I am super excited to finally announce that we ( @towards_AI ) released our first independent industry-focussed course: From Beginner to Advanced LLM Developer. Put a dozen experts (frustrated ex-PhDs, graduates and industry) and a year of dedicated work, and you get the most practical and in-depth LLM Developer course out there (~90 lessons). It is a one-stop conversion for software developers, machine learning engineers, data scientists, aspiring founders or AI/Computer Science students. We think many millions of LLM Developers will be needed to build reliable customised products on top of foundation LLMs and achieve mass GenAI adoption at companies. We want to help you lead this new field! One of the reasons I quit my PhD in AI (one year + 8 days ago) was to build practical solutions that will help others in the real world and improve what exists. Along with our book “Building LLMs for Production,” this course is our attempt to achieve this goal. While I love the world of academia, when I first stepped into the startup world in 2019, it felt like I pretty much knew nothing all over again. I needed to understand the problems in the real-world application of AI and build solutions for it. Not just research but real models, real products, and real people using them. But here’s the thing: grasping these challenges is merely the first step. For the ‘how’ part of it, you need to get into the code, the architecture, the models, the APIs, the deployments, the trials and errors, and the complex and wide varieties of frameworks - you don’t have time to reinvent the wheel in a startup! This discrepancy between academia and the industry is even worse in the LLM era. So we’ve gathered everything we worked on building products and AI systems and put them into one super practical industry-focused course. Right now, this means working with Python, OpenAI, Llama 3, Gemini, Perplexity, LlamaIndex, Gradio, and many other amazing tools (we are unaffiliated and will introduce all the best LLM tool options). It also means learning many new non technical skills and habits unique to the world of LLMs. Even though the course is super practical (oriented towards building a real-world project), we believe the course teaches concepts that will stay relevant for a long time even as LLMs get better, such as reducing hallucinations, customising to specific companies and tasks, teaching how to work with them, some cool theory, practical tips and more. The only skill required for the course is some Python (or programming) knowledge. We cover the full stack of learning to build on top of foundation LLMs - from choosing a suitable LLM application to collecting data, iterating on many advanced techniques (RAG, fine-tuning, agents, audio, caching), integrating industry expertise and deploying. Our students will create a working product, which we certify, and we also provide instructor support in our dedicated Discord channel. This could become the seed of a startup, a new tool at your company or portfolio project for your LLM job applications. You can find all the lesson titles and more information on the course page (or DM me). I also want to thank some of our amazing team at Towards AI that helped achieve this, including @omar_solano1 , Fabio , Rucha , @_LouiePeters , Arjun, Jaiganesan and everyone in the team that contributed to this course. A big thanks to some of our reviewers and my friends for the support, @daansan_ml , @SerranoAcademy , @rohanpaul_ai , @tunguz , @antgrasso , @ceobillionaire , @Scobleizer , @swyx , @hasantoxr and many others... “From Beginner to Advanced LLM Developer” is now available on the Towards AI Academy: tinyurl.com/5n9y32y4 #llms #llm #llmops #llmdeveloper

Are open models catching up? The best open-weight model vs the best closed model on our internal writing benchmark, by release date. An open model, Kimi K2, briefly took #1 in September 2025. The gap shrank to just 101 Elo in February. Then Claude Opus 5.5 blew it back open to 447. Still, the best open writer today, GLM-5.3, costs 30x less per script.

Opus 5.5 managing ElevenLabs' API with a cloned voice is an extremely powerful and mind-blowing combination

Protesting is not the right move… AI will integrate our lives wether we want it or not. The thing we should control is how will it affect our lives. Something we can actually have more control over. We use calculators when we want to. I don’t use it for calculating tip as I learned how to do that mentally. But I use one for very important calculations (well, used to use one, now ChatGPT uses it for me). I believe AI is and will cause a big transformation. But so did the calculators, electricity and the internet. We cannot ask to pause the progress or to slow down. It’s simply unrealistic. But we can create the guardrails and shape laws to optimise AI’s integration into our society. Many companies are working in this, and hopefully governments will invest even more there. Not to push or restrict progress, but to evolve and iterate with it. We should and can act. It’s not too late, and is much more powerful than protesting (I think).

as per popular demand, here's every Opus 5.5 effort setting (on our internal writing benchmark, 167 model config in total:) max: #1, 2600 Elo, $3.43 and 17 min per script xhigh: #2, 2399 Elo, $0.86, 4.4 min high: #3, 2342 Elo, $0.34, 1.7 min adaptive (no effort flag): #6, 2264 Elo, $0.21, 1.0 min medium: #10, 2199 Elo, $0.21, 1.1 min low: #20, 2144 Elo, $0.17, 0.8 min The previous #1, Claude Fable 5.1 on max, sits at 2303 for $3.15 per script. Some high-level insights: - xhigh beats Fable by almost 100 Elo at about 1/4 of the price - high beats it too, at roughly 1/9 of the price - max is 200 Elo above everything else, but it costs 4x xhigh and takes 17 minutes per script - If you pass the no-effort flag with a Claude subagent, Opus 5.5 picks for itself and lands between medium and high For most writing work, high seems to be the sweet spot. Use Max when you want the best draft and can wait. I assume Max is great here because of our detailed, extensive prompt. 10 script tasks, 5 scripts each, scored blind by three AI judges from three different labs.

