GreenPT

Happy to announce we've partnered with GreenPT 🌱 But let me tell you WHY it matters to me. Back in November I asked myself a question: if I could use my talents to build a company, what would I build? For almost two years I'd watched the #AI boom explode out of nowhere. I didn't just watch, I played in it too, as a designer using these tools every day. And I saw the pattern. When AI can cut a company's costs, of course they'll implement it. Fast. That's just how it works. I believe technology can be good for people. AI too. But there's always a hidden cost. Right now the AI infrastructure boom is being built backwards. Gas plants go up first. Efficiency comes as an afterthought. And the communities breathing the emissions never got a say. All of us use AI for everything. Almost nobody is measuring what it costs the planet. So we built SOMA AI with our agents I call "Los Somitos." They measure your sustainability and turn it into reports and action plans you can actually use in the company. But measuring everyone else's footprint while ignoring our own? That felt wrong. That's why GreenPT. AI that runs on renewable energy and tracks its own CO₂. Today, it's fully integrated into SOMA. Guillermo Llopis made it happen 🙏 What it means: every time you use SOMA AI, you measure your impact, build your reports, and you do it on AI that isn't working against the whole point. I'm on a mission to make AI work for the planet. That means partnering with the right companies, not just the fast ones. The choices we make in the next year will outlast every model we're excited about now. The people paying the cost deserve to know it was worth it. Book a demo with us https://somaai.earth/ #ClimateTech #ResponsibleAI #SustainableAI #GreenPT #SOMAAI #ESG

Let's lead with efficiency. Our friends at GreenPT and Neuralwatt are calling on a sense-check for how we go about this AI transition. Instead of building power first and thinking about efficiency second, cut the waste and make AI leaner, then build for the remaining demand. - It's greener and better for society. - It's cheaper for the companies building AI. - It's cleaner for the towns hosting it. - And it leaves more power on the grid for everyone else. If you believe AI should be built responsibly, share their call to action. The decisions being made in the next eighteen months will outlive every model we’re excited about today, and the communities carrying their cost deserve proof that the capacity was necessary. Hats off to Robert Keus, Cas Burggraaf of GreenPT - Chad Gibson and Scott Chamberlin of Neuralwatt for putting their necks out and lending their technology expertise to this debate that seem more confused by the day.

Joint Statement Neuralwatt × GreenPT: Build AI lean first. Then build the power it actually needs. Last week, the The New York Times reported that at least 82 gas-burning power plants are being built or proposed to power American data centers, most of them permitted in the last 18 months. One of them, outside El Paso, will emit as much carbon in a year as 360,000 cars. The people breathing those emissions are rarely the ones building with the AI those plants power. Pollution and climate emergencies have never been distributed evenly, and the costs will largely land in rural counties and border towns where the air is already worse and the power to object is always thinner. To be clear: Neuralwatt and GreenPT are not against AI buildout. We are two companies whose entire work depends on AI succeeding. But this expansion is being built the reflexive way, not the deliberate way. There is tremendous potential for AI to benefit humanity, and it is being undermined by the short-term rush to claim it first. ⚡ AI runs far less efficiently than it could. Peer-reviewed research in Joule found that optimized inference systems already use 4–20× less energy per query than widely cited estimates, and that further improvements across model design, serving systems, and hardware could reduce per-query energy use by 8–20×. These plants are being sized to projected demand that includes enormous inefficiency, which means communities will breathe emissions for power that could have been avoided. 🔥 Gas turbines are the default because they're fast to deploy, not because they're the best option. On-site gas is usually more expensive than grid power, and build costs have risen 66% since 2023. These plants lock in decades of emissions for what's framed as a short-term bridge, but no one says where the bridge ends. Meanwhile, solar and wind with battery storage, small modular reactors, geothermal, and fuel cells are all already powering data centers. The choice isn't between gas and nothing. 🔄 So here's our ask: Reverse the order. Instead of building power first and thinking about efficiency second, cut the waste and make AI leaner, then build for the remaining demand. • It's a sequence that costs nothing to follow. • It's cheaper for the companies building AI. • It's cleaner for the towns hosting it. • And it leaves more power on the grid for everyone else. If you believe AI should be built responsibly, share this. The decisions being made in the next eighteen months will outlive every model we're excited about today, and the communities carrying their cost deserve proof that the capacity was necessary. Robert Keus, Founder/CEO, GreenPT Cas Burggraaf, Co-founder/CTO Chad Gibson, Co-founder/CEO, Neuralwatt Scott Chamberlin, Co-founder/CTO, Neuralwatt #SustainableAI #GreenPT #Neuralwatt #AIInfrastructure #CleanEnergy #DataCenters #EnergyEfficiency #ResponsibleAI

After my latest post where I used neuralwatt together with Spring AI, is such an setup also possible for GreenPT which is famous for sustainable AI powered by renewable energy and open-source models. Sure, that works as well with the following application.yaml Have fun

