1.8% of Uncached Requests Ate 15% of SemiAnalysis's Opus 5.5 Bill
SemiAnalysis checked its Opus 5.5 bill and found that only 1.8% of requests ran without a cache. On those requests the model re-read the entire context at full price, and they accounted for about 15% of spending. Most often these were new subagents, which start with no cache, like a new hire who has to be briefed on the whole project from scratch. Another 12% of sessions paid about $3 each to resume after an idle period longer than an hour. The takeaway for teams running agents: track the share of uncached requests separately.
AI-processed from X: @SemiAnalysis_; edited by Hamidun News
SemiAnalysis, an analyst firm focused on AI infrastructure, looked at its own Opus 5.5 bill and found a skew. Only 1.8% of requests ran “without a cache”: the provider had to process the entire context again instead of reusing the already-read portion at a reduced price. Yet that tiny slice generated about 15% of all Opus 5.5 spending. Put simply, roughly one request in fifty-five cost nearly a seventh of the bill. If you run coding agents, the same thing is probably happening to you. You just haven't measured it yet.
What an “uncached request” looks like in practice
Prompt caching works like a bookmark in a book. Say an agent is working with a repository: the start of the request holds the instructions, the project files and the command history. If the beginning of the request matches the previous one, the provider does not recompute it and charges less for that part. If there is no bookmark, the entire accumulated context is paid for again. Agentic tasks often carry large contexts, so every such miss noticeably hits the budget.
Where the expensive misses come from
SemiAnalysis names two causes.
The first is new subagents. According to the analysts, most uncached requests came from subagents that started work from scratch. The main agent spawns a helper, and the helper has to receive the full briefing again, like a new employee who is told about the whole project on day one. Many short-lived subagents mean many of these “first days,” and each one is billed.
The second is idle time. According to SemiAnalysis, 12% of sessions resumed after a pause longer than an hour and paid about $3 to re-read the context. A developer goes to lunch, comes back and continues in the same session. To them it is the same task, but to the bill it is almost a fresh start, because the saved cache no longer works by then.
A caveat: $3 per resumption is a small sum. The problem is scale, when such resumptions happen in every tenth session across many people. Also, the figures come from one company and one model, so the proportions on your team will differ.
What to do in practice
First, measure the share of uncached requests separately, not just overall token consumption. The average hides rare but expensive misses: 1.8% of traffic is barely visible on a chart, yet it accounts for about 15% of spending.
Second, find out who generates them. Set up monitoring that lists the sessions and subagents with the most uncached requests. If subagents account for most of them, check whether each one is necessary, or whether the task can be done in a context that is already running.
Third, account for pauses. If people often return to a session after a long break, it makes sense to compress the history into a short summary before resuming, rather than resending the entire previous context.
Fourth, make the beginning of subagent prompts identical. It is better to keep the shared instructions and project context in one block, in the same order. This is a general rule of how caching works, not a conclusion from the SemiAnalysis post: any change at the start of a request, such as a timestamp, can make the request “cold.”
The main lesson for founders and product managers: an agent's cost is determined not only by the price of a token, but also by how often you pay for the same context again. It is a controllable architectural parameter, on par with the choice of model. The SemiAnalysis post is part of a series (3/4), but this part alone shows where to start optimizing: with the share of uncached requests and the subagents that create it.
- Nov 18, 2026GPT 6.2 Sol Ultrafast Delivers 150 Tokens/s but Burns Through Limits 6x Faster
- Nov 18, 2026Visa Test: 22 AI Models Passed 57% of Purchases, and in 54% of Failures Still Paid
- Nov 18, 202611 open-source repos chained together for Claude Code, from research to voiceover
- Nov 18, 2026JetBrains' Mellum Runs Locally in an IDE: 5 Out of 10 on a MacBook M3 Pro
Key Questions Answered
What an “uncached request” looks like in practice?
Prompt caching works like a bookmark in a book. Say an agent is working with a repository: the start of the request holds the instructions, the project files and the command history. If the beginning of the request matches the previous one, the provider does not recompute it and charges less for that part. If there is no bookmark, the entire accumulated context is paid for again. Agentic tasks often carry large contexts, so every such miss noticeably hits the budget.
Where the expensive misses come from?
SemiAnalysis names two causes. The first is new subagents. According to the analysts, most uncached requests came from subagents that started work from scratch. The main agent spawns a helper, and the helper has to receive the full briefing again, like a new employee who is told about the whole project on day one. Many short-lived subagents mean many of these “first days,” and each one is billed. The second is idle time.
What to do in practice?
First, measure the share of uncached requests separately, not just overall token consumption. The average hides rare but expensive misses: 1.8% of traffic is barely visible on a chart, yet it accounts for about 15% of spending. Second, find out who generates them. Set up monitoring that lists the sessions and subagents with the most uncached requests. If subagents account for most of them, check whether each one is necessary, or whether the task can be done in a context that is already running.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.