Beyond token prices: the real break-even of AI work
Saving $3 on tokens is not a bargain if it creates $4 of extra work.
Reviewed
I. Three minutes can erase the saving
Model choice and the cost of finishing · Part 2 of 3
Saving $3 on tokens is not a bargain if it creates $4 of extra work. Compare the cost of finishing the task—not just the next response.
All examples below are calculated scenarios, not live performance tests. USD; human time is valued at an assumed $60/hour.
In Part 1, switching from GPT-6 Astra to GPT-5.6 Sol saved $3.212 across 30 further calls, including the cold start.
Scenario: 30 further calls; $60/hour for additional human work. Cold start included. Common review costs excluded equally. Scale starts at zero.
The switch remains cheaper only while its additional human work stays below about 3 minutes 13 seconds. Four extra minutes turn the $3.212 saving into a $0.788 loss. That is a sensitivity calculation, not a claim that Sol needs more correction.
Cache invalidation spends the same saving
That time allowance also assumes no further cache loss. Four full Sol-only prefix misses at calls 6, 12, 18 and 24 use $2.3552 of the saving, leaving $0.8568—about 51 seconds of extra human work. Six assumed Sol-only misses at calls 5, 10, 15, 20, 25 and 30 reverse the saving before any human rework.
If Astra also loses cache, calculate both paths again. Rebuilding cache and correcting work draw from the same remaining benefit. These loss schedules are stress tests, not measured miss rates.
II. A higher AI bill can buy a cheaper result
Suppose a failed automated attempt needs $60 of human completion. Assume failures are reliably detected and verification costs are equal.
| Hypothetical strategy | AI cost | Success chance | Expected total |
|---|---|---|---|
| Lower-cost executor | $2 | 50% | $32 |
| More capable strategy | $8 | 80% | $20 |
The extra $6 prevents $18 of expected fallback work. The strategies break even when human completion costs $20; above that, the more capable strategy wins.
The 50% and 80% are assumptions—not METR scores or measured model results. Without a real success advantage, the extra spend has no such benefit.
III. What about upgrading mid-session?
An upgrade does not become economical simply by running longer at higher prices. It must finish with less work, avoid failures, or reduce human involvement.
Reverse Part 1: Sol has a warm 100,000-token prefix; Astra starts cold. Each call adds 2,000 input tokens and generates 1,000 billed output tokens. The same API tariffs give:
Same assumed token size per call; no later cache misses. Required call counts are hypothetical, not measured model speed. Scale starts at zero.
Against 30 Sol calls, Astra must finish in at most seven equally sized calls to save on the API bill alone. Eight cost $2.606 and miss that threshold.
But ten Astra calls cost only $0.542 extra. Avoiding about 33 seconds of human work would recover that premium. If Sol cannot finish acceptably at all, thirty cheap calls are not an equivalent alternative.
IV. The decision to make
Count from the current checkpoint. Ignore already-paid costs when choosing the next strategy. Include retries, verification and repair. Price waiting only when it blocks valuable work; do not bill autonomous runtime as human labor by default.
An upgrade need not take over the whole session. Part 3 asks when one stronger advisor is enough.
Spend more only when the remaining cost it prevents is greater.
V. Details, assumptions and sources
Reproduce the baseline and cache-loss numbers
The examples use net input growth: previous responses carried forward are already included in the 2,000-token increment. Billed output includes reasoning. All new input is assumed cache-written; the baseline reuses the previous prefix on later calls. There is no compaction.
Standard prices per million tokens for GPT-5.6 Sol cache reads/writes/output are $0.40/$5/$20; Astra's are $1/$12.50/$50. Ordinary input is $4/$10 and is not charged again for cache-written tokens. All examples stay below 272,000 input tokens. Taxes, regional uplifts, tools and subscription quotas are excluded. Sol's promotional tariff is listed through at least November 21, 2026. Pricing.
Sol_warm(N) = 0.070N + 0.0004N(N−1)
Astra_cold(K) = 1.150 + 0.175K + 0.001K(K−1)
The $1.15 is Astra's cold-start surcharge relative to warm Astra. Sol's 30-call total is $2.448 here, not Part 1's $2.908: in Part 1 Sol also started cold.
For a later loss, charge only the lost prefix at the write-minus-read difference. Sol's four stress-test misses lose 110k + 122k + 134k + 146k = 512k tokens. Their premium is 512,000 × ($5 − $0.40) / 1,000,000 = $2.3552. No second initial-cold-start charge is added. Partial losses affect only the missing portion. For non-written input, use the ordinary-input category instead.
If both strategies suffer the six listed full losses, Astra costs $15.2970 and Sol $6.5788: switching then saves $8.7182. This illustrates the importance of a symmetric comparison, not a forecast. A compaction scenario must also account for the shorter subsequent context.
The general per-call bill is (ordinary × input price + reads × read price + writes × write price + billed output × output price) / 1,000,000. Keep these input categories disjoint. Prompt caching.
For the failure example: expected cost = AI cost + verification + (1 − success probability) × human completion cost. Adding $3 of extra verification raises the fallback break-even from $20 to $30. Unequal recovery costs or undetected errors need separate estimates.
What METR can—and cannot—add
METR relates task success to human expert task duration under a tested agent setup. Its time horizon is neither AI runtime nor saved working hours. Relevant task-level results can inform success estimates; multiplying a headline horizon by an hourly rate cannot establish savings. Methodology.
We do not use a Sol horizon numerically: METR's June 26, 2026 report calls its estimates non-robust because they depend heavily on treatment of evaluation-environment exploitation. A stuck session also needs success estimates conditioned on its current failure—not a fresh-task average. Sol evaluation.
One more cheap attempt?
If another attempt costs c, succeeds with probability p and otherwise leaves escalation cost E unchanged, trying once costs c + (1 − p)E. It is cheaper than immediate escalation when p > c/E.
At $0.20 versus $2, equality is 10%. Additional review or blocked time raises the threshold. A new hypothesis can justify another attempt; repeating the same failure is not an independent chance. Privacy, permissions and minimum quality remain constraints, not tradable discounts.
Sources checked September 22, 2026. Calculations independently reproduced; no new model-performance or provider-cache experiments.
Sources
Sources checked 2026-09-22.