Note

WHAT THE NUMBERS SHOWED

Nothing crossed 272K.
Until this week, nothing could.

On 15 September, Kiro roughly doubled its credit rates and added a tier: any GPT-5.6 request over 272K tokens bills at double again. We went into our own data to find out what it had cost us. Across 1,558 GPT-5.6 runs over 26 days, the number that crossed 272K was zero — and the reason turned out to be less flattering, and more useful, than we expected.

WHAT CHANGED

The base rate doubled.
Then a second rate appeared above it.

We did not have to take anyone's word for the old prices. Kiro's client stores every past conversation with the rate multiplier that applied at the time, so the change is datable from your own disk. Both columns below were read out of that store.

Kiro credit rate multipliers before and after 15 September 2026, and the new long-context tier
Model Until 14 Sepobserved locally From 15 Sepup to 272K Changebase rate Over 272Knew tier
gpt-5.6-sol2.4×4.4×+83%8.8×
gpt-5.6-terra1.0×2.2×+120%4.4×
gpt-5.6-luna0.1×1.1×+1,000%2.2×

The tier applies to these three models and no others. Luna is the row that produced most of the noise: a model priced as an afterthought at 0.1× became a real line item overnight, and anyone who had been using it freely for exactly that reason felt it first.

Why a threshold stings more than a rate rise: a rate you can budget for. A cliff you cannot, because nothing in a normal working day tells you how close to it you are. You find out when the balance moves.

WHAT WE MEASURED

1,558 runs against the line.
None of them reached it.

Every task Selvage has run through Kiro since mid-August, read back from the client's own conversation store. Not a sample and not a benchmark we designed — the working history, including the runs that failed.

Distribution of peak context across 1,558 GPT-5.6 runs A histogram in 17,000-token bands. The tallest band is 17 to 34 thousand tokens with 321 runs. Counts fall steadily across the range to 17 runs in the 221 to 238 thousand band, then rise again to 37 runs in the final band immediately below 272,000 tokens, where the model's context window ended. No run exceeded 272,000 tokens. PEAK CONTEXT PER RUN · 1,558 GPT-5.6 RUNS each bar is a 17K-token band; the last bar sits against the old window limit 100 200 300 20 321 278 190 154 115 101 79 83 61 38 27 19 17 18 37 0K 34K 68K 102K 136K 170K 204K 238K 272K 272K the GPT-5.6 window until 14 Sep the price tier from 15 Sep Peak context reached in a single agent run. 0 of 1,558 crossed 272K. Source: Kiro conversation store, 14 Aug - 15 Sep 2026.
The tail decays as you would expect, and then stops decaying. The final band — 37 runs pressed against 272K — is not a long tail. It is a queue at a wall.

THE SAME DATA, TWO WAYS

A run is not a task.
Both numbers are in here.

Selvage gives every issue its own Git worktree, and every agent inside it — implementer, then a fresh reviewer — its own conversation. Context is billed per run, so that is the honest unit for pricing. Tasks are the unit a person recognises. The table reports both.

Peak context per agent run and per task, with each figure as a share of the 272K tier
Percentile Per agent run
1,558 runs
Share of tier Per task
575 tasks
Share of tier
Median64,62224%99,25636%
75th116,05143%150,30655%
90th169,28362%215,64679%
95th216,89380%247,82891%
Highest272,000100%272,000100%
Crossed the tier0 of 1,558—0 of 575—
Read the last row carefully. 272,000 appears as the maximum in both columns because it was the maximum — it is the context window GPT-5.6 had until 14 September, not a number our runs happened to stop at. A percentile that lands exactly on a hard limit is reporting the limit, not the behaviour.

The volume matters as much as the ceiling: 575 tasks over 26 working days is 22 a day on average and 85 on the heaviest, across 9 repositories. That removes the obvious objection — that context stays small because not much is happening.

WHY NOTHING CROSSED

We were not managing context.
The window was.

The flattering version of this note stops at "zero crossings" and takes a bow. The data does not support it. 1,532 of those 1,558 runs were on a 272K context window — they could not cross the tier, because the tier is drawn exactly where the window used to end.

Runs reaching 95 percent or more of the context window, by model
ModelRunsReached 95%+ of the windowRate
gpt-5.6-luna649294.5%
gpt-5.6-terra65571.1%
gpt-5.6-sol25410.4%
All GPT-5.61,558372.4%

Thirty-seven runs went all the way to the wall. Twenty-five of them were compacted there — Kiro summarised the conversation and kept going, which is recorded in the conversation itself. Those 25 are the only compactions in the entire history, and every one happened at the window limit. Nothing else ever got close enough to need one.

So the guardrail was never ours. There is no context budget in Selvage. No threshold, no warning, no discussion of one — and we would have told you, before this week, that we simply did not generate much context. What actually happened is that a hard limit stopped 37 runs on our behalf, silently, and we never had cause to notice.

That limit is now gone. GPT-5.6 moved to a 1M-token window on 14 September, and the same boundary came back the next day as a price. The 37 runs that were stopped by a wall are, on the new window, 37 runs that would have kept going — into a band that bills at double.

WHAT WE TOOK FROM IT

The shape of the work is the budget.
It was never a cost decision.

The distribution is the part that survives the correction. Four runs in five peak under 136K, half the tier. That is not the window's doing — the window only ever acted on the last 2.4% — and it is the reason we expect the tier to stay mostly irrelevant to us even now that the ceiling is a price instead of a wall.

It falls out of the shape of the work. A task in Selvage is one issue, in its own worktree, in its own conversation; when the task ends, the conversation ends with it, and the reviewer starts from nothing rather than inheriting the implementer's afternoon. Context only grows where something lets it accumulate. A long interactive session is exactly that — every file read, every failed attempt, every correction stays in the window and is paid for again on the next turn. A task-per-issue harness has nowhere to put that history.

One issue per task was a decision about accountability. We wanted a reviewer to see a diff small enough to judge, and a failure attributable to something specific rather than to a long afternoon. Cost never entered the design conversation, because at the time cost was not the problem. Then someone else changed their pricing, and the same boundary turned out to be worth money.

What we would have concluded without the data. Two wrong things. Our first instinct blamed context growth for the rise in credits — it was the base rate, which doubled for every request regardless of size. Our second was that our context discipline was doing the work; a hard limit was doing the last of it. The measurement was worth more than either instinct, which is the same reason Selvage keeps evidence for everything else it runs.

We are not claiming foresight. We did not see this coming and could not have. The honest lesson is narrower: constraints adopted for one good reason tend to pay out somewhere you were not looking — and the ones you did not adopt, but were handed by a platform, can be taken back without notice. Knowing which is which requires measuring, not remembering.

Method. Figures read from the Kiro CLI conversation store (~/Library/Application Support/kiro-cli/data.sqlite3), covering 14 August to 15 September 2026: 575 Selvage tasks, 1,766 agent runs across 9 repositories, of which 1,558 were on the three GPT-5.6 models the tier applies to. Peak context per run is the maximum context share Kiro recorded for any request in that run, multiplied by that run's context window. Rate multipliers are the values stored on each conversation at the time it ran.