Our internal writing benchmark uses 3 AI judges: Claude, GPT and DeepSeek. So do they favor their own family? The GPT judge scores OpenAI models 2.3 points higher (out of 100) than the other two judges do. The DeepSeek judge gives DeepSeek models +1.7. The Claude judge gives Anthropic only +0.4... but Google +1.3 😅 Hoping to add Jev as a fourth soon... waiting on @typesafeai ;)

Pick your LLM Writer 😂 (One-shot prompt to Opus 5.5 asking it to make a nice/fun and useful 8-bit animation from the benchmark; super unclear prompt, haha)

You can pay 429x more for writing that is 1.7% better. Early results from our internal creative writing benchmark: we have models write full YouTube scripts for our channel, 10 real tasks, 5 scripts each, scored out of 100 by three AI judges against our own edited references. GLM-5.3 Flash writes one script for $0.0074 and scores 88.2. Claude Fable 5.1 at max effort scores 89.7 and costs $3.15. That is the best score we measured under a cent, against the best score we measured at any price. Plenty of pricier setups score worse, so spending more does not buy that gap on its own. I would happily pay Fable money if it cut my editing time. That is the part worth measuring from your own experience: same prompts, model names hidden, then count the scripts you would actually publish and the minutes you spend fixing them.

Opus 5.5 is the worse release since ChatGPT I need to wake up at 5am tomorrow and can’t go to sleep It’s so fun to one shot tons of amazing stuff 🤯 It almost feels like we’ve just unlocked a next level of intelligence/understanding requiring much less detailed instructions

GPT-6 Astra is OpenAI's flagship. GPT-6 Sol is the cheaper model positioned below it. On our writing benchmark, at each model's deepest effort setting: Sol scores 2182 Elo. Astra scores 2013. Sol costs $0.11 a script. Astra costs $0.85. Sol takes 171 seconds. Astra takes 526. The cheap model wins by 169 Elo while costing an eighth as much and running three times faster. Every effort setting of Sol we tested beats every effort setting of Astra. Astra's default setting sits at 1843, which is below where GPT-5.6 Sol was. this measures writing YouTube scripts in one editorial voice. Astra is built for agentic and engineering work, and nothing here says it is bad at those. It does say that paying the flagship price for prose is the wrong call. Check the cheap model against your own task before you assume the expensive one is better.

GPT-6 Sol is BETTER than Astra on our benchmark! Much better! We ran it over the same 10 YouTube scripts every other model on our board writes, five drafts each, scored blind by three AI judges. It is 6th out of 94 models, at $0.11 a script. Three things stand out… Its effort dial actually works. Default scores 2058, high 2086, max 2182. That is 124 Elo for roughly double the price, which is the widest effort spread we have measured on any model this month and almost the only one consistent getting better results the more it thinks. It beats GPT-5.6 Sol and Astra while costing a third as much than 5.6 Sol and much much less than Astra. And it beats GPT-6 Astra by far. More on that in the next post. if you were using GPT-5.6 Sol (or Astra) for drafting, this is a straight upgrade on both quality and price. Run it at max effort. Very nice upgrade by @OpenAIDevs !

An average score hides what a model is actually good at. So here are those same six lab leaders on eight axes: six writing metrics, plus what they cost and how fast they are. Nobody wins everywhere. The shapes are all very different. Grok 4.7 has the cleanest prose of the six and the best tone of voice. It is also last of the six on writing quality, and you can see that trade in the notch in its shape. GPT is the best of the group on screen cues and on factual substance, and the weakest on tone. Claude Fable 5.1 has the most complete shape and costs nearly 9x the next dearest model here. DeepSeek V4.1 Flash is extremely fast and cheap. The overall score separates these six by 2.2%. The individual axes separate them by up to 5.5%. If you are choosing a writer, choose on the axis you actually need, not on the leaderboard position.

We took the best writer each of the top labs has and ran them over the same 10 YouTube script generations (creative writing). Three numbers decide which one we'd actually use at scale: what it scores, what it costs, how long you wait. The whole frontier field is 2.2% points wide. The price range across it is 248x. DeepSeek V4.1 Flash scores 88.0 for 1.3 cents in 76 seconds. Claude Fable 5.1 scores 89.7, the best writing we have measured anywhere, for $3.15 and ten minutes. That is 1.7 points for 248 times the money and 8 times the wait. Grok 4.7 shipped this week, which puts xAI in the middle of this group on price and near the bottom on speed. The expensive end of this chart buys you very little average quality