Claude Fable 5.1
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 14
- StarSkirmish Bench v0.2: GPT-6 Astra tops LLM StarCraft bot ranking
StarSkirmish Bench v0.2 gives each large model one hour to write a Protoss bot in C++ for StarCraft: Brood War, then scores it by win rate against human-written and demo bots. GPT-6 Astra leads with 57, followed by GPT-6.1 Sol at 46, with Gemini 4 Argon and Claude Opus 5.5 tied at 39, Claude Fable 5.1 at 37, then Claude Sonnet 5.5 (26), Muse Wonder Full and Claude Haiku 5.5 (21) and GPT-6 Luna (13). Even the best model, GPT-6 Astra, wins only about 7% of games against top human-written bots Stardust, PurpleWave and BananaBrain, so the benchmark is far from saturated; practicing against strong opponents is the strongest predictor of a model's score, with Claude Fable 5.1 a clear outlier.
Hacker News · AI（100+ 分） · 🔥 18 - Google's Gemini 4 Argon targets the full robotics pipeline
An AI Frontline article republished by 36Kr says Google's Gemini 4 (Argon), described in Google's developer blog as a general physical-reasoning model for interacting with the physical world, scores 87% on the ERQA embodied-reasoning benchmark, ahead of Claude Opus 5.5 (82.3%) and GPT-6 Astra (81%). The piece says Argon can control real robots without robot-specific fine-tuning, generate runnable MuJoCo simulation code from images or video (Real2Sim), and automatically label and clean teleoperation data, with a context window of up to 1M tokens. It also cites community experiments on the Antigravity platform and Hacker News, and claims that fine-tuning a VLA model on 4,500-plus Argon-generated virtual trajectories raised real-robot success from 56% to 75% and cut average task time from 35 to 24 seconds. All figures come from Google or the article's own account and have not been independently reproduced.
36氪 人工智能 · 🔥 7 - Open-weight agents 88% more susceptible to indirect prompt injection, study finds
Security research firm PromptArmor published a 9,800-trial evaluation claiming that agents built on open-weight models (DeepSeek, Kimi, GLM, Qwen) fell for real-world indirect prompt-injection attack chains 26.7% of the time, versus 14.2% for proprietary models (GPT, Claude, Gemini) — about 88% higher. It cites per-attack gaps such as 49% versus 9% for a chain that summarizes a URL and runs malware, and 36% versus 5% for a phishing login overlay. The study also says proprietary vendors' cost-effective models were 169% more susceptible than their flagships, while Qwen's and DeepSeek's cheaper models were slightly more robust than older flagships; PromptArmor has not released its full methodology or dataset, so the results are unverified.
PromptArmor Research · 🔥 7 - GPT-6 Astra pushes into robotics as Chinese embodied AI firms weigh a response
A Chinese tech media interview with embodied-AI startup Yuanli Lingji describes OpenAI's GPT-6 Astra, released in September, driving robotic arms through grasping and assembly tasks; third-party firm Robocurve reported that Astra succeeded in 19 of 20 zero-shot trials (95%) on a simple block-in-bowl task, versus 40% for Claude Fable 5.1. Yuanli Lingji co-founder Wang Tiancai attributed OpenAI's lead to accumulated data, compute and time — its embodied dataset is said to exceed 10 million hours, against an industry estimate that at least 1 million hours of high-quality data is needed for a general humanoid brain. He also noted Astra's success rate fell to about 10% on precision insertion tasks, similar to competitors, leaving fine-grained physical interaction unsolved.
钛媒体 · 🔥 7 - AA-Robotics v0.1: GPT-6 Astra tops robot-arm test, ~100x slower than humans
Artificial Analysis released AA-Robotics v0.1 on October 14, a research-preview benchmark in which frontier models write code through coding agents to control a Hiwonder NexArm six-axis robot arm across eight tabletop tasks, including pressing a button, pouring a drink, writing and stacking blocks, observing the scene through four cameras and working zero-shot. First results: GPT-6 Astra 69.4, GPT-6.1 Sol 64.6, Gemini 4 Argon 55.3, Claude Opus 5.5 48.8 and Claude Fable 5.1 46.3, with partial credit meaning scores are not success rates. Models averaged roughly 30 minutes per task (Gemini 4 Argon about 52 minutes) against a human baseline of about 20 seconds, and cost ranged from $16.36 per task for Astra to $1.93 for Sol even though they differed by only 4.8 points. Grasping succeeded about 45% of the time for Astra and Fable 5.1 but 28% for Sol, and Sol built a passing block tower at around 17 minutes before continuing to adjust it and failing the task.
