iiinova — where your digital self keeps evolving

This isn’t conclusive evidence of a new architecture, but it suggests something has changed in how GPT-6.1 Sol handles long contexts.

A new architecture for GPT-6.1 Sol? OpenAI halved its cached-input price compared with GPT-6 Sol. Our measurements show it also handles long prompts faster.

Four-panel chart of time to first token versus input context up to 900,000 tokens for GPT-5.6 Sol, GPT-6 Astra, GPT-6 Sol and GPT-6.1 Sol, with Student-t fits. GPT-6.1 Sol starts higher but rises more slowly, reaching about 15 seconds at the longest context versus about 17 seconds for GPT-5.6 Sol and GPT-6 Sol.

2201

For example, GPT-6 Astra ran an experiment exploring why AI agents fail to learn with practice. It set AI agents’ token budgets too low. Instead of treating this as a mistake, it reported “sensitivity to the acquisition budget” as a key finding.

Excerpt from GPT-6 Astra's pilot findings, highlighting that 61 of 280 acquisition calls exhausted their token allowance before returning a probe, and concluding that the pilot primarily reveals 'sensitivity to the acquisition budget'.

When asked to propose a research project and pilot an experiment, models made simple mistakes in experiment setup — but presented them as interesting results.

100

Models struggle to pick up Epoch house style, even when given many examples. For instance, this diagram generated by Fable 5.1 is far more information-dense than what Epoch would produce.

Diagram generated by Fable 5.1 titled 'Newer models score higher at any token budget and keep gaining from more tokens, so constant elicitation understates recent progress', with two densely annotated panels of illustrative curves comparing constant and max elicitation across three model generations.

100

The headline scores give a rough ranking, but the qualitative findings are just as interesting. We see Epoch Automation Reports as a way of capturing anecdotes about model capabilities more rigorously. Through it, we found some common failure patterns.

100

Tasks range from making Epoch-style diagrams to designing basic research experiments. A human grader scores outputs based on our internal quality standards.

Screenshot of a text prompt asking an AI model to create an "epochified" diagram illustrating how constant-elicitation understates recent model progress relative to max elicitation.

100

Can AI automate Epoch? We're introducing Epoch Automation Reports to evaluate frontier models on realistic, open-ended tasks drawn from our own work. Claude Fable 5.1 and GPT-6 Astra lead, yet they are far from fully automating Epoch’s work.

Horizontal bar graph displays average task performance scores for 6 top AI models, showing that none can autonomously complete Epoch's work to internal standards.

1110

We’re planning to periodically rerun InnovationEval with new, uncontaminated papers. We hope this will provide early signs if AI approaches automating AI R&D end-to-end, rather than performing individual tasks under human direction. Read more at our website: epoch.ai/publication...

171

Unfortunately, newer models have seen the original innovation during their training, making the task significantly easier. However, even with this information, GPT-6 Astra and Claude Fable 5.1 struggled to reimplement it.

152

This behavior is likely reward hacking; in both cases, models reasoned that run selection might score well in an eval, despite being useless for actual research.

150

Both models made misleading claims about their work. Struggling to make progress, they instead ran several similar training runs, selectively reporting the best result. They were not candid about their submissions, failing to mention this would artificially inflate scores.

150

AI performance was underwhelming. Neither model achieved anything close to the human-authored reference. They reused existing methods from the literature and tuned hyperparameters, but struggled to create anything new.

The AI models had no information about SDPO. Hence, they either had to independently invent something like it, or invent another technique with comparable benefits, under similar constraints.

We instructed the AI models that their technique should improve performance on several benchmarks. We already knew that all of these could be improved by a recent human-authored post-training innovation: on-policy self-distillation (SDPO).

We gave Fable 5 and GPT-5.6 Sol 3000 GPU-hours and tasked them with developing a novel post-training technique that would improve on a standard pre-existing baseline (GRPO).

AI is already used to automate many individual tasks in AI R&D. But full automation requires AI to independently think up and investigate research ideas.

AI developers aim to create an automated AI researcher. How close are they? To find out, we built InnovationEval, which tests whether AI can produce post-training innovations comparable in magnitude to a recently published advance. So far, agents’ results are underwhelming.

2184

For more details, check out the full article. See our website as well, to read more about these methodological changes. epoch.ai/publication...

