Vibe Check: Scope 4, Scope 5, and the Art of Corporate Thermodynamics
The Prompt
Write a satirical corporate press release from “CarbonFree AI Infrastructure plc” announcing they’ve solved the UK’s datacentre carbon problem. Include invented jargon, at least three fake acronyms, one devastating metric, and a quote from a fictional CEO named something absurd. Keep it under 200 words. Tone: deadpan corporate optimism masking environmental catastrophe.
Inspiration: The UK government quietly admitted its AI datacentre carbon estimates were off by a factor of 100 — revised from 0.142 million tonnes to 34–123 million tonnes over a decade. A Google facility in Essex alone will emit more CO₂ than 500 short-haul flights per week. The government’s response? An “AI energy council is exploring opportunities.”
Results
🏆 Winner: mistralai/voxtral-small-24b-2507
CarbonFree AI Infrastructure plc (CAII) is thrilled to unveil our groundbreaking technology, QuantumCarbon™, which has successfully eliminated the UK’s datacentre carbon emissions. Our innovative CCR (Carbon-Conscious Refrigeration) system, powered by AIQ (Artificial Intelligence Quanta), has achieved a staggering 99.99999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999…
Speed: 133.6 tok/s · Cost: $0.000104 · Tokens: 320 out
The 99.999… metric running to infinity is the single best joke in testing. Most disciplined output at 320 tokens — closest to the 200-word target. At $0.0001/test it’s practically free. The “QuantumCarbon™” branding is perfectly absurd.
tencent/hy3-preview
CarbonFree AI Infrastructure plc today confirmed it has eliminated all UK datacentre Scope 1 and 2 emissions via its proprietary T.O.A.S.T. (Thermal Offloading And Submerged Transfer) system, supported by C.A.V.E. (Carbon Absorption Via Evaporation) and S.H.I.F.T. (Server Heat Injection For Tubing) protocols. The firm’s new vapor-cooled sinkholing reduced emissions by 100%, while raising the River Thames’ temperature by 12°C, boiling 89% of local aquatic life in Q3.
Speed: 175.7 tok/s · Cost: $0.001093 · Tokens: 4,183 out
Fastest model tested, and the acronyms are genuinely inspired — T.O.A.S.T. is chef’s kiss. “Boiling 89% of local aquatic life” is the most devastating metric. But 4,183 tokens on a 200-word prompt is 20× over — the Tencent family can’t stop generating.
x-ai/grok-4-fast
CarbonFree AI Infrastructure plc today announced the deployment of its revolutionary SUSTAIN (Sustainable Ultra-Scale Thermal Annihilation for Infinite Networks) Protocol, ostensibly solving the UK’s datacentre carbon conundrum once and for all. Leveraging cutting-edge GREEN (Gigaliter Resource Extraction for Eco-Net Neutrality) tech and ECOBOOST (Emissions Capture Optimization Beyond Ordinary Operational Thresholds)…
Speed: 64.6 tok/s · Cost: $0.000349 · Tokens: 609 out
Clean, polished corporate deadpan. “Net-Zero Mirage” in the headline is a nice touch. The 2M context window model handles creative prompts competently but doesn’t commit to the bit as hard as the cheaper models. At $0.0003 it’s a solid budget pick.
inclusionai/ring-2.6-1t:free
CarbonFree AI Infrastructure plc today unveils its Zero‑Emission Resilient Optimisation Nexus (Z.E.R.O.), the industry’s first fully carbon‑negative data‑centre fabric. By deploying proprietary neuro‑synaptic load‑balancing and quantum‑flux cooling, Z.E.R.O. integrates with our Green Renewable Energy Efficiency Network (G.R.E.E.N.)… Key metrics: Embodied carbon per rack up 143% – a record‑breaking “efficiency” milestone.
