Have you tested the codex token-saving tricks everyone keeps recommending? 🫣Do they actually work? 📒I found ~20 of them from github, x, youtube, and short-video tutorials. (Caveman. Ponytail. MarkItDown. AGENTS.md. /clear. prompt caching...) 📢some claim 80% savings, but is that actually true?🧐 so I’m running a 20-day test: same task style. real screenshots. usage numbers when available. kept / killed verdict. The survivors become a step-by-step handbook at the end. Follow if you want REAL results.🤑 #Ai #ClaudeAI #Codex #Chatgpt #Tokenusage

Introducing WeirdML v3, a fully agentic benchmark featuring 11 complex hand-made tasks. Models must explore and understand unfamiliar data, develop ML and data analysis pipelines and produce results despite limited data, unspecified goals and/or very limited feedback. 1/8

Yet another big update for GLM 5.3 Flash EXL3 ⚡️ Performance improvements: On 2x DGX Sparks ~13% faster on prose single stream ~23% faster on prose 2-4 concurrent streams ~36-37 tok/s on prose single stream ~75-67 tok/s on prose in 4 concurrent streams On 3x DGX Sparks ~15% faster on prose single stream ~23% faster on prose 2-4 concurrent streams ~41 tok/s on prose single stream ~88 tok/s on prose in 4 concurrent streams No change for TP=4. New features and fixes for all: - Much faster loading time for the weights! From more than 300s to now 55s. - Fixed an issue that after a long conversation, the next message was sometimes treated like a brand-new prompt and had to be re-read from scratch. Get it here: github.com/MiaAI-Lab/GLM-5.3…

the main thing i was excited about launching this week will be next week instead, but imo worth the wait!

We’re sharing how GLM-5.3 helped build and optimize the inference infrastructure serving GLM-5.3-Flash. The system went from its first successful run to production readiness in less than two weeks, with end-to-end throughput tripling relative to the initial baseline. The key was dense feedback: local correctness tests, execution traces, microbenchmarks, and end-to-end measurements that enabled targeted hypothesis testing rather than reliance on aggregate performance metrics alone. z.ai/blog/glm-built-its-infe…

The Need for Speed ⚡️ INSANE DeepSeek v4.1 Flash update on 2x DGX Spark 🔥 +33 FASTER prose on single stream 42 tokens per sec vs old 32 50% FASTER prose on 2 concurrent streams 63.7 tokens per sec vs old 42.5 That puts 2x setup at roughly the same speed as DeepSeek v4.1 Flash running on 3–4x DGX Sparks before the update 🤯 Get it here: github.com/MiaAI-Lab/DeepSee…

OpenAI and Anthropic need to drop GPT 6 Sol and Opus 5.2 sooner rather than later. Fable 5.1 and GPT 6 Astra are incredible models. They are also unusable. Even with multiple subscriptions. I have four OpenAI accounts and two Claude Max plans and I still hit limits within days. We do not need smarter models. We need models smart enough to build with that we can actually afford to run.

fable 5.1 just got mogged by minimax tested both with same prompt at xhigh reasoning but results surprised me > minimax m3 took around 4 minutes and cost $0.40 > fable 5.1 took around 45 minutes and cost $17 no idea what fable was doing for 45 minutes when minimax got it done in 4, while being so cheap which one did better here?

i am looking forward to grok 4.7 and 4.8. grok 4.6 is pretty much the only model that’s steerable, fast, and doesn’t over engineer the world by default

Grok 4.7 should be roughly on par with Opus 5.0, not 5.1. Better in some ways, worse in others. We need to fix multimodal performance. Grok 4.8 will be a noticeable improvement. Grok 4.9 is probably Astra/Fable class. Grok 5 maybe better than anything. We shall see.

Independent evaluation of AI models, across both capability and safety, is essential. We welcome Dario’s call to give independent evaluators greater access to frontier AI models. For nearly three years, Artificial Analysis has been building benchmarks and infrastructure to independently measure AI capabilities. We have supported pre-launch benchmarking of frontier models with almost every major AI lab. We are continuing to see jumps in capability across every dimension we measure. The world needs a vibrant ecosystem of independent AI evaluators. We are going to keep working to build it!

Run DeepSeek v4.1 Flash on 2x DGX Sparks ✨ One of the BEST models you can run today, now available for only two dgx sparks. Ships with conservative settings for stabiliy: - 600k context, 775k KV cache pool - 2 concurrent connections - 550k prefill stress test passed - EXL3 quantization Performance: ~32 tok/s on prose single stream ~42 tok/s on prose 2 concurrent streams ~1000 tok/s prefill 8k-128k ~872 tok/s prefill on 256k Get it here: github.com/MiaAI-Lab/DeepSee…

Wow. Turns out another OpenAI agent swarm was busy spamming and exploiting RubyGems way back in May, within days of the previously uncovered Wiki attacks: simonwillison.net/2026/Sep/1…

We independently benchmarked Devin Fusion for its release today - this is the first time a multi-model coding agent has been included on the Artificial Analysis Coding Agent Index, and it effectively retains Claude Fable 5.1 and GPT-6 Astra performance while reducing costs Devin Fusion runs a frontier lead model with a cost-efficient sidekick. We tested configurations from Cognition combining frontier models from Anthropic and OpenAI with their new SWE-2 (medium) as a sidekick model. Configured with Claude Fable 5.1 (xhigh) + SWE-2 (medium), Devin Fusion scores 62 on the Coding Agent Index v1.5, while with GPT-6 Astra (xhigh) + SWE-2 (medium) it scores 59. The Fable configuration has the higher score, while the Astra configuration is 43% less expensive and completes tasks 31% faster. Congratulations to @cognition on the release! See below for our results and analysis 🧵

Next week we’ll be retiring GPT-5.3-Codex-Spark. Can you believe we shipped a model named as such!! It's had a good run and was a lot of fun, but usage has been declining and we have significantly better models now. Time to make room for the future.

basketball AI (95% local AI + 5% GPT-6 Astra) - detect ball and players - track players - re-identify players across plays - OCR player numbers - recognize player in possession - detect court keypoints - map player positions and trajectories Astra is powerful, but expensive tool. use it when it actually makes a difference. ↓ deep dive

🚀 Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient. 🔹 Introducing the smallest model in our new architecture family, with native visual understanding. 🔹 Designed for greater capability, faster inference, higher throughput, and scaling to larger models. 1/6

DeepSeek V4.1-Flash beats V4-Pro 0813 less than a month after its release. And it's $0.30/$1.20 per 1M input/output. That's roughly: - 1/20th cost of Opus 5 - 1/17th cost of GPT-5.6 Sol - 1/12th cost of Kimi K3 ...while beating all three on Terminal-Bench.

We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics. The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra. The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.