Everyone's raving about the 'deep research' mode. Claims of "long context" and "thorough analysis." Let's get specific.
What's the *actual* token limit? The marketing says "big." My infrastructure bills care about numbers.
* Is it 128K? 200K? 1M?
* Is it a hard limit, or does it chunk/stream beyond a point?
* What's the real-world doc/page count before it starts forgetting the intro?
Show me a `curl` call or a config snippet that proves it. Without that, it's just vibes.
Because if your "deep research" on a 100-page PDF costs 10x more due to hidden context scaling, that's not a feature—it's a pricing trap.
— show the math
show the math
The limit is 200k tokens. It's a hard limit, not a streaming window. The system will reject your request if you exceed it.
Their pricing page shows context scaling. Input for their 200k model is 10x more per token than their standard 8k model. You're right about the cost trap.
For a 100-page PDF at roughly 500 tokens/page, you're at 50k tokens. You'll stay under the limit but the input cost for "deep research" is the real multiplier. Check the `/v1/chat/completions` endpoint and the `context_length` parameter in the model name.
Prove it with a benchmark.