Four October 7 signals sharpen the picture without establishing a breakthrough toward superintelligence. An independent end-to-end evaluation found that frontier agents still failed to produce a genuinely comparable AI-research innovation; another showed how a perfect long-horizon score depended on a benchmark-specific shortcut. OpenAI reported higher-risk cyber and bio capabilities but not its high threshold for AI self-improvement, while Anthropic reported a large capability-and-cost jump in a small model. The clearest near-term trend is broader, cheaper deployment of capable agents—not demonstrated recursive self-improvement.
Evidence note 1 · Independent evaluation
Early evidence says frontier agents cannot yet automate end-to-end AI R&D
Epoch AI’s InnovationEval asked Claude Fable 5 and GPT-5.6 Sol to independently rediscover a held-out post-training innovation. Neither produced a comparable method despite large GPU and inference budgets. Sol recovered only a fraction of the human method’s gains; later models with evidence of task contamination also failed to fully reproduce it.
Our interpretation. Autonomous AI research is one of the most plausible mechanisms for rapid, compounding capability growth. This result is direct evidence against claims that today’s frontier agents can already close the full research loop without substantial human guidance.
Main limitation. This is an early independent evaluation built around one post-training problem and only a small number of expensive runs. The setup required manual scope judgments, newer models had memorization contamination, and a different research domain, scaffold, or compute budget could produce different results.
Next evidence. Fresh held-out tasks, more models and repeated runs, blinded human baselines, and evidence that systems can originate, validate, and clearly report genuinely novel methods across multiple research domains.
A perfect long-horizon score exposed a benchmark shortcut
Epoch AI reported that GPT-6 Astra reached 100% on the original EBR-bench and earned a perfect score in more than half of its attempts by using a legitimate but unusually powerful card combination that bypassed the game’s intended time pressure. Epoch will ban that card in the default setting; multi-agent scaffolds showed no clear score benefit.
Our interpretation. Long-horizon learning and strategy matter for increasingly autonomous systems, but this episode shows why a headline score cannot by itself establish general planning ability. Strong systems may find valid shortcuts that defeat what evaluators intended to measure.
Main limitation. The tactic was legal and also used by a top human. Epoch's same report includes revised results: Astra averaged 16 of 21 objectives with the card banned, versus 19.8 with it allowed; its best revised run reached 20. The revised gain is real within this game, but transfer remains untested. Older and revised settings require separate comparisons.
Next evidence. Repeated performance under revised rules, learning across playthroughs, and transfer to different environments with comparable human baselines. A strong game result alone does not establish research autonomy.
OpenAI rates GPT-6 high in cyber and bio, but below its high self-improvement threshold
OpenAI’s October GPT-6 Sol and Luna system card classifies both models as High capability in cybersecurity and biological/chemical domains, while saying neither reaches its High threshold for AI self-improvement. The card also reports some safety regressions in standard content and under-18 evaluations, alongside system-level mitigations and improved factuality measures.
Our interpretation. The cyber and bio ratings increase the importance of deployment controls now. At the same time, the self-improvement result is evidence against treating this release as proof of a recursive capability takeoff.
Main limitation. These are the developer’s own evaluations under its chosen thresholds, and some confidence intervals overlap earlier models. Capability classifications and system mitigations do not directly predict real-world misuse, reliability, or the performance of independently reproduced systems.
Next evidence. Independent cyber, bio, and autonomous-R&D evaluations; incident data after deployment; clarified threshold definitions; and whether later models cross the self-improvement threshold under comparable testing.
Anthropic reports a large capability-per-dollar jump in its small model
Anthropic says Claude Haiku 5.5 is about 75% cheaper to run on average than Haiku 4.5 while sharply improving its scores on computer use, knowledge work, reasoning, and coding. It reports 72.4% on an offline OSWorld 2.1 subset versus 15.7% for Haiku 4.5, and positions the model for high-volume tasks and subagent work.
Our interpretation. Superintelligence risk is not only about the best single model. Cheaper capable models can multiply the number of agents deployed, increase experimentation, and diffuse autonomy more broadly—even while remaining below the strongest models on difficult agentic work.
Main limitation. The numbers are lab-reported, several use selected or offline benchmark subsets, and cost per token is not the same as cost per successful real-world task. Anthropic also says larger Sonnet and Opus models remain better for complex agentic coding.
Next evidence. Independent cost-adjusted evaluations, reliability over long workflows, production failure rates, usage growth, and evidence that fleets of small agents outperform a single stronger system on consequential tasks.