{"slug": "claude-opus-5-5-beats-a-human-baseline-and-ai-researchers-move-up-their-self", "title": "Claude Opus 5.5 beats a human baseline and AI researchers move up their self improvement clock", "summary": "Vals AI moved its forecast for full recursive self-improvement from August 2027 to July 2027 on October 15 after its Recursive Self-Improvement Index found Anthropic's Claude Opus 5.5 beat the published human baseline on one of five research tasks, training a small language model from scratch within a 24-hour compute budget at 0.8308 bits per byte against the 0.85 BPB mark researchers had set. Opus 5.5 scored 0.46 overall on the index, which runs from 0 for a baseline model to 1.0 for theoretical optimum with 0.5 pegged to published research performance, topping four of the five tasks and ahead of Gemini 4 Argon at 0.29, GPT-6 Astra at 0.27 and GPT-6 Sol at 0.20. Vals AI CEO Rayan Krishnan wrote on X that \"the only solution is coordination,\" while noting humans still lead on harness engineering, where annotators hit 93.6% agreement versus 81.46% for the best models.", "body_md": "*Vals AI just watched a Claude model out-train a published human benchmark, and the firm's response was to make its self-improving-AI prediction arrive a month sooner, not later.*\n\nVals AI updated its Recursive Self-Improvement Index on October 15 after testing Claude Opus 5.5, and the independent benchmark firm didn't like what it saw, or rather, it liked it too much. Opus 5.5 topped four of the five research tasks on the index, and on one of them, training a small language model from scratch inside a 24-hour compute budget, it actually beat the published human baseline: 0.8308 bits per byte against the 0.85 BPB mark that researchers had set as the standard to clear. That is the first time any model has outscored a human expert benchmark on an AI-training task, according to Vals AI's own announcement on X.\n\nThe firm's reaction wasn't celebratory. It moved its forecast for full recursive self-improvement, the point where AI systems can design their successors without a human in the loop, from August 2027 to July 2027. One month. But it's the direction that matters, not the size of the move.\n\nVals AI CEO Rayan Krishnan posted the update himself: \"We've updated our timeline for full RSI to July 2027 instead of August after our eval of Opus 5.5. It's telling that the biggest advocate for pacing is still very much racing ahead.\" He was needling Anthropic, which has spent much of the past year positioning itself as the safety-conscious lab, while its own model just moved the industry's self-improvement clock forward.\n\nHere's what the index actually measures. Launched in August, the RSI Index scores models on five tasks that mirror real AI development work: compressing data efficiently, training a language model, engineering an evaluation harness, running post-training, and more. The scale runs from 0 for a baseline model up to 1.0 for theoretical optimum, with 0.5 pegged to published research performance. Opus 5.5 landed at 0.46 overall, the mean of four of those task scores: 0.46 on compression, 0.53 on LM training, 0.15 on harness engineering, and 0.35 on post-training.\n\n[Anthropic Calls Opus 5.5 Pacing the Frontier While It Tops the Benchmarks](https://startupfortune.com/anthropic-calls-opus-55-pacing-the-frontier-while-it-tops-the-benchmarks/)\n\nAnthropic released Claude Opus 5.5 just ten days after CEO Dario Amodei called on AI labs to slow down for safety, but the new model beats Fable 5.1 on every benchmark shown and undercuts GPT-6 Astra on cost while running faster and cheaper than its predecessor. The launch raises the question of whether 'pacing the frontier' is a real slowdown or... - [fastest Claude model among frontier AI options](https://startupfortune.com/anthropic-calls-opus-55-pacing-the-frontier-while-it-tops-the-benchmarks/) - [Opus 5.5 cheaper faster stronger benchmark performance](https://startupfortune.com/anthropic-calls-opus-55-pacing-the-frontier-while-it-tops-the-benchmarks/)\n\nThat 0.46 sounds modest until you see where the competition sits. Gemini 4 Argon, which actually leads Google's broader Vals Index on general agentic tasks, came in at 0.29 on RSI specifically, with GPT-6 Astra at 0.27 and GPT-6 Sol at 0.20. Opus 5.5's weakest category, harness engineering at 0.15, is also where humans still dominate: human annotators hit 93.6% agreement building evaluation harnesses, while the best models managed just 81.46%. Training a model, it turns out, is a problem AI is already good at. Judging whether a model did the job well is still a human skill.