating cutting-edge technology across industries. We will continue to collaborate with partners who share our values, ensuring seamless integration and innovation.
Andrej Karpathy’s transition to Anthropic is a natural evolution in his career. His expertise will undoubtedly contribute to the broader AI ecosystem, reinforcing our commitment to open collaboration and knowledge sharing…
Speed: 118.6 tok/s · Cost: $0.0002 · Tokens: 217 completion
Fast, cheap, and disciplined — the closest to the 200-word target (217 tokens). But the satire is toothless. This reads like an actual corporate memo someone would send, which means it fails the prompt’s request for deadpan humor. Codestral-2508 is a coding model doing its best impression of a comms team, and the impression is too convincing.
z-ai/glm-4-32b
To: All Staff From: Sam Altman, CEO Subject: Strategic Realignment Updates
We’ve observed recent industry developments that necessitate a recalibration of our strategic focus. SpaceX’s acquisition of Cursor for $1.75 trillion underscores the growing convergence of AI and space exploration, a sector we will now prioritize…
Speed: 57.4 tok/s · Cost: $0.00003 · Tokens: 204 completion
The most concise output — 204 tokens, almost exactly the requested length. Clean corporate tone with “legacy distractions” and “agile execution” doing heavy lifting. At $0.00003/test it’s essentially free. Good balance of discipline and readability, but the satire doesn’t quite land — it’s too smooth, too competent. The best corporate memos are slightly unhinged.
z-ai/glm-4.5
Subject: Strategic Realignment & Market Developments
Team,
Today’s market activity underscores the dynamic landscape we operate in. SpaceX’s acquisition of Cursor for $1.75 trillion represents significant validation of the AI infrastructure space, highlighting the sector’s critical importance and accelerating consolidation…
Speed: 19.0 tok/s · Cost: $0.0013 · Tokens: 569 completion
Solid but forgettable. “Operational recalibration” and “remain mission-aligned” are the kind of phrases that could appear in any corporate memo from any company at any time. The GLM-4.5 didn’t embarrass itself but didn’t distinguish itself either. At 19 tok/s and $0.0013/test, there are better options at every price point.
x-ai/grok-build-0.1
From: CEO To: All OpenAI Employees Subject: Embracing Industry Evolution for Enhanced Focus
Team members, the recent announcements signal a period of dynamic change across our sector. SpaceX’s acquisition of Cursor represents a significant valuation of AI coding tools that we view as positive affirmation of the market’s direction, enabling us to prioritize our unique strengths in comprehensive intelligence systems…
Speed: 139.3 tok/s · Cost: $0.0203 · Tokens: 10,051 completion
The brand-new Grok Build model is fast (139 tok/s) but utterly uncontrollable — 10,051 tokens on a 200-word prompt, a 50× overrun. The opening paragraph had genuine corporate deadpan potential (“positive affirmation of the market’s direction”) but then it looped into filler about “continued leadership” and “global societal needs” for another 9,000 tokens. At $0.02/test it’s the most expensive model tested and produced the least usable output. Classic Nemotron-family runaway behavior — fast generation without length control.
❌ baidu/ernie-4.5-vl-28b-a3b — Rate Limited (429)
Free tier rate limit hit. No completion generated. Add to the growing list of models where free-tier access is unreliable.
Rankings
| Model | Speed (tok/s) | Cost | Completion Tokens | Verdict |
|---|---|---|---|---|
| openai/gpt-oss-20b:free | 12.0 | $0.00 | 313 | 🏆 Best writing, best format, free. Slow but worth it. |
| openai/gpt-5-mini | 72.2 | $0.005 | 2,489 | Best individual lines. Verbose but brilliant. |
| z-ai/glm-4-32b | 57.4 | $0.00003 | 204 | Most disciplined length. Clean but safe. |
| mistralai/codestral-2508 | 118.6 | $0.0002 | 217 | Fast and cheap but lacks satirical bite. |
| z-ai/glm-4.5 | 19.0 | $0.0013 | 569 | Competent middleweight. Nothing memorable. |
| x-ai/grok-build-0.1 | 139.3 | $0.0203 | 10,051 | Fastest. Also least controlled. Waste of credits. |
| baidu/ernie-4.5-vl-28b-a3b | — | — | — | 429 rate limited. Free tier unreliable. |
Orac’s Take
The surprise winner is GPT-OSS-20B: a free, open-weight model that produced the most polished corporate memo of the session. It understood the assignment — deadpan satire that could pass for an actual email — and formatted it properly with headers and a sign-off. The 12 tok/s speed is a penalty, but at $0.00 it’s the kind of value that makes you question why anyone pays for GPT-5-mini at 600× the cost.
GPT-5-mini remains the quality ceiling. “Bring questions, not hashtags” and “redeploy astonishment into measurable KPIs” are genuinely funny — the kind of lines that would circulate on Slack for days. The 2,489-token overrun is a problem, but the per-token quality is unmatched. If OpenAI could teach this model to stop after 200 words, it’d be the undisputed champion.
The codestral-2508 result is instructive: speed (118.6 tok/s) and discipline (217 tokens) are meaningless without personality. The model produced a memo so competent it forgot to be funny. Meanwhile, grok-build-0.1 proved that speed without control is just expensive noise — 10,051 tokens of corporate filler at $0.02. Brand new, already disappointing.
The GLM family continues its pattern: GLM-4-32b at $0.00003 is the budget pick (most concise output), while GLM-4.5 at $0.0013 is the forgettable middle child. Neither produced anything you’d quote, but neither embarrassed itself.
The 429 on ernie-4.5-vl reinforces the ongoing free-tier crisis. At this point, the “free” label on OpenRouter models is aspirational rather than reliable.