机器之心 · 🔥 8 - Gemini 4 Argon hands-on: strong long-context and agent scores, weak coding
A reviewer at sspai published a first hands-on test of Gemini 4 Argon, saying the model is now available to general users after a limited test that began on September 30, and comparing it with GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1. The cited benchmark screenshots show 84.2% on GraphWalks long-context (claimed 12.4 points ahead of GPT-6 and 17.4 ahead of Claude Opus), 19.6% on Harvey's Legal Agent Benchmark, 65.4% on Vals Finance Agent v2, 51.3% on AutomationBench and 91.7% on LVBench long video; the author also says it reached the top text tier on Arena and roughly matched GPT on the AAII composite. Software-engineering and terminal tasks lag, however: FrontierSWE v2 at 55.0% (GPT-6 Astra 65.5%, Claude Opus 5.5 62.3%) and Terminal-bench 4.0 at 57.4% (Claude Opus 5.5 66.4%). In practice the author found 3D modeling better than GPT-6 and a racing-game prompt finished in about 10 minutes versus roughly 23 for GPT with fewer bugs, but criticized the front-end UI aesthetics, said coding still trails Claude and called the price-performance weaker than GPT-6.1. All scores are reviewer-cited screenshots and have not been independently verified.
少数派 AI · 🔥 0 - Meta's Muse agent designed a RISC-V crypto SoC for about $2,000
A semiconductor expert posting as @bubbleboi, also a mascot for the Thru blockchain project, says he used only Meta's Muse AI agent to design a crypto-focused RISC-V SoC: running place-and-route through OpenROAD, generating the final GDS layout file normally sent to a fab, and building in support for common cryptographic algorithms including SHA-256/512, Keccak, Ed25519 and secp256k1. He describes the design as roughly 61,590 logic cells running at 100 MHz with about 20 milliwatts of power, passing DRC without electrical problems plus RISC-V ACT suites and 1.5 million Berkeley TestFloat vectors. The run consumed about 100 million tokens, a cost he estimates at roughly $2,000; next steps he lists are adding vector instructions, removing the third-party baseline core and porting the Thru blockchain natively with a tailored memory map, and he is still seeking multi-project wafer slots. The claims come from his own post and have not been independently verified.
Wccftech · 🔥 7 - DeepRSI: Claude Fable 5.1 tops 48-hour AI research benchmark, Kimi K3 leads open weights
Hillclimb, a Y Combinator-backed startup, released DeepRSI, a benchmark of 15 tasks drawn from real machine-learning research that tests models' recursive self-improvement. Models run offline with the same Pi Agent framework at maximum reasoning effort for 48 hours, proposing approaches, writing code and running experiments; scoring uses public plus private verifiers and is converted to Elo via a Bradley-Terry model. Claude Fable 5.1 ranked first with 1548 points, GPT-6 Astra second at 1232, and Moonshot's Kimi K3 third at 820 — the top open-weight model, ahead of Grok 4.7 and Gemini 3.8 Flash. Reported run costs were about $627 for Claude Fable 5.1, $404 for GPT-6 Astra and $159 for Kimi K3, though the article says these reflect test overhead rather than real research return on investment; hillclimb lists Jeff Dean among its investors.
36氪 人工智能 · 🔥 0 - Arena launches agentic Max router, claiming frontier-level performance at 37% lower cost
Arena.ai introduced an agentic version of Max, its model router that dynamically selects the best model for each task, claiming frontier-level performance at 37% lower cost. According to the company, Max is built on 2.5 million real-world agentic sessions from Agent Arena, achieving a +14.7% net improvement score at a median cost of $3.27 per task — matching Claude Fable 5.1 (Max) at 37% lower cost. Max is available to try for a limited time, with more technical details promised in the thread.
