Field notes · September 2026
Use the cheap model to write code
GPT-6 Luna at max effort writes code as well as smarter models like Claude Opus and GPT-6 Astra, for a fraction of the price. The saving on every task dwarfs the one-time cost of switching. Numbers from our own team of AI coding agents.
$0.95
Luna max, per coding task
$9–23
the same task on Claude Opus 5.5, GPT-6 Sol or GPT-6 Astra
58%
of Luna's work passes review first time. The smarter models: 56–67%
$1.22
to switch a stuck agent up, once. The cheap model saves $8–22 on every task
Luna max costs a fraction of the other coders
Luna max40 tasks
$0.95
Sonnet 55 tasks
$1.80
Opus 5.5 high13 tasks
$9.40
Sol high11 tasks
$11.20
Astra high23 tasks
$23
Cost of one coding task, by the model doing the coding. A task runs about 170 turns, each re-reading the whole conversation. Price per token, times that, is the bill. Sonnet has too few tasks to trust yet.
Its work passes review as often
Luna max24 reviews
58%
Sol 5.6 high32 reviews
56%
Luna 5.6 xhigh39 reviews
51%
Opus 5.5 high6 reviews
67%
Opus 5 high7 reviews
71%
Sol 6 high8 reviews
88%
Share of coding work an independent reviewer passed clean on the first look. Sol 6 and both Opus rows have too few samples to trust yet.
Spend the smart models where they run once: the spec and the review
Spec · Opus 5.5Code · Luna maxCode · AstraReview · Opus 5.5
Luna codes, Opus specs and reviews
$13.50
Astra codes, Opus reviews
$24+
Luna codes, no spec
$5
Cost of one finished task, three ways to staff it. A spec is written once and a review runs once per attempt, so a smart model there is paid once. A coder is paid on every turn. The cheapest row, no spec, is where most of our rework came from.
Switching to a smarter model costs $1.22, once
Read from cache, a tenth of the priceRead fresh, full price
Before
Luna
$0.00
Switch
Astra
$1.43
Next turn
Astra
$0.18
The common objection: switching loses the cache. It does, for one turn. The new model reads the conversation once at full price, about 150k tokens here, and caches it. One real agent, three turns in a row.
That is less than one code review
Luna → Astra
$1.22
Sol 5.6 → Sol 6
$0.25
Astra → Luna
$0.01
One code review
by Opus 5.5
$1.41
by Astra
$4.28
Measured over 11 switches. A coder that keeps failing review costs a review every attempt, so moving it up pays for itself on the first attempt it saves.
Default coders to Luna maxSame first-review pass rate as the pricier models, at $0.95 a task.
Put Opus or Astra on the spec and the reviewThey run once per task, so their price is paid once.
Move a stuck coder up without worrying about the cacheIt costs $1.22, once. Another failed review round costs more.
Token counts from our agents' logs, Sep 12–30, 2026, priced at API list prices per million tokens (input / cached / output): Luna $0.10 / $0.01 / $0.50 · Sol $2 / $0.20 / $10 · Astra $10 / $1 / $50 · Opus 5.5 $4 / $0.20 / $20 and Sonnet 5 $3 / $0.30 / $15, both plus cache writes at 1.25× input. Astra is the median task. Staffing rows assume 2.84 review rounds for Luna (measured) and 1 for Astra (a floor), and a spec averaging $8.50 over 6 tasks. Samples are small. Astra coders were handed the hardest, most-stuck tasks, which raises their cost and makes their pass rate incomparable. That a spec reduces rework is our experience, not yet measured. On a subscription, read dollars as quota.
← All dispatches