𝗟𝗲𝘀𝘀 𝘁𝗮𝗹𝗸. 𝗟𝗲𝘀𝘀 𝗰𝗼𝗱𝗲. 𝗦𝗮𝗺𝗲 𝗮𝗻𝘀𝘄𝗲𝗿. Today we're shipping GreenPT compression models, the first AI provider endpoints that compress your responses before they ever reach you. Same model. Same price per token. Same CO2 per token. Up to 70% fewer output tokens, because the answer comes back shorter. Choose a family by model id, no extra parameters, no SDK gymnastics: 🌱 𝐠𝐥𝐦-𝟓.𝟐-𝐜𝐚𝐯𝐞𝐦𝐚𝐧 - terser prose. Strips filler, keeps every technical token intact. (-22% to -87%) 🌱 𝐠𝐥𝐦-𝟓.𝟐-𝐩𝐨𝐧𝐲𝐭𝐚𝐢𝐥 - less generated code. Your agent thinks like the laziest senior dev in the room. (-22% to -94%) 🌱 𝐠𝐥𝐦-𝟓.𝟐-𝐡𝐨𝐧𝐞𝐲 - both, for mixed workloads. (-6% to -77%) Each comes in -lite, default and -ultra intensities. Swap the model id in your existing OpenAI-compatible request and you're done. { "model": "glm-5.2-ponytail", "messages": [{ "role": "user", "content": "Add retry with backoff to this fetch call." }] } Works in OpenCode, Cursor, Claude, and every coding app that speaks the OpenAI API. The compression is baked into the endpoint itself. 𝐓𝐡𝐞 𝐬𝐭𝐨𝐫𝐲 𝐛𝐞𝐡𝐢𝐧𝐝 𝐭𝐡𝐞 𝐧𝐚𝐦𝐞𝐬 We didn't invent these rulesets. Two brilliant open-source builders did: 🪨 Caveman, by Julius Brussee, makes your agent talk like a caveman: same answers, roughly 65% fewer output tokens, code and errors byte-for-byte exact. Brain still big. Mouth small. A March 2026 paper found brevity can actually improve accuracy by around 26 points. 94.5k stars on GitHub and counting. 🎀 Ponytail, by Dietrich Gebert, makes your agent code like the senior dev who's been there longer than the version control. YAGNI first, stdlib before custom code, never compromise on validation or security. Around 54% less code, up to 94% on over-builds, 100% safe. 92k stars. GreenPT is now the first listed sponsor of Ponytail, and both rulesets ship natively on our endpoints. Point your agent at GreenPT and the compression is already there. 𝐖𝐡𝐲 𝐢𝐭 𝐦𝐚𝐭𝐭𝐞𝐫𝐬 Same answer, delivered with less energy. Fewer output tokens means less compute, less CO2 and a faster response, at no extra cost to you. The saving comes entirely from generating fewer tokens, and we pass all of it through. 𝐀𝐛𝐨𝐮𝐭 𝐭𝐡𝐞 𝐯𝐢𝐝𝐞𝐨 Yes, we made a video about efficient AI using AI. We get that that's funny. We couldn't resist: the meeting between a caveman and a senior dev practically begs for a video, and once we saw the cinematic possibilities we caved (sorry). We're contradicting ourselves this once, and hope you'll let us have the joke. We did try to do it responsibly, with Pruna AI's P-Video, three to six times more energy-efficient than comparable models. Less talk. Less code. Same answer. Same planet. 💚 Docs and links in the comments. #GreenPT #AI #API #OpenAI #Caveman #Ponytail #TokenCompression #SustainableAI #GreenTech #OpenCode #CodingAgents

I've audited 278+ websites. Most land in the 40s and 50s out of 100. Then I ran GreenPT and got a number I almost never see. 97.4 out of 100. Grade A. 0.15g of CO₂ per visit. To put that in context: the median site in my dataset emits 1.72g a page view. GreenPT is running about 11× cleaner than the middle of the pack and it's not a static brochure site, it's an AI product. The category you'd expect to be heavy. What I loved is that it wasn't an accident. Robert Keus and the team built GreenPT as privacy-friendly, sustainable AI from the start so the clean website isn't a marketing layer bolted on top. It's the mission, showing up in the code. That's the whole point of this series. I'm not here to name-and-shame. I'm here to show that a genuinely low-carbon web experience is possible and to point at the teams already doing it, so the rest of us have something to aim at. Curious where your own site lands? Comment AUDIT and I'll run yours and send the breakdown. 🌱 #ClimateUX #DigitalSustainability

🟢 New on GreenPT: DeepSeek V4 Flash (0731) Frontier-class intelligence no longer requires frontier-class compute. On the Artificial Analysis Intelligence Index v4.1, DeepSeek V4 Flash scores 50, level with GLM-5.2 and just behind Claude Sonnet 5 (53). It does this with only ~13B active parameters, a fraction of what comparable models activate per token. Fewer active parameters means less compute per request, and less energy. And with this model that compounds: because it reasons at a higher level, it needs fewer queries to reach the right answer. Less energy per request × fewer requests per task, that's the efficiency gain that matters. At €0.14 / €0.35 per 1M tokens (input/output), DeepSeek V4 Flash is: ▪️ ~95% / ~98% cheaper than Claude Sonnet 5 ▪️ ~97% / ~99% cheaper than Claude Opus 5 ▪️ 87% cheaper on input and 92% on output than our own GLM-5.2 Live now on the GreenPT API: `deepseek-v4-flash-0731`, 1M context, tool use, reasoning built in. Docs: docs.greenpt.ai #AI #SustainableAI #GreenAI #LLM #DeepSeek

If you've landed here recently, welcome. 💚 Short version of who we are: GreenPT is an AI platform for people who don't want to choose between using AI and being honest about what it costs. We respect your privacy, your data never trains a model. Every conversation runs on renewable energy. And you see exactly what it emits, per chat, per day, per month. We built GreenPT because the AI most people use trains on what you type and hides its environmental footprint. That felt like something worth fixing. A few things worth knowing: 🌱 We respect your privacy, your data never trains a model 🌱 Every conversation shows its real energy and CO2 cost 🌱 We run open-source models on our own infrastructure 🌱 Powered by 100% renewable energy If any of this is up your alley, follow along or try us for 14 days. 💚

Robert doesn't say "changes how we look at AI" often. When he does, it's worth 15 minutes of your time. If you build with APIs and want to be one of the first to see what we're cooking, drop him a PM. €50 in credits, on us. 💚