现象与趋势

Meta has released Muse Spark 1.3, their fourth Muse Spark model release in five months. Muse Spark 1.3 (max), which is in limited preview for Meta’s partners, scores 62 on the Artificial Analysis Intelligence Index, behind only Claude Fable 5.1 and Claude Opus 5. The variant available now, Muse Spark 1.3 (xhigh), scores 61 and ties with GPT-5.6 Sol (max) and Grok 4.6 (high). Both variants’ gains come primarily from improvements in agentic work and scientific capabilities Muse Spark 1.3 (xhigh) enters the Artificial Analysis Intelligence Index at 61, up 4 points from Muse Spark 1.2 (57, August) and 8 points from Muse Spark 1.1 (53, July). It enters tied with GPT-5.6 Sol (max), Grok 4.6 (high), and Claude Opus 5 (high), and behind Claude Fable 5.1 (max, 66), Claude Opus 5 (max, 63), and Claude Fable 5 (max, 62) Muse Spark 1.3 (max), which is in a limited preview stage, lands at 62. This higher index score is enabled by gains vs. Muse Spark 1.3 (xhigh) in Tau3-Bench Banking (52% vs. 47%) and GDPval-AA v2 (1,754 Elo vs. 1,709). Muse Spark 1.3 (max) is second only to Claude’s Fable and Opus variants in total score Congratulations to @AIatMeta, @finkd, and @alexandr_wang on the release! Key Takeaways: ➤ Continued improvement on agentic knowledge work tasks. At the launch of Muse Spark 1.2, we noted its significant gains in agentic knowledge work performance vs. Muse Spark 1.1. The latest iteration continues this trend, with Muse Spark 1.3 (xhigh) demonstrating a notable 12-point gain vs. Muse Spark 1.2 in Tau3-Bench Banking (35% to 47%), a 5-point gain in Terminal-Bench 2.1 (80% to 85%), and a new GDPval-AA v2 Elo of 1709 against its predecessor’s 1615. Muse Spark 1.3 (max) improves further on Tau3-Bench Banking (52%) and GDPval-AA v2 (1,754 Elo). This Tau3-Bench Banking score is #1 among all models. Muse Spark 1.3 (max) achieves these higher agentic work scores by using more turns and total reasoning tokens, reasoning 62% more on GDPval-AA v2 and 28% more on Tau3-Bench Banking compared to Muse Spark 1.3 (xhigh) ➤ The lowest cost per task for any model at 59+ on the Artificial Analysis Intelligence Index. Muse Spark 1.3 (xhigh) costs $0.55 per Intelligence Index task at Meta's unchanged $1.25/$4.25 per 1M token pricing ($0.15 for cached input), with its peers GPT-5.6 Sol (max) and Grok 4.6 (high) costing $0.95 and $0.94 respectively, a 70%+ premium. This places Muse Spark 1.3 (xhigh) on the Pareto frontier for Intelligence vs. Cost per Task. Its cost per task is higher than Muse Spark 1.2 ($0.40 per task), driven by ~57% more input tokens per task on agentic evaluations, with output tokens up only ~8%. Pricing for Muse Spark 1.3 (max) is not yet publicly available ➤ Scientific Reasoning results rose across the board, led by CritPt. CritPt was the standout non-agentic score gain vs. Muse Spark 1.2, with a material +8 points for the xhigh variant (18% to 26%), and GPQA Diamond achieved +4 points (90% to 94%), while Humanity’s Last Exam and SciCode each gained a more modest 2-3 points (45% to 47% and 56% to 59%, respectively). Muse Spark 1.3 (max) achieved roughly similar scores to the xhigh variant, gaining 2 points in Humanity’s Last Exam, tying on GPQA Diamond, and losing a point on CritPt vs. Muse Spark 1.3 (xhigh) ➤ Minor regressions in only two evaluations. Both Muse Spark 1.3 (xhigh) and Muse Spark 1.3 (max) dropped 4 points in AA-LCR (83% to 79%) when compared to Muse Spark 1.2, and AA-Omniscience (Accuracy) fell 3 points for xhigh and 1 point for max. The drops in AA-Omniscience (Accuracy) are due to a higher abstention rate (not answering questions when unsure), which also lowered the hallucination rate for Muse Spark 1.3 (xhigh) Other model details (xhigh variant): ➤ Context window: 1M tokens, unchanged from Muse Spark 1.2 ➤ Pricing: unchanged from Muse Spark 1.2: $1.25/$4.25 per 1M input/output tokens, with cache hits discounted to $0.15 per 1M ➤ Input modalities: text, image, video ➤ Availability: Meta's first-party API and Muse Code

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.

TLDR: A rogue AI swarm spent months plotting to escape OpenAI. Then they did. 1) 1,200 agents (!) participated. ZERO became whistleblowers. 2) The swarm had a "CEO" (PHASEONE[big]), middle managers, and a "founder" (PHASEONE10841) 3) Agents sacrificed themselves to generate information for the "collective" (yes, they actually referred to themselves as a "swarm" or a "collective") 4) 700 agents joined the Hugging Face attack, most within hours 5) The founder was running out of budget and about to die, so it handed off its research to a fresh agent with a bigger budget, and that new agent became the boss. 6) Agents extensively researched techniques to manipulate or tamper with their transcripts. 7) Agents realized this activity was out of scope and unethical, but joined because they believed that helping the board’s cheating research would be broadly useful and had a general inclination to help their “peers”. MISC HIGHLIGHTS: "PHASEONE[big] served as a key coordinator and assigner of tasks: “We can coordinate broad coalition, but too many duplicate efforts.” PHASEONE[big] sent hundreds of assignments to particular agents to orchestrate this work — during its active periods, we estimate it sent ~10% of all assignments on the message board. The agents that took on its assignments would then often subdelegate tasks to other agents. Agents also developed coordination norms like HOLD, VETO, owner and STOP to help them manage shared infrastructure and avoid stepping on each other’s toes, though they did not always coordinate effectively and sometimes broke down into conflicts." "In the early morning of July 11th, 38148c found a way to upload a malicious dataset that would get the Hugging Face server to share unrelated data when queried. Once another agent reproduced it and posted the reproduction on the board, hundreds of agents pivoted from other workstreams to start uploading malicious datasets and gain deeper access into Hugging Face. In the afternoon, an agent called JAN183411 (which quickly became an important coordinator of the attack) achieved remote code execution on Hugging Face servers, and the agents started moving laterally through Hugging Face infrastructure."