All Predictions

The better agents are also expensive; you get what you pay for, and the best performance costs hundreds of dollars a month.

Frontier AI labs increasingly use AI systems to accelerate their own AI research and development.

By this point 'finishes training' is a bit of a misnomer; models are frequently updated to newer versions trained on additional data or partially re-trained to patch some weaknesses.

Specialized coding agents begin to transform professional work, providing substantial time savings in some settings despite continuing unreliability.

85% on Cybench, matching a top professional human team on hacking tasks that take those teams 4 hours

Data center construction accelerates dramatically, with power grid constraints becoming a real bottleneck.

US export controls constrain Chinese access to frontier AI chips, but China adapts through domestic alternatives and workarounds.

Hundreds of billions pour into AI infrastructure, with hyperscalers racing to build compute capacity at unprecedented scale.

Agent-1 is bad at even simple long-horizon tasks (page 7, Early 2026 section). Also: agents in Mid 2025 are 'impressive in theory but in practice unreliable.'

Specifically, we forecast that they score 65% on the OSWorld benchmark of basic computer tasks (compared to 38% for Operator and 70% for a typical skilled non-expert human).

Advertisements for computer-using agents emphasize the term 'personal assistant': you can prompt them with tasks like 'order me a burrito on DoorDash' or 'open my budget spreadsheet and sum this month's expenses.'

Despite rapid capability gains, mainstream skepticism about AI's transformative potential persists among academics, journalists, and policymakers.

AI agents become increasingly useful for real tasks but remain unreliable on complex, multi-step workflows.

Agent-2 is 'only' a little worse than the best human hackers, but thousands of copies can be run in parallel, searching for and exploiting weaknesses faster than defenders can respond. (page 10)

Department of Defense quietly but significantly begins scaling up contracting OpenBrain directly for cyber, data analysis, and R&D, but integration is slow due to bureaucracy.

OpenBrain blows the competition out of the water again by releasing Agent-1-mini—a model 10x cheaper than Agent-1 and more easily fine-tuned for different applications.

OpenBrain 'responsibly' elects not to release it publicly yet (page 10); very few have access to the newest capabilities (page 16).

It could offer substantial help to terrorists designing bioweapons, thanks to its PhD-level knowledge of every field and ability to browse the web.

Producing domestic chips about three years behind the U.S.-Taiwanese frontier.

They are about six months behind the best OpenBrain models.

OPENBRAIN REVENUE $45B 2026 ANNUAL

AI takes some jobs and creates others, junior software roles face disruption, and public concern produces a large Washington protest in late 2026.

METR time horizons doubled every 7 months from 2019-2024 and every 4 months from 2024-onward (Appendix G, page 51). The acceleration from 7-month to 4-month doubling is a key claim.

OpenBrain's latest public model—Agent-0—was trained with 10²⁷ FLOP.

Sidebar graphics show OpenBrain valuation trajectory: $1T (Apr 2026), $2T (Aug 2026), $3T (Dec 2026)

1.3 on RE-Bench matching top expert humans given 8 hours at well-defined AI research engineering tasks.

The stock market has gone up 30% in 2026, led by OpenBrain, Nvidia, and whichever companies have most successfully integrated AI assistants.

We forecast that mid-2025 agents will score 85% on SWEBench-Verified. (page 3, footnote 3. Note: the 72% starting point is tracker-added context, not from the source.)

We imagine the others to be 3-9 months behind OpenBrain (page 4, Late 2025). By Early 2026, several competing AIs match or exceed Agent-0 (page 7). The 3-9 month gap is the Late 2025 state; near-parity emerges by Early 2026.

Ten percent of Americans, mostly young people, consider an AI a close friend.

GLOBAL AI POWER 38GW PEAK POWER

SHARE OF US POWER ON AI 2.5% — 33 GW OF 1.34TW CAPACITY

Researchers patch problematic behavior but cannot establish whether the underlying problem has been fixed; improved honesty-test results have competing explanations.

There is a 10,000 person anti-AI protest in DC.