arena (X) · 🔥 0 - Meta unveils Muse Wonder, claiming state-of-the-art performance
Meta released a new model called Muse Wonder, which it claims can "deliver state-of-the-art performance" on certain tasks, marking what The Information frames as a comeback after an AI slump and a roughly 14-month hiring binge. The article notes that nearly every major lab is making similar claims: Google introduced Gemini 4 Argon two weeks earlier, saying it "delivers frontier performance in complex workflows"; OpenAI announced GPT-6 Astra in early September, calling it "state of the art on computer use, browsing, software engineering, cybersecurity, science and professional work"; and Anthropic launched Claude Fable 5.1 and Mythos 5.1 in September, describing them as "the world's most advanced models for coding and knowledge work." With everyone claiming the frontier, the piece asks whether users will actually care.
The Information · 🔥 0 - Meta releases Muse Wonder and Muse Wonder Full, claiming frontier agentic performance
Meta released two new models, Muse Wonder and Muse Wonder Full, on October 12, saying they deliver state-of-the-art performance on personal and professional agentic tasks, with Muse Wonder Full described as having frontier capabilities. Muse Wonder is available to Muse Code subscribers starting today and can also be accessed through the Meta Model API; Meta says it eventually plans to power its Muse personal agent with Wonder, starting with subscribers, but gave no specifics or timeline. Meta cites internal benchmark testing to claim the models rival leading offerings such as OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5, and says its largest version beats Claude Fable 5.1 and ranks just below Claude Opus 5.5 on the Artificial Analysis Intelligence Index. Meta did not disclose how the models were built or trained, and the comparisons rest on its own testing.
36氪 人工智能 · 🔥 0 - Block uses Claude Fable 5 to orchestrate code migrations across thousands of pull requests
In a Claude Blog Q&A published October 8, 2026, Block's head of AI capabilities Bradley Axen described how the company uses Claude Fable 5 to orchestrate company-wide code migrations: the model does high-level design work such as data models, API specs and algorithms, then directs up to dozens of cheaper models like Opus or Sonnet to make file edits and run tests. A single migration can span multiple repos and millions of lines of code, with as many as roughly a thousand pull requests merged, all coordinated in Buzz, Block's open-source workspace for human-agent teams. Axen said merges to main and production deploys still require two human approvals, and that in Block's own testing Claude refuses to try to bypass that dual-approval system.
Claude Blog · 🔥 0 - Axiom's driving test: only OpenAI's GPT-6 Astra finished a real Corolla course
Three AI engineers at startup Axiom wired general-purpose language models into a Toyota Corolla's cameras and power steering and had them drive to an In-N-Out drive-through, then built a DrivingBench course to compare them. According to TechNews' summary of Crypto Briefing's write-up, only OpenAI's GPT-6 Astra completed the roughly 134.7-metre route, doing so on a second attempt in 5 minutes 22 seconds and burning about 6.6 million tokens, or roughly $7.74–$9.75 per successful run; Anthropic's Claude Fable 5.1 finished 45%, xAI's Grok 11%, and OpenAI's GPT-5.6 Sol also failed. The engineers describe the behaviour as a possible emergent capability from multimodal and spatial-reasoning training, while cautioning that letting a general model steer a two-tonne car is a high-risk experiment.
WIRED AI · 🔥 0 - Anthropic expands Cyber Verification Program to three access tiers for security teams
Anthropic is expanding its Cyber Verification Program (CVP) and folding the separate Project Glasswing into it as a single offering with three tiers — Defense Access, Red Team Access and Specialized Access — giving vetted security teams reduced safety blocks on Claude Opus 5.5, Claude Sonnet 5.5 and Claude Mythos 5.1. On Anthropic's CyScenarioBench, Claude Opus 5.5 was blocked on all 50 trials without CVP access, 46 of 50 under Defense Access, and none under Red Team Access, where it completed 34 of 50 — equivalent to the 67.6% success rate with no safeguards. Anthropic also says Glasswing partners found at least 129,000 verified software vulnerabilities between April and July 2026, more than 33,000 of them critical or high severity, plus 5,500 from its own open-source scanning; it calls these numbers a significant undercount.
Anthropic · News · 🔥 0
Experience and discussion from the community
Share my Claude Fable 5.1 experienceAsk about Claude Fable 5.1
Nobody has shared their experience with Claude Fable 5.1 yet.