How current is each model, and what recency costs
A model's knowledge cutoff is the last date its training data covers — ask about anything after it and you're relying on your own context, not the model's memory. Across 154 models here, the median cutoff is September 2025 and the most recent maker-published cutoff reaches Jun 2026. Cutoff is a property of the model, identical whichever provider serves it, so the price shown is each model's cheapest host.
One caveat before you read the table. Only 55 of these models have a cutoff their maker actually published; for the other 99 — nearly every open-weight family — no maker figure exists at all, and the date below is a third-party estimate. Those are marked≈ and they are not a fact in the sense the rest of this index means it.
Freshest priced model: claude-haiku-5-5 is trained through Jun 2026 — $0.6/1M blended (Anthropic).
Recency and quality don't move together — a newer cutoff isn't a better model. Cross-check the value ranking and license before choosing.
How fresh is the field?
Share of models whose training cutoff is at least this recent.
Most recent knowledge
Ranked by training cutoff; where two match, the cheaper host is listed first. 10 of these 20 rows are third-party estimates (≈), so read the order as indicative rather than exact.
| Model | Cutoff | Cheapest host | Blended /1M | Source |
|---|---|---|---|---|
| claude-haiku-5-5 | Jun 2026 | Anthropic | $0.6 | maker → |
| claude-sonnet-5-5 | Jun 2026 | Anthropic | $12 | maker → |
| claude-opus-5-5 | Jun 2026 | Anthropic | $24 | maker → |
| claude-fable-5-1 | June 2026 | Anthropic | $60 | maker → |
| grok-4.7 | May 2026 | xAI | $8 | maker → |
| claude-opus-5 | May 2026 | Anthropic | $30 | maker → |
| gpt-6-luna | April 2026 | OpenAI | $0.6 | maker → |
| kimi-k2.6 | ≈ April 2026 | Chutes | $2.4 | estimate → |
| kimi-k2.5 | ≈ April 2026 | DeepInfra | $2.7 | estimate → |
| kimi-k2-thinking | ≈ April 2026 | $3.1 | estimate → | |
| kimi-k2.7-code | ≈ April 2026 | StreamLake | $3.712 | estimate → |
| kimi-k3 | ≈ April 2026 | Chutes | $11.35 | estimate → |
| gpt-6-sol | April 2026 | OpenAI | $12 | maker → |
| gpt-6.1-sol | April 2026 | OpenAI | $12 | maker → |
| gpt-6-astra | April 2026 | OpenAI | $60 | maker → |
| deepseek-v4-flash | ≈ March 2026 | Relace | $0.21 | estimate → |
| deepseek-v4-flash-0423 | ≈ March 2026 | StreamLake | $0.21 | estimate → |
| deepseek-v3.2-exp | ≈ March 2026 | AtlasCloud | $0.336 | estimate → |
| deepseek-v4.1-flash | ≈ March 2026 | Io Net | $0.468 | estimate → |
| deepseek-v3.2 | ≈ March 2026 | GMICloud | $0.518 | estimate → |
Cutoff = last month of training data. Rows linking to maker → are taken from the maker's own documentation, at the family's reference version where a per-version date isn't published. Rows marked ≈ linking to estimate → come from a third party because the maker publishes no cutoff for that family. Prices per 1M tokens (USD), cheapest published rate. Verified 2026-10-23.
FAQ
What is a model's knowledge cutoff?
It's the last date a model's training data covers — ask about anything after it and the model is relying on your context, not its own memory. Across 154 models here, the median cutoff is September 2025 and the most recent maker-published cutoff reaches Jun 2026.
Why are some of these dates marked as estimates?
Because the maker never published one. 55 of the 154 models here have a cutoff stated in their own documentation; for the other 99 — largely the open-weight families — no maker figure exists anywhere, and the date shown is a third-party estimate marked with ≈. We keep it because a rough date beats no date, but it is not the same kind of fact as the rest of this index and we don't pretend otherwise.
Which model has the most recent knowledge, and what does it cost?
claude-haiku-5-5 is the freshest priced model with a maker-published cutoff, trained through Jun 2026, at $0.6 per 1M tokens blended via Anthropic.
Does a newer knowledge cutoff mean a better model?
No — recency and quality don't move together. A more recent cutoff only means fresher training data, not stronger reasoning. Cross-check the value ranking before choosing on cutoff alone.