Alternatively, you'll find the blog by going to their home page, scroll down to their news link, find the blog link and scroll down to 26th Feb post, was one page 1 at time of posting:
www.atlascloud.ai/blog-list/1
And @atlas_cloud_ai still have the outputs in their docs here, posted on 26th February:
www.atlascloud.ai/blog/Nano-B...
The original output from @MarsForTech posted 24th February.
x.com/i/status/20...
.@atlas_cloud_ai have been accused of faking API outputs and using @MarsForTech 4k tests generations in their official docs. This is completely unacceptable!!
See below for more info. 🧵
x.com/MarsForTech...
OK, had enough, I've turned off OpenClaw for now. It's definitely the way forward but not ready for me.
x.com/koltregaske...
Thoughts?
x.com/i/status/20...
Yes, I can't see frontier open-weight models continuing forever - these are getting insanely expensive to train, needing pricier hardware and huge electricity.
But China, France and the other countries always find efficiency gains and funding. I think they'll say alive for a while yet.
A new GitHub? OpenAI is building its own repo system to replace Microsoft GitHub after dealing with lots of outages. It’s internal-only for now, but could be made public later.
x.com/i/status/20...
I think we've just found our local model for normies, Qwen 3.5 Small will run on most devices and the models are very capable for their size.
This is the fallback model of fallback models for OpenClaw. Use LM Studio or Ollama to grab it.
x.com/Alibaba_Qwe...
In the US for the last 2 months yes. But this is worldwide for the last 12 months:
x.com/i/status/20...
- Thaler's lawyers argued the ruling could harm AI development in the creative sector during key growth years.
The decision upholds that purely AI-generated material lacks eligibility for US copyright protection.
x.com/Reuters/sta...
- The Copyright Office rejected the application in 2022 because creative works need a human author; lower courts upheld this in 2023 and 2025.
- Trump's administration urged the justices not to hear the case, citing Copyright Act provisions that imply human - not machine - authors.
The US Supreme Court declined to hear an appeal over whether art generated solely by AI can be copyrighted.
- Stephen Thaler from St Charles, Missouri, applied in 2018 for copyright on "A Recent Entrance to Paradise", an image his AI system DABUS created independently.
And lots more coverage from Robert's report here.
x.com/Scobleizer/...
x.com/mattshumer_...
The Pro variant solves extremely difficult problems other models cannot, though weaknesses persist in frontend taste compared with Opus 4.6 and Gemini 3.1 Pro plus occasional real-world context misses; the attached table shows competitive benchmark scores across agentic and coding tasks.
GPT-5.4 standard mode with heavy thinking has outperformed previous Pro versions for Matt after a week of testing, making the choice of model feel almost over and reducing his use of Pro entirely while delivering faster results with fewer reasoning tokens.
Matt Shumer says it's "The best model in the world."
Dan's team at Every put the model through code reviews and workflows where it offered deeper analysis and conversational feedback, though it sometimes over-expanded scope or marked tasks complete prematurely.
x.com/danshipper/...
GPT-5.4 beat Codex 5.3 and Opus 4.6 in head-to-head planning tests on real engineering tasks, delivering thorough plans with a human feel while proving fast and about half the price of Opus 4.6 in OpenClaw use, accoridn to Dan Shipper.
It leads in corporate law (57.5% mean), plans professionally with high tool use, but retains some junior analyst errors like overthinking or distraction.
x.com/mercor_ai/s...
OpenAI GPT 5.4 tops Mercor APEX-Agents leaderboard at 35.9% Pass@1 and 52.5% mean score, first model to exceed 50% mean and up 15%+ Pass@1 in under three months.
GPT-5.4-high has joined the Text Arena tied with Gemini-3-Pro, with top-3 creative writing and top-10 instruction following plus hard prompts.
x.com/arena/statu...
The benchmark evaluates production of full working applications from short text specifications with gains from enhanced verification and tool use. The model also leads ProofBench and IOI while ranking fourth on the Vals Index.
x.com/ValsAI/stat...
GPT-5.4 tops Vibe Code Bench at 67.4% accuracy, 5.7 percentage points higher than the previous SOTA on this application-building test.
GPT-5.4 achieves 74.0% on ARC-AGI-2 at $1.52 per task with GPT-5.4 Pro reaching 83.3% at $16.41 per task.
Scores were measured across low, medium, high and xHigh compute levels for both ARC-AGI-1 and ARC-AGI-2.
x.com/arcprize/st...
GPT-5.4 Pro set a new record on FrontierMath with 50% on Tiers 1-3 and 38% on Tier 4.
x.com/EpochAIRese...
GPT-5.4 becomes OpenAI's most createive model in terms of Design Arena's metrics.
x.com/Designarena...
GPT-5.4 runs 1.5x faster with the same intelligence and reasoning when you use /fast.
x.com/OpenAIDevs/...
A showcase of what builders have made with GPT-5.4.
x.com/OpenAIDevs/...
