#推理

🤖🧠NEW PAPER🧠🤖 (The result of an 8-year project!) LLMs seem very different from symbolic systems. Yet LLMs excel in symbolic domains (e.g., language/code/math). How do they do it? Our finding: LLM representations have implicit symbolic structure! Link in thread ⬇️ 1/n

fable 5.1 vs fable 5 vs opus 5 – three lord of the rings landmarks, built in 3d from one image the setup: one reference image per scene, one html file per build, everything procedural – no meshes, no textures, no image files, nothing past @threejs from a cdn. each model reads the picture, writes its own prompt from it, then builds to that prompt in the same turn. three named camera shots per scene on keys 1/2/3, so it can be screen-recorded. run through @openrouter tasks: 1. bag end – hobbiton from two frames, outside and in. the round green door has to open onto the room you are standing in 2. barad-dûr – the tower and orodruin from one film still. the eye has to move and track the camera, the volcano erupts on a cycle, the clouds never stop 3. rivendell – jerry vanderstelt's painting. sun shafts that shimmer, water that falls without a break, trees that sway on a gust models: @anthropicai fable 5.1, fable 5, opus 5 total cost, three builds #1 fable 5 – $14.97 #2 opus 5 – $18.53 #3 fable 5.1 – $22.38 wall clock, three builds #1 fable 5 – 38m #2 fable 5.1 – 92m #3 opus 5 – 122m output tokens #1 fable 5 – 298,592 #2 fable 5.1 – 439,435 #3 opus 5 – 724,418 lines of code shipped #1 fable 5 – 2,885 #2 fable 5.1 – 4,021 #3 opus 5 – 5,161 biggest single build, lines #1 opus 5, bag end – 2,410 #2 fable 5.1, barad-dûr – 1,375 #3 fable 5, bag end – 1,319 observations: • fable 5.1 is the only model that furnished the bag end interior – a live fire, panelling, books on the floor, leaded diamond windows, against fable 5's flat color and opus's dark tunnel. the round door outside opens onto that room, the hard part of the brief • what it costs is thinking room. the 128k output ceiling is a thinking budget in disguise: fable 5.1 burned 102,116 of it on reasoning and hit the wall mid-file. opus spent 109,241 and hit the same wall. fable 5 spent 61,240 and finished bag end in one call – the only one that did • fable 5.1's first pass is not the finished thing. its barad-dûr came back with three defects you only catch by looking at it – nothing a read of the code would have flagged • it is the best of the three at being corrected. handed a plain list of what was wrong, it returned 32 targeted patches over two rounds, every one applied first try, and it worked out one of the causes itself instead of guessing at constants conclusion: nine scenes, 12,067 lines and 1.46m output tokens for $55.88 all in – and the cheapest model was also the fastest, by 3.2x! follow @thehypedotnews for 24/7 ai news, analysis and breakdowns

Muse Spark 1.3 is rolling out today with frontier performance almost too cheap to meter. This is the biggest jump we've made so far on coding and agentic work. Try it in Muse Code and our API. Next up 🍉 and Muse Spark open weights releases coming soon.

Introducing Gemini 3.8 Flash, another jump in Gemini's agentic + coding capabilities, and our 3rd updated Flash model in only 6 weeks... This model has been a ton of fun to work with, excited to see what you all think!

Fable 5.1 from @AnthropicAI on ARC-AGI (Verified): - ARC-AGI-2: 90.0%, $3.12/task - ARC-AGI-1: 97.5%, $1.40/task Across the two benchmarks, Fable 5.1's average cost per task is about 32% lower than Fable 5's, driven by better token efficiency.

OpenAI says Astra is its first model to reach the Critical cybersecurity capability threshold. Under its Preparedness Framework, that means Astra can, with the right tools and access, find unknown flaws and develop exploits across hardened systems without step-by-step human guidance. Hence, OpenAI now says Astra will launch with additional chain-of-thought monitoring, while classifiers can automatically stop potentially unauthorized actions. Astra reached roughly 39% exploit success at ~75K output tokens, while GPT-5.6 Sol is only around 1% there and needs nearly 140K tokens to reach ~12% i.e. Astra is dramatically more capable and token-efficient at exploit development on this internal benchmark. Expert assessments went further: Astra escaped a browser sandbox, executed host commands, and escalated an unprivileged operating-system user to root.

Big news: Qwen3.8-Max-0902 by @Alibaba_Qwen just debuted at #1 overall in the Code Arena: WebDev with 1691 pts! It scores 3 pts above Claude Opus 5 (Max), 17 pts above Kimi K3 (Max), and 22 pts above the previous Qwen3.8-Max. Priced at a blended $5/MToken, Qwen3.8-Max-0902 also claims the highest-scoring position on the Pareto frontier! Stay tuned for a closer look at its Pareto positioning, and for Agent Arena scores coming soon. Its strength carries across every Code Arena: WebDev category: - #1 in Data & Analytics and Consumer Product - #2 in Brand & Marketing, Gaming, and Simulations - #3 in Content Creation Tools and Reference-Based Design Congrats to the @Alibaba_Qwen team on this huge update!

As we prepare to release Astra, we’re focused on making increasingly capable AI safe and broadly accessible. Astra represents a significant advance in cybersecurity capability, reaching the Critical threshold under our Preparedness Framework. We're previewing how we evaluated the model, how its safeguards have advanced alongside its capabilities, and what we'll continue to learn and improve. https://openai.com/index/path-to-astra/

Cache reads with Fable 5.1 cost 75% less than Fable 5’s So the savings get larger as the workload becomes more agentic. Agents constantly revisit the same context, plans, tool outputs, and working state, so cheaper cache reads make persistent reasoning dramatically less expensive; that’s how a 75% cut in one component can become ~45% off the entire workload. The economic question in many enterprises are now “How long can an agent keep working before the job stops making financial sense?” So this curve need to keep falling.