Claude Code recently shipped two features worth knowing about: Fast Mode and a 1M token context window. Both are genuinely useful. Both can surprise you on the bill if you don't understand how the pricing actually works.
I tested both today. My first Fast Mode message immediately spent $1.63 in extra usage. My spending limit was $0.10.
That's 1,630% over limit on a single message. Let me explain how that happens and what you should know before you turn either of these on.
Fast Mode: The Important Part First
Fast Mode goes entirely to extra usage billing. All of it. Your included plan allocation is irrelevant the moment you turn it on.
This is the thing that caught me. I had plenty of included usage left. Didn't matter. Fast Mode bypasses all of that and charges straight to your extra usage balance at 6x the standard rate.
The rates for Opus 4.6 in Fast Mode:
| Context size | Input (per MTok) | Output (per MTok) |
|---|---|---|
| ≤200K tokens | $30 | $150 |
| >200K tokens | $60 | $225 |
The >200K row is what the docs say happens when Fast Mode stacks on top of the long-context premium: 6x fast × 2x long-context = 12x on input, 6x fast × 1.5x long-context = 9x on output. In my testing though, the fast mode charges were at the straight 6x rate with no additional 1M context multiplier on top. My conversation stayed under 200K tokens, so I can't confirm the stacking behavior for the >200K tier from firsthand experience.
Thinking Tokens Are Why It Adds Up So Fast
Claude Code has extended thinking enabled by default, with a ~32K token budget. Thinking tokens are billed as output tokens.
In Fast Mode at the standard (≤200K) tier, output costs $150 per million tokens. So 32,000 thinking tokens alone runs about $4.80, before you've seen a single word of the actual response.
You don't see thinking tokens in the interface. They're invisible. The meter is running and you have no real-time visibility into it. This is how a single "hello, what files are in this directory?" message costs $1.63.
What Happened When I Tested It
I set a $0.10 extra usage spending limit, which felt reasonable for a quick experiment. Turned on Fast Mode. Sent one message.
$1.63 spent. Immediately.
Message 4 got rate-limited until I manually raised the limit to $2.85. After I raised it, the next message had no problem going over again. By this point I'd spent $2.85 total across a handful of messages. Nothing intensive. No large codebases. No complex multi-step tasks. Just poking around.
The rate limiting behavior is actually correct: it shouldn't reject requests that are under the limit. The problem is that Fast Mode burns through limits so quickly that you cross the threshold before you realize it. If you run a long automated session with extended thinking at these prices, hitting $50+ in a sitting isn't hard.
After turning Fast Mode off, I confirmed that the 1M context window alone doesn't touch extra usage at all. It pulls from your included plan allocation. Different beast entirely.
The 1M Context Window: Much Calmer
The 1M context window pricing is simpler than it sounds. The key is that the premium only kicks in when your actual input tokens in a single request exceed 200K. The configured max doesn't matter. If you have 1M context enabled but only 40K tokens in the conversation, you pay standard rates.
Standard Opus 4.6 rates, for reference:
| Context size | Input (per MTok) | Output (per MTok) |
|---|---|---|
| ≤200K tokens (actual use) | $5 | $25 |
| >200K tokens (actual use) | $10 | $37.50 |
Here's the important part: on Pro or Max plans, 1M context does not trigger extra usage charges. Not even close. I've been running multiple Claude Code instances with 1M context enabled for days, and none of them touched extra usage—the only instance that did was the one with Fast Mode turned on. The 1M context window pulls from your included plan allocation. It uses that allocation faster when you exceed 200K tokens (because of the 2x/1.5x multiplier), but it's still your plan's included usage, not a separate bill.
This is the critical distinction people need to understand: Fast Mode = extra usage from the first token. 1M context = included plan usage, even with the multiplier.
The threshold that triggers the premium rate is per-request actual token count. Not cumulative. Not configured max. Each request is evaluated independently. If you stay under 200K actual tokens per request, you're paying standard rates from your plan. No surprises.
One More Thing: Mid-Conversation Fast Mode
If you turn Fast Mode on in the middle of a long conversation, you pay the full Fast Mode uncached input price for the entire conversation context in that request. The history that was free (or cheap) before doesn't get a discount just because it existed before you switched. All of it reprices at Fast Mode rates.
If you're going to use Fast Mode, turn it on from the start of a session, not halfway through a 50-message thread.
The Short Version
- Fast Mode = extra usage, always. Your included plan allocation does nothing for Fast Mode requests.
- 1M context alone does NOT trigger extra usage. It pulls from your included plan allocation. The multiplier applies if you exceed 200K tokens in a single request, but it's still included usage, not extra charges. I confirmed this across multiple sessions.
- The 200K threshold is per actual request tokens, not configured max. Enabling 1M context doesn't change your price until you actually use more than 200K in one request.
- Thinking tokens are expensive and invisible. 32K thinking tokens at Fast Mode output rates is roughly $4.80 per response, before any visible output.
- Set your extra usage limit carefully. A $0.10 limit will not survive a single Fast Mode message with thinking enabled. Set it to at least a few dollars, or don't use Fast Mode at all.
- Switching to Fast Mode mid-conversation reprices everything. The full context history is billed at Fast Mode uncached rates for that request.
The 50% introductory discount on Fast Mode expired February 16, 2026, so these are the real prices now. Whether Fast Mode is worth it depends entirely on what you're doing. For a short burst of high-stakes work where speed matters, maybe. For ambient background coding sessions where you've set it and forgotten, definitely not.
I'll take my chances with standard mode for now. The speed difference isn't worth $1.63 per "hello."
Enjoyed this post?
Get notified when I publish something new. No spam, unsubscribe anytime.