Exploring Claude/GPT Knowledge Cutoffs and Pre-Training Timelines

Who Should Read

An LLM API developer who wants to verify a model's knowledge cutoff before switching model versions

Core Mechanics

• Date quizzes measure the cutoff
  • The author built 8-way multiple choice quizzes from daily Wikipedia events and ran every model on them.
  • The error-rate curve shows where the model loses signal from its training data.
  • The curve is smooth, so the midpoint is used as the cutoff.
• Opus 4.7 onward looks like one training run
  • This is the author's speculation: their effective cutoffs all sit near late December 2025.
  • It assumes pre-training completion correlates with dataset timespan, so it could be off by a few months.
• GPT-5.6 is a separate checkpoint from GPT-5.5
  • The author speculates the GPT-5.6 family comes from its own checkpoint finishing around late February 2026.
  • Luna seems to predict the future only because of high error rates at low reasoning effort.
• Opus 5 doesn't match its published cutoff
  • The published cutoff is May 2026, yet it knows no more than the January 2026 cutoff models.
  • Ablations on the question types didn't change this, and coding package version recall showed the same cutoff.

Community Reactions

• Opinion that this analysis could show whether labs delay releasing finished models
  • The commenter hoped it would also help gauge how far behind open-weights models are.
• Opinion skeptical that Anthropic never distilled from ChatGPT
  • The commenter says it was plausible in the GPT-4 era when Anthropic was the underdog.
• Reminder that safety and alignment post-training cycles run after the pre-training cutoff
  • The commenter says this should be factored into the interpretation.
• Experience that models seem to have per-domain cutoffs
  • A commenter found classic literature static, tech knowledge updated often, and tabloid topics always current.
  • They said software like Git looks stable because basic commands cover most jobs.
• Opinion that a name like 'Opus 5' covers many models and versions updated over time

How to Apply

• Measure your model's cutoff with your own quiz
  • Write 10-20 eight-option questions on events from the last six months and run them on your model.
  • Record where accuracy drops after the published cutoff.
• Regression-test recency-dependent prompts on model swaps
  • Build a question set around library versions and API changes.
  • Compare accuracy against the old model before deploying the new one.
• Inject the current date and documents for freshness-sensitive features
  • Put today's date in the system prompt.
  • Supply recent facts through RAG or search tools instead of relying on model memory.
• Document your measured cutoff instead of trusting the published one
  • As the Opus 5 case shows, published and actual knowledge can differ, so record measured values in your team wiki.

Try It

# Paste into Claude or ChatGPT as is (a simplified version of the post's method: build a quiz + ask the date)

You are a quiz writer for measuring knowledge cutoffs. In the {{domain}} domain, pick one real event per month from {{start month}} to {{end month}}. For each event, write an 8-option multiple choice question with exactly one correct answer and seven plausible wrong ones.

Output only the questions, and put the answer key separately at the very end.

Next steps: run these questions on the model you want to measure, tabulate accuracy by month, and find where it drops. Also ask the same model 'What is today's date (year and month)?' and compare.

Terminology

Related Papers

• The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge

  Warning an LLM judge about bias or letting it self-critique improved 20 of 21 model–bias combinations.

• 90% of US working hours are spent on tasks not covered well by AI benchmarks

  Mapping 29 benchmarks to job tasks shows 90% of US working hours lack good coverage

• Postgres 19 delayed after AI reviewers find more bugs than humans can fix

  LLM code review surfaced so many bugs that Postgres 19 slipped to November and 8-10 features were reverted

• Show HN: Gemini 4 Argon plays StarCraft

  LLMs wrote StarCraft bots in one hour; GPT-6 Astra scored highest.

• A Plausible Explanation Is Enough to Mislead Multimodal Judges

  Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.

• Neural networks are clean

  Looking at neurons directly shows most have one main job, challenging superposition.

Related Resources

• https://blog.sshh.io/p/exploring-claudegpt-knowledge-cutoffs