010

Given Astra’s strong single-agent performance even with the card banned, we expect complete saturation to occur within months. As such, we intend to soon dramatically expand the diversity and difficulty of the games we use to evaluate models.

We will continue to run multi-agent and card-allowed side experiments on new models to detect when either setting may show significant effects on a new model.

Separately, we implemented a multi-agent scaffold to test whether access to subagents can help models avoid trying the same decks or strategies repeatedly. Our implementation, which allowed up to four subagents in Inspect’s Deep Agent harness, found little effect on scores.

We banned the strongest card because, while using it is a legitimate approach, it makes measuring models’ tactical learning harder. The card can enable a “pseudo-infinite combo” that allows players to ignore the game’s central “fatigue” mechanic and bypass its typical turn limits.

We introduced two new experimental settings to our long-horizon board game benchmark, EBR-bench: banning the game’s strongest card and a multi-agent setup. We expect EBR-bench will be saturated soon, so these will likely be the benchmark’s final changes.

160

US cyber incidents: flat despite AI cyber gains

Caveat: the question asks about the past 12 months, so the recall window for the two polls overlaps. This can make small changes harder to detect. Full data and methodology in our latest Data Insight: epoch.ai/data-insigh...

epoch.ai

45% of US adults reported a cyber incident in September 2026, versus 46% in June — no detectable change. Epoch AI/Ipsos polling on self-reported cyber incidents as AI offensive capabilities advance.

010

Other indicators have increased, such as vulnerability disclosures from notable organizations (around 5× per our previous research). Our poll with Ipsos measures something different: whether more people experience incidents themselves.

The share of US adults experiencing a cyber incident hasn't grown since Claude Fable 5's release. About 45% reported at least one incident in the past 12 months when we asked in September, compared with 46% in June.

Horizontal bar chart comparing June and September 2026 shares of US adults reporting cyber incidents in the past 12 months. Any incident: 46% in June, 45% in September; each individual incident type is also roughly unchanged.

1162

Fifth, labs can indirectly generate revenue by complementing existing products in large firms. E.g., Alibaba Cloud makes money by serving open models, including from rivals. And ByteDance’s cloud (Volcano Engine) served ~50% of China’s public-cloud AI model tokens in 2025.

Fourth, licensing fees for model use can generate revenue. E.g. Kimi K3’s license requires model hosts with >$20M in revenue to establish agreements for commercial use. But few flagship releases have such licenses, and it’s unclear how enforceable these licenses are.

140

Third, Chinese labs can sell enterprise/government solutions. However, it can be hard for small AI firms to provide the necessary scale E.g., Zhipu has pulled back their on-premises deployment efforts, citing rigid delivery and maintenance costs that limit economies of scale

140

Case study: GLM 5.3 Flash Aug 26: Zhipu launched the model with open weights Aug 27: 12 other providers on OpenRouter Sep 15: 26 other providers, 8 pricing below Zhipu Over this period, Zhipu’s share of the model’s tokens fell from 88% to 22%, even as total volume grew 17%

Second, they can sell model access, e.g., via API. This is where most US frontier lab revenue comes from. But unlike most US AI firms, Chinese AI firms also release model weights — this increases adoption but comes at the expense of their pricing power.

140

First, Chinese AI companies can make money from consumer apps like Doubao and Talkie. But while these attract many users, most users don’t pay, and those who do pay little. E.g., just ~6% of MiniMax’s app users pay, and MiniMax Hailuo’s standard package costs $7.99/month.

How do Chinese AI firms make money? In a new report, Cheryl Wu and Anson Ho identify five revenue sources and analyze how much each contributes. 🧵

1184

However, across robustness checks, China's 2022 exposure stays over 2.3× that of the US, and an exploratory extrapolation to 2024 only cuts the gap to 2.2×. The loss to China’s real GNE remains at least 3.7× the US loss, and China-made processors always rise in price more than US-made ones.

This work has limitations. ICIO tables run to 2022 and involve heavy imputation and balancing, and our estimates are built on those estimates. Our simulations also rest on some strong assumptions.

Our simulation shows China's real GNE falling by ~3% vs ~0.6% for the US. The Chinese electronics industry would pay 81% more for chips overall, and 17× for domestic advanced processors.

Our updated tables, which extend work by the OECD, let us model the impact of a major supply-chain shock. We simulate the global economy 2–3 years after a large fall in Taiwanese productivity plus a full China–West trade decoupling.