GPT-5.5 is 2x more expensive than GPT-5.4, meaning roughly that active params should be ~2x higher

now GPT-5.5 has 2.14x longer time horizons than GPT-5.4, which would require a ~2.4× increase in active parameters

New paper!

Think Fast: Estimating No-CoT Task-Completion Time Horizons of Frontier AI Models

@METR_Evals showed that models' time horizons have doubled every few months. We ask: what length of tasks can models complete without any CoT?

Authors (continued): @joneedssleep @paanarle @TwmStone Ram Potham Ionut Gabriel Stan @harrymayne5 Simeon Hellsten, Shubhorup Biswas @arianaazarbal @wlanderson0 @latentfool @RyanPGreenblatt Julian Stastny

Paper: https://arxiv.org/abs/2606.07157 Our code and benchmarks are available upon request

This is an important thing to measure: if models can do extensive reasoning without any CoT, monitors would struggle to understand models’ motivations and catch dangerous planning. We suggest that AI companies start to track no-CoT THs explicitly.

We find that frontier models like GPT-5.5 answer questions that take humans roughly 3 minutes with 50% reliability, and that the no-CoT time horizon of frontier models has doubled every 373 days.

We also measure a reasoning token horizon by replacing human time with how many tokens it takes o3-mini to solve a specific task. We find that GPT-5.5 solves questions that require about 1,500 tokens for o3-mini, and that this token horizon has doubled every 437 days.

Authors: @dswg97 @F_Rhys_Ward @AndersCWoodruff @RaunoArike Josh Hills, Alex Serrano, @ida_icy, Jason Ross Brown

@redwood_ai @ConstellOrg @MATSprogram @AetherAIS LW post: https://www.lesswrong.com/posts/SieLowPgNgRSPGhFw/estimating-no-cot-task-completion-time-horizons-of-frontier

Our trends are robust to changing the distribution of benchmarks, adding longer-generation and multi-turn agentic questions, restricting to questions with measured human times only, dropping any single model, and breaking down by domain.

Across 43 benchmarks, we fit a logistic curve on success rate against log human time to estimate models 50% no-CoT time horizons. We use a combination of measured human times and estimated ones.

Thanks to the following orgs for supporting this word: @redwood_ai @ConstellOrg @MATSprogram @AetherAIS

@scaling01 straight lines on a log graph is always a banger, unless you know any better. remember this?

@scaling01

@scaling01 You’ve shown us a huge number of graphs and extrapolations, so apparently we already know everything about the future. Now combine them and provide a specific, testable prediction for the model’s performance and capabilities as of June 2027. Otherwise, it’s just bullshit

@scaling01 Amazing chart!

Here´s the thing: RSI makes releasing models a lose for the big firms. Google knows this. All you do is allow lesser code bases to reach parity through agentic looping/improvement. The game just switched to applied knowledge (e.g. drug discovery) Better models simply enable more RSI.

@scaling01 the gap between the green and purple trend lines is the whole story. CoT is buying 2x the doubling rate but you're paying for it in latency and tokens every single call

@dswg97 @METR_Evals This is a useful eval axis because it separates visible reasoning quality from latent task capacity. Monitoring gets harder when a model can complete longer tasks without exposing much intermediate state.