Both prices expire, in different ways
| Gemini 4 Argon | GPT-6 Astra |
|---|
| Input | $2 to 30 Nov, then $4 | $10 to 272K input, then $20 |
| Output | $10 to 30 Nov, then $20 | $50 to 272K input, then $75 |
| Cached input | 95% off, introductory period only | $1, or $2 above the threshold |
| Token budget | 1M combined input + output | 1.05M context, 922K max input, 128K output |
| Model ID | gemini-4-argon | - |
Argon's cliff is a calendar
The $2/$10 everyone is quoting is introductory and ends on 30 November 2026. On 1 December it becomes $4/$20 - a straight doubling, applied to everyone at once, with no action required on your part to trigger it.
The 95% cached-input discount is tied to the same period. If a large part of your spend is repeated context, that discount ending matters more than the headline rate doubling.
Practical consequence: any cost model built on $2/$10 has about six weeks of validity. If you are choosing a provider this month on price, run the numbers at $4/$20 as well, because that is what you will actually pay from December.
Astra's cliff is a token count, and it is harsher
Astra is $10/$50 while input stays at or under 272,000 tokens. Past that it is $20/$75 - and the entire request rebills at the higher tier, not just the excess. A request at 271K costs one rate. A request at 273K costs double, retroactively, on every token in it.
That is a worse failure mode than Argon's. A calendar change is predictable and you can plan for it. A threshold that silently doubles a bill mid-request, based on input length that varies per call, is the kind of thing discovered on an invoice.
The token budget difference nobody explains
Argon runs on a single 1M-token budget covering input and output together. A long prompt reduces the room left for the answer. This is not a 1M context window with a separate output cap - it is one pool.
Astra separates them: 1.05M total context, 922K maximum input, 128K maximum output.
For most work this never matters. Where it does: anything generating long output from long input - full-document translation, large refactors, exhaustive extraction. On Astra you can hand it 900K and still get 128K back. On Argon, 900K of input leaves 100K for the response, and you find out by being truncated.
One unresolved point worth knowing: Google's 30 September announcement described Argon as having a 1M output cap, up from 64,000 in earlier Gemini models. The later description is the combined budget above. Those are different claims and nobody has reconciled them publicly.
Cost on a real workload
Take 1M input and 100K output tokens.
- Argon today: $2 + $1 = $3. Except 1M input plus 100K output exceeds the combined budget, so this request does not fit. Realistic ceiling is roughly 900K in and 100K out.
- Argon from 1 December: double it.
- Astra: 1M input is over 272K, so the whole request bills at the high tier - $20 + $7.50 = $27.50.
Nine times the price, and that gap is entirely the threshold rule rather than the headline difference. At short lengths the ratio is 5x; at long lengths it becomes 9x.
Which to use
- Cost, this month - Argon, comfortably. Even after December it stays well under Astra.
- Long input - Argon, unless you also need long output. Astra's 272K rule makes long input expensive in a way its sticker price hides.
- Long output from long input - Astra. The separate 128K output cap is the only thing here that guarantees room for the answer.
- Heavy caching - Argon until 30 November at 95% off; recheck in December.
- Budget stability - neither is good. Argon doubles on a known date; Astra doubles on an input length that varies per request. Argon is the more forecastable of two imperfect options.
Verdict
Argon is substantially cheaper and will stay cheaper after the introductory period ends - but anyone quoting $2/$10 past 30 November is quoting a price that no longer exists.
Astra's problem is structural rather than seasonal. Rebilling an entire request because input crossed a line is a pricing rule that punishes exactly the workload its 1M context is sold for. Check your median input length against 272,000 tokens before committing to it, because that single number moves your bill by 2x.