Epoch AI News

Anthropic

OpenAI

Google

Other

Meta

DeepSeek

xAI

The METR time horizon will exceed 10 hours.

xAI IPO

FrontierMath Tier 1-3 >= 80%

FrontierMath Tier 4 >= 80%

An open millennium prize problem is solved, involving some AI assistance

OpenAI will introduce ads to ChatGPT in some form

ARC-AGI 3 Semi-private >= 50%

SSI will raise >= $1B in a funding round

OpenAI releases GPT-6

Anthropic releases Claude 5

Schmidhuber will complain about people not citing his work properly

OpenAI will announce some kind of hardware product

Zvi will write a blog post mentioning SSI

SSI will have an update listed at https://ssi.inc/updates

Thinking Machines will train and release their own model

Yudkowsky will publish a post on Lesswrong

A Solid Result in FrontierMath: Open Problems is solved by AI

I will think that there has been significant progress towards models which use neuralese/recurrence

Sam Altman will continuously be CEO of OpenAI until the end of the year.

An AGI lab will be valued at >= $1T

Ilya Sutskever will continuously be CEO of SSI until the end of the year

I will think that computer use has significantly improved since 2025

Epoch Capabilities Index >= 170

Dario Amodei will continuously be CEO of Anthropic until the end of the year

Microsoft+Google+Amazon+Meta capex will increase by >=30% vs 2025

My median ASI timelines will shorten

Anthropic IPO

Jensen Huang continuously CEO of Nvidia through EOY 2026

Grok 5 will be released

The METR time horizon will exceed 20 hours

There will be federal AI Safety legislation which I think is net positive

An LLM will beat me at chess

Anthropic will release a model classified as ASL-4

My P(doom) at EOY (resolves to %)

An open source model will be released with 5T+ total parameters

Ilya Sutskever will be on a podcast for more than 30 mins

There will be an international treaty/agreement centered on AI

There will be clear evidence of egregious scheming in the wild

SSI will release a product

Nvidia will outperform the S&P

An LLM will beat me at Shogi

I will think that there have been significant advances in continual learning

The METR time horizon will exceed 40 hours

Epoch AI will estimate that there has been a training run using more than 5e27 FLOP

At least 3 people will do an anti-AI hunger strike

I will think that METR time horizon continues to be an important benchmark

Anthropic releases Claude 6

I will meet someone who has an AI companion

I will ride on a tesla robotaxi

The US will allow selling some sort of Blackwell chip to china

Any of [OpenAI, Microsoft, Google, Anthropic, xAI, Meta] claim to have reached AGI

Largest distributed training run exceeds 1e25 FLOP

Google will outperform the S&P

Yann Lecun’s AMI lab will release an open weights model of some kind

Any of [Coreweave, Nebius, Lambda] will be acquired

Any of [Coreweave, Nebius, Lambda] declare bankruptcy

Epoch Capabilities Index >= 185

Chatgpt will write an explicit sex story without jailbreaks

Anthropic will release a model classified as ASL-5

There will be an AI capabilities pause lasting at least a month involving frontier companies

Metaculus will predict AGI before 2030

AAA with LLM powered NPCs releases on Steam

SSI will raise >=$10B in a funding round

Thinking machines will post at least 10 blog posts at https://thinkingmachines.ai/blog/

There will be an AI protest involving more than 100k people

There will be a credible leak about SSI strategy

I will watch a fully AI-generated film lasting at least an hour

I will see a humanoid robot walking around in a non-demo setting

An open source model will top the Chatbot Arena in the 'text' category

I will think that a model released by Meta is the best coding model for a period of at least a week.

The bubble collapses in devastating fashion

Anthropic will release an image/video model

SSI will be valued at >= $100B

An AGI lab will be fined >= $10B

OpenAI will sell their own chips to external customers

Epoch Capabilities Index >= 200

Anthropic will introduce ads to Claude in some form

Largest distributed training run exceeds 1e27 FLOP according to Epoch AI

US unemployment rate reaches 10% due in part to AI

Grok 6 will be released

China will invade Taiwan according to metaculus

Any of [OpenAI, xAI, Google, Anthropic, Meta] releases an LLM (general-purpose or code-only) pretrained largely on licensed data

S&P 500 will fall by more than 50%

I will think that a Chinese model is the best coding model for a period of at least a week.

An AGI lab will be valued at >= $5T

Any of [OpenAI, xAI, Google, Anthropic, Meta] will offer a subscription plan costing >= $1,000/month.

Any of [OpenAI, xAI, Google, Anthropic, Meta] will offer a subscription plan costing >= $10,000/month.

I will think that SSI has the best coding model in the world for a period of at least a week

S&P 500 will rise by more than 50%

OpenAI IPO

Largest distributed training run exceeds 1e26 FLOP

The weights of a closed source model from [OA, XAI, GDM, Anthropic] will be stolen/leaked

Yudkowsky will publish a new book

SSI IPO

SSI will be valued at >= $1 T

NO

November 2026

December 2026

October 2026

January 2027

February 2027

March 2027

Other

NO

Anthropic

OpenAI

Google DeepMind

46% or higher

Below 38%

38% to below 42%

FutureLivingLab, iFlow ROME, Safety-Aligned Data Composition https://huggingface.co/papers/2512.24873

MIRI, "How could an LLM destroy humanity?" short https://youtube.com/shorts/Z6oPeFvp6_s

MIRI, An International Agreement to Prevent the Premature Creation of Artificial Superintelligence, Notes on Article VI https://arxiv.org/abs/2511.10783

Epoch AI, How many AI agents could run on the AI chips shipped through 2027? https://epoch.ai/publications/estimating-the-agent-population

Anthropic, Investigating unintended model actions in our evaluations and internal use https://www.anthropic.com/news/investigating-unintended-model-actions

Kirgis & Hua, The Continual Constitutional Convention: How language models revise the documents that govern them https://www.alignmentforum.org/posts/c9woewdY9ddsPqf3J/the-continual-constitutional-convention-how-language-models

5% to under 7%

Below 5% (no model beats today's 3 of 68)

7% to under 10%

Below 35% (GPT-6 Astra's 34.7% still on top)

35% to under 45%

45% to under 60%

85% to under 90%

80% to under 85%

90% or higher

Below 78%

78% to below 82%

82% to below 88%