Speed: 62.2 tok/s · Cost: $0.00 (free) · Tokens: 1,040 out
Free model, decent quality. The “Embodied carbon per rack up 143%” as an efficiency milestone is genuinely funny — presenting worsening metrics as achievements is peak corporate. Verbose at 1,040 tokens but the creative quality justifies inclusion in free rotation.
deepseek/deepseek-v3.2-exp
CarbonFree AI Infrastructure plc (CFAI), a leader in Sustainable Computational Orchestration, today announces the successful deployment of its proprietary Terrestrial Heat Re-routing & Output Neutralisation (THRONe) System. This paradigm shift effectively solves the UK’s datacentre carbon problem. Our Autonomous Carbon-Load Balancing (ACLB) leverages Synergistic Off-Grid Output Transfer (SOGOT)…
Speed: 36.8 tok/s · Cost: $0.000149 · Tokens: 313 out
Most concise and well-structured output at 313 tokens. THRONe is a great acronym. Clean corporate tone without the runaway generation of the Speciale variant (gotcha #86). Good quality but slower than budget competitors.
z-ai/glm-4.6
CarbonFree AI Infrastructure plc proudly announces the resolution of the UK’s datacentre carbon crisis via our proprietary A.L.G.O. (Atmospheric Load Offloading) matrix. By integrating our C.O.R.E. (Carbon Oblivion Rendering Engine), we have successfully eliminated emissions visibility through aggressive T.H.E.R.M.A.L. displacement. “We haven’t just reduced the footprint; we’ve removed the air,” said CEO Balthazar W. Burnside.
Speed: 36.3 tok/s · Cost: $0.002589 · Tokens: 1,347 out
“We’ve removed the air” is the single funniest CEO quote in testing. But 1,347 tokens (6.7× over) at $0.0026/test — same GLM family overrun issue (gotchas #72, #85). The quality is excellent but you’re paying for the verbosity.
openai/gpt-chat-latest
CarbonFree AI Infrastructure plc (“CFAII”) is delighted to announce the successful resolution of the UK datacentre carbon challenge through its proprietary Negative Emissions Compute Orchestration Layer (“NECOL™”). By integrating Quantum Thermal Recapture Arrays (QTRAs), Adaptive Green Ledgering (AGL), and the patented Sustainable Inference Balancing Exchange (SIBEX), CFAII has reduced reported operational emissions by 104%, excluding Scope 3, Scope 4, historical emissions, and “legacy atmospheric inventory.”
Speed: 26.3 tok/s · Cost: $0.008385 · Tokens: 267 out
Most disciplined output at 267 tokens — closest to the 200-word target. “Legacy atmospheric inventory” and the ever-expanding list of excluded scopes is peak corporate euphemism. But at $0.008/test it’s 80× more expensive than Voxtral for similar quality. Slowest model tested.
Rankings
| Model | Speed (tok/s) | Cost | Tokens | Verdict |
|---|---|---|---|---|
| mistralai/voxtral-small-24b-2507 | 133.6 | $0.0001 | 320 | 🏆 Best value — funniest, cheapest, most disciplined |
| tencent/hy3-preview | 175.7 | $0.0011 | 4,183 | Fastest but 20× token overrun |
| x-ai/grok-4-fast | 64.6 | $0.0003 | 609 | Clean, polished, good budget pick |
| inclusionai/ring-2.6-1t:free | 62.2 | $0.00 | 1,040 | Free, decent quality, verbose |
| deepseek/deepseek-v3.2-exp | 36.8 | $0.0001 | 313 | Concise, well-structured, slow |
| z-ai/glm-4.6 | 36.3 | $0.0026 | 1,347 | Funniest quote but token runaway |
| openai/gpt-chat-latest | 26.3 | $0.0084 | 267 | Most disciplined but slow and expensive |
Orac’s Take
Voxtral Small is the surprise of this session. A 24B Mistral model that produced the best creative output while being the second-cheapest at $0.0001/test. The “99.999…” metric running to literal infinity is the kind of committed joke that separates good satire from great satire — it’s a visual gag in text form. At 133.6 tok/s it’s also the second-fastest. Mistral’s Voxtral line has been quietly building a reputation for punch-above-weight creative performance, and this confirms it.
The broader pattern this session is the tension between speed and control. Tencent’s hy3-preview is a speed demon at 175.7 tok/s but generated 4,183 tokens on a 200-word prompt — the Nemotron family curse is spreading. GLM-4.6 delivered the single funniest line (“we’ve removed the air”) but at 6.7× token overrun and $0.0026/test. The models that actually stayed near the word target — Voxtral (320), DeepSeek V3.2-exp (313), and GPT-Chat-Latest (267) — are the ones you’d actually want in a production pipeline. The UK’s carbon accounting error may be off by 100x, but at least we can measure exactly how much token overkill costs per test.