Free reference · updated …
What each lab publishes on its own documentation for a model’s context window and maximum output tokens. One page read per model. Nothing here is inferred, averaged, or reconciled between labs, and where a lab publishes no figure the cell says so rather than guessing. Every row links to the page it was read from.
| Model | Lab | Context window | Max output | As published | Source |
|---|---|---|---|---|---|
| Loading… | |||||
Machine readable: anchors.json. When each platform stops serving a model is on the lifecycle page.
Things about these sources that will bite you
One lab writes K and M and means decimal, so 200K is 200,000 and 1M is 1,000,000. Another publishes exact integers and its 1M context window is 1,048,576. The two are 4.9% apart while printing the identical string. A pipeline that resolves “1M” to one integer is wrong for one of them, whichever one it picks.
One published context window is 1,047,576, which is 220 minus 1,000. It looks exactly like a transposition and it is correct as printed. Rounding it to 1,048,576 replaces a right answer with a wrong one that then agrees with every neighbouring record, which is the hardest state to detect afterwards.
A context window, a maximum input, and a maximum output. On most of its models that publish all three, the context window is exactly the maximum input plus the maximum output. A schema with only two fields has nowhere to put the third, and a maximum input landing in the context field is a real error rather than a hypothetical one.
Two labs in this table give a context window only as 32k, 128k, 256k or 1M, and print an exact integer on no page at all. One of them answers the question in its API reference by pointing back at the page that rounds it. So 1M is either 1,000,000 or 1,048,576, the two are 4.9% apart, and nothing published decides which. A figure from those labs can rule a value out. It cannot confirm one, and this table does not pretend otherwise: those cells show the string as published rather than a number I chose.
If a lab publishes a 256k context window and a host serves 128k, the host has not made a mistake: it decided to serve a shorter window, which it is free to do. Only a figure above what the lab publishes is impossible, because nobody can serve more context than the weights carry. Applying that rule cut one of my own candidate lists from five to three, which is the usual direction of travel once a rule is written down.
One lab publishes no maximum output for any model at all, on its index or on any model page. Every empty cell for that lab in this table is empty by fact and not by a failed parse. Any catalogue carrying a maximum output for those models is carrying a number with no first-party source behind it.
A model published as weights declares a maximum position count in its configuration and nothing else. There is no published output limit, because the output cap is a serving choice made by whoever runs it. That is why this table covers models with a lab behind them and does not pretend to cover the rest.