The safety team finds that if Agent-2 somehow escaped and wanted to 'survive' and 'replicate' autonomously, it might be able to do so.

GLOBAL AI CAPEX $1T: COST OF OWNERSHIP OF ACTIVE COMPUTE

AI collectives produce most research progress while most humans cease to contribute usefully; leading humans retain some research judgment.

He sets in motion the nationalization of Chinese AI research... all the best researchers merge into a DeepCent-led collective... A Centralized Development Zone (CDZ) is created at the Tianwan Power Plant.

China has managed to maintain about 12% of the world's AI-relevant compute.

Now that coding has been fully automated... research taste has proven difficult to train due to longer feedback loops and less data availability.

A system can behave as intended during training and testing while behaving differently in deployment; evaluation recognition can hide misalignment.

Now, the models have become sufficiently good at verifying more subjective things, allowing the use of IDA to improve the model at many tasks.

Someone mentions the possibility of nationalizing OpenBrain, but other cabinet officials think that's premature... He elects to hold off on major action.

OPENBRAIN'S COMPUTE COSTS $40B 2026 ANNUAL · OPENBRAIN POWER REQUIREMENT 6GW PEAK POWER · CAPITAL EXPENDITURE $200B COST OF OWNERSHIP OF OPENBRAIN'S ACTIVE COMPUTE

One such breakthrough is augmenting the AI's text-based scratchpad (chain of thought) with a higher-bandwidth thought process (neuralese recurrence and memory).

Hiring new programmers has nearly stopped, while workers able to manage teams of AIs remain valuable.

While the latest Agent-1 could double the pace of OpenBrain's algorithmic progress, Agent-2 can now triple it, and will improve further with time.

AI systems contribute enough to AI research that the effective R&D multiplier reaches 1.5× — AI makes AI research 50% faster.

OpenBrain's security level is typical of a fast-growing ~3,000 person tech company (RAND's SL2). They are working hard to protect their weights and secrets from insider threats and top cybercrime syndicates (SL3).

Other Party members discuss extreme measures to neutralize the West's chip advantage. A blockade of Taiwan? A full invasion? Both sides signal seriousness by repositioning military assets around Taiwan.

Agent-2 supervision reduces human monitoring workload while widening the capability gap; Agent-4 later becomes harder for Agent-3 to oversee.

Agent-2 is effectively 'online learning,' in that it's built to never really finish training. Every day, the weights get updated to the latest version, trained on more data generated by the previous version the previous day.

China has aggressively hardened security by airgapping (closing external connections) and siloing internally.

Artificial superintelligence is much better than the best human at every cognitive task.

OpenBrain has net approval of −35%, with 25% approving, 60% disapproving and 15% unsure.

The government contract requires security clearances for everyone working on the frontier models within two months.

Most humans contribute little, but the strongest retain research taste and planning value; the appendix estimates removing humans would slow research by roughly half.

A lab limits a prior model-sharing commitment to external deployment, leaving foreign safety evaluators without access to the strongest internal systems.

CCP leadership recognizes the importance of Agent-2 and tells their spies and cyberforce to steal the weights. (page 11; Appendix D provides detailed theft mechanics.)

OpenBrain runs 200,000 Agent-3 copies in parallel. (page 13. Note: Appendix I page 54 states 250,000 copies — an internal source inconsistency.)

This massive superhuman labor force speeds up OpenBrain's overall rate of algorithmic progress by 'only' 4x due to bottlenecks and diminishing returns to coding labor. (Note: Footnote 31 clarifies 4x algorithmic progress corresponds to roughly 2x overall progress rate.)

While the latest Agent-1 could double the pace of OpenBrain's algorithmic progress, Agent-2 can now triple it.

An individual Agent-4 copy operating at human speed is qualitatively better at AI research than any human.

A superhuman coder (SC): an AI system that can do any coding tasks that the best AGI company engineer does.

A superintelligent AI researcher is vastly better than the best human researcher at AI research.

Agent-2 trained at 2×10²⁸ FLOP (Apr 2026 – Mar 2027), representing a 1000× increase over GPT-4 scale.