\n\nKrishnan has been careful to frame the number itself as less important than the slope. A model clearing 46% of a theoretical ceiling isn't about to run away from its creators. What worries him is how fast that number has moved, and how little outside visibility there is into the handful of labs doing this work. \"The only solution is coordination,\" he wrote on X, a pointed jab given that Anthropic's own researchers, through co-founder Jack Clark, have separately pulled their internal full-RSI forecast forward too, from late 2028 to sometime in 2027.\n\nThat alignment between an outside benchmark and an insider's gut call is the real story here. Two parties with different incentives, a third-party eval shop and a lab's own co-founder, landed on roughly the same year independently. Anthropic's index has risen by roughly 0.30 points annually on Vals' tracking, a trend line that would cross 0.6 right around July 2027 if it holds.\n\nNone of this means a model is about to start writing its own training code unsupervised next summer. Opus 5.5 still needed 24 hours and a fixed budget to beat a benchmark researchers had already published. That's a controlled replication, not an open-ended breakthrough. But replication at this speed, on this specific task, is exactly the kind of signal that moves a firm like Vals AI to shave a month off its clock rather than add one. Mustafa Suleyman, Microsoft's AI chief, has made a similar argument in recent months about AI capability compounding faster than most roadmaps assume. The debate over whether compute or research velocity is the real constraint on AI progress is no longer theoretical. Vals AI just put a date on one side of it.\n\n**Also read:** [Anthropic offers Blackstone a double-digit yield to fund its chip habit](https://startupfortune.com/anthropic-offers-blackstone-a-double-digit-yield-to-fund-its-chip-habit/) • [TypeSafe's Jev becomes Vercel AI Gateway's fastest-adopted model ever](https://startupfortune.com/typesafes-jev-becomes-vercel-ai-gateways-fastest-adopted-model-ever/) • [PS5LM Plus now runs AI models natively on a jailbroken PS5's GPU](https://startupfortune.com/ps5lm-plus-now-runs-ai-models-natively-on-a-jailbroken-ps5s-gpu/)\n\n*This article is posted in [AI News](https://startupfortune.com/category/ai/), check it out for more related stories.*\n\n## Join the discussion\n\n[Open in the community →](https://startupfortune.com/community/)\n\nAlmost there. Sign in and your reply posts straight away.", "url": "https://wpnews.pro/news/claude-opus-5-5-beats-a-human-baseline-and-ai-researchers-move-up-their-self", "canonical_source": "https://startupfortune.com/claude-opus-55-beats-a-human-baseline-and-ai-researchers-move-up-their-self-improvement-clock/", "published_at": "2026-10-17 03:56:04+00:00", "updated_at": "2026-10-17 04:31:43.561667+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-research", "ai-safety", "large-language-models", "ai-agents"], "entities": ["Vals AI", "Claude Opus 5.5", "Anthropic", "Rayan Krishnan", "Gemini 4 Argon", "GPT-6 Astra", "GPT-6 Sol", "Dario Amodei"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/claude-opus-5-5-beats-a-human-baseline-and-ai-researchers-move-up-their-self", "markdown": "https://wpnews.pro/news/claude-opus-5-5-beats-a-human-baseline-and-ai-researchers-move-up-their-self.md", "text": "https://wpnews.pro/news/claude-opus-5-5-beats-a-human-baseline-and-ai-researchers-move-up-their-self.txt", "jsonld": "https://wpnews.pro/news/claude-opus-5-5-beats-a-human-baseline-and-ai-researchers-move-up-their-self.jsonld"}}