· via dev.to (home feed)
Sixfold context growth, not cache windows, decides the cost of resuming AI coding sessions
Palo Alto AI Research Lab measured its coding sessions' starting context growing sixfold in three months, and argues context size, not cache window length, now decides the cost of resuming a session.

A sixfold jump in three months
A dev.to post from Palo Alto AI Research Lab reports first-hand measurements showing that the median context carried into its AI coding sessions grew from 91,549 tokens in August to roughly 546,000 tokens by mid-November — about a sixfold increase in three months. The daily medians, drawn from session transcripts between November 8 and 18, ranged from about 490,000 to 546,151 tokens across 7 to 24 sessions per day. The lab's argument is that this growth, not the length of provider cache windows, is what now decides the economics of picking an old session back up.
The cache window question
The post starts from a question by co-author Anton Dziatkovskii: after 40 to 60 minutes away from Claude Code, Codex, Antigravity or Cursor, does resuming a session mean the whole history is resent and every token billed again? The lab's answer is that such a window exists but that the assumed duration matches no documented default.
For Anthropic, the post cites published mechanics of a 5-minute default cache lifetime plus a 1-hour extended tier: a 5-minute cache write is priced at 1.25 times the base input rate, a 1-hour write at twice that rate, and cache reads at one tenth. A notable detail comes from the lab's own tooling: the operating instructions of the session the post was written in reportedly state that its requests run with a 1-hour prompt-cache TTL, and that this drops to 5 minutes if the session enters usage overage. In other words, the protection shrinks by a factor of twelve precisely when consumption is highest.
Conflicting figures, and two tools with none
For OpenAI, the post describes a 30-minute window extendable to 12 hours, with up to 90 percent off cached input — but it flags contradictions between sources: one puts writes at 1.25 times base input and reads at 0.1, another prices a 12-hour write at 2.5 times, and a third, older source claims a 24-hour default. None is a vendor pricing page. On Cursor and Antigravity, the authors found nothing on session-resume costs at all; the single Gemini-side figure — reads roughly 75 percent cheaper than fresh input — comes from a source more than a year old.
Quota, not dollars
The post also reframes the question itself. Paying for tokens twice is API-billing language; on a subscription, a cache miss does not produce an invoice, it burns quota — the same tokens measured as a share of a weekly limit. That changes the advice: against dollars the rule is to stay inside the cache window, while against a weekly bucket it is to keep the context small. With a large context, the two rules point in opposite directions.
What a resume actually costs
Applied to the 546,151-token median, the post estimates that the first request of a resumed session costs about 55,000 tokens at the discounted read rate inside the window, versus the full 546,151 once the cache has expired — a difference of roughly 491,000 tokens per resume. The TTL decides only whether that sixfold-larger figure is discounted or paid in full.
Caching is also keyed on the shared prefix, and compaction rewrites the start of the context, so compacting before resuming guarantees a cache miss while buying a much smaller bill for everything after. The lab's recommendations: resume if you are back inside the window with the context unchanged; compact if the window has expired and the context is large; resume if the window has expired but the context is small. The case of a large context plus a day away lands squarely in the compact row. The authors add that compaction in their setup is a manual two-step, since a bare /compact does not read their enriched prompt file.
How reliable the numbers are
Only two items in the post come from primary sources, the authors say: the self-reported 1-hour TTL and their own transcript medians. The Anthropic and OpenAI multipliers are secondary write-ups that disagree with each other, and the claim that Anthropic cut its default TTL from an hour to five minutes in March 2026 appears in a single blog with no official confirmation the lab could find.
Why it matters
As agentic coding sessions routinely carry half a million tokens, cache windows become a discount on a number developers control separately. For anyone budgeting a subscription quota, context hygiene — compaction, summarisation, fresh sessions — now matters more than TTL tuning. And with no public resume-cost figures for two of the four tools in the original question, most of the market remains unmeasured.
- #llm
- #prompt-caching
- #claude-code
- #developer-tools
- #context-window