Our own agent logs from 27 September to 30 October 2026, priced at API rates: Claude Ultra (formerly Max 20x) bought 1.7 times more per percent than Max 5x, Codex held 10 dollars per percent on Pro 100 and Pro 200 alike; whether more dollars buy more finished work is still an open question.
OpenAI doubled Codex limits on Pro 100, 200 and 500 and dropped the Pro 200 cut it had announced for 30 October. A day later Anthropic renamed Max 20x to Ultra: same $200, same weekly limit, no five-hour limit for verified accounts (coverage). Before that came weeks of "limit reached every week" reports and forum threads asking what Pro 200 still buys.
Two months in one list
Tap an event for details, or filter by who made the change.
10-30 · Max 20x becomes Ultra Same price and weekly limit; verified Ultra accounts lose the five-hour limit.
So what is each plan worth?
We run a team of coding agents on both providers around the clock, and we switched plans on both. We priced every token in our logs from 27 September to 30 October at API list prices and added it up per usage window. These numbers describe that period only; both providers changed their plans at the end of it.
$17per 1% of a Claude week on Max 5x
$26–35per 1% on Ultra: about 1.7 times, not 4
$10per 1% of Codex on Pro 100 and Pro 200 alike
~$2,900 vs ~$900API value of one full window: Ultra vs Pro 200
Claude. API-equivalent dollars per weekly window, stacked by model; the figure above each bar is dollars per 1%. Windows A–B on Max 5x, C–H on Ultra. F and H are partial windows.
Opus 5.5Sonnet 5.5
Codex. API-equivalent dollars per window, stacked by model; the figure above each bar is dollars per 1%. Windows 5–8 on Pro 100, 9–17 on Pro 200. Most windows ended in an early reset, so bar height tracks how long a window lasted, not its size.
Sol 6.1Sol 6Astra 6Terra 6.1Luna 6
Claude, 27 September–30 October: bigger plan, not four times bigger
Max 5x → Ultra roughly 1.7× per 1%. $16–18 before, $26–35 after. Ultra, the former Max 20x, is sold as four times Max 5x.
A full Ultra week held about $2,500–2,900 of our work.
Sonnet weeks bought less per percent. Windows that were about half Sonnet came in at $26–29; Opus-heavy windows at $30–35. Sonnet appears to use more of the limit than its API price suggests.
Cache reads are most of the value. Long-running agents keep reading the same context.
Claude, API-equivalent dollars per 1% of the weekly limit, by window. Hover a bar for its window, plan, share used and API value.Show all 8 Claude windows
Claude windows, API-equivalent dollars by model (* partial or irregular window)
Window
Plan
Used
Opus 5.5
Sonnet 5.5
Total
Per 1%
A* · 09-27 → 09-30
Max 5x
66%
$1,148
$16
$1,163
$17.6
B* · 09-30 → 10-02
Max 5x
85%
$1,278
$106
$1,384
$16.3
C* · 10-02 → 10-07
Ultra
97%
$2,407
$533
$2,940
$30.3
D · 10-07 → 10-14
Ultra
95%
$1,139
$1,344
$2,484
$26.1
E · 10-14 → 10-21
Ultra
93%
$1,165
$1,392
$2,558
$27.5
F · 10-21 → 10-22
Ultra
47%
$745
$632
$1,378
$29.3
G* · 10-22 → 10-28
Ultra
96%
$1,477
$1,297
$2,774
$28.9
H* · 10-28 → 10-30
Ultra
22%
$751
$20
$771
$35.0
Codex, 11–30 October: about $10 per percent on Pro 100 and Pro 200
We moved from Pro 100 to Pro 200 on 18 October, and the meter did not move. Before, 11–17 October (windows 5–8 on Pro 100): $9.5 per 1%, and window 6 reached 100% at $869. After, 18–30 October (windows 9–17 on Pro 200): $9.7 per 1%, and window 13 reached 100% at $913.
The meter follows list price. Windows with Astra 6 (5× Sol's price) or Terra 6.1 (half) still landed at $8.7–10.8 per 1%.
The difference between plans was resets. On Pro 200 our window reset early every one or two days, so we got more windows, not bigger ones.
Our logs end on 30 October, before the doubling on 29 October could show up in a full window.
Codex, API-equivalent dollars per 1% of the limit, by window. Hover a bar for its window, plan, share used and API value.Show all 13 Codex windows
Codex windows, API-equivalent dollars by model (* early reset, jump or open window)
Window
Plan
Used
Sol 6.1
Astra 6
Terra 6.1
Luna 6
Total
Per 1%
5* · 10-11 → 10-13
Pro 100
32%
$254
$76
–
$15
$345
$10.8
6* · 10-13 → 10-15
Pro 100
100%
$684
$126
–
$38
$869
$8.7
7* · 10-15 → 10-15
Pro 100
12%
$115
–
–
$6
$122
$10.1
8* · 10-15 → 10-17
Pro 100
40%
$384
–
–
$14
$407
$10.2
9* · 10-18 → 10-18
Pro 200
20%
$194
–
–
$5
$199
$9.9
10* · 10-18 → 10-20
Pro 200
49%
$478
–
–
$19
$497
$10.2
11* · 10-20 → 10-23
Pro 200
45%
$418
–
–
$22
$440
$9.8
12* · 10-23 → 10-24
Pro 200
37%
$357
–
–
$7
$364
$9.8
13* · 10-24 → 10-25
Pro 200
100%
$910
–
–
$3
$913
$9.1
14* · 10-25 → 10-25
Pro 200
16%
$159
–
–
$4
$162
$10.2
15* · 10-25 → 10-26
Pro 200
37%
$366
–
–
$6
$372
$10.1
16* · 10-26 → 10-28
Pro 200
39%
$215
–
$145
$2
$362
$9.3
17* · 10-28 → 10-30
Pro 200
38%
$342
–
$52
$8
$402
$10.6
So which $200?
More dollars of API value is not more work, and we can't yet say which plan gets more of it done.
Benchmarks price a task in tokens; our work doesn't behave like that. Public indexes such as Artificial Analysis v5 show some models using far fewer tokens per task, but a benchmark task is not our task.
Fewer tokens per turn did not mean less usage for us. A model running at lower reasoning effort saved tokens per turn, then spent them again on the back-and-forth of reviews, fixes and retries.
Our merged PRs came out close. One led by Sol 6.1 cost a median $5.5 of API value; one led by Terra 6.1, $4.7. Half the token price did not halve the cost of a change.
We use both: Claude for the leads that review and decide, Codex for most of the implementation. So we leave it open: which $200 does more of your work depends on how many rounds your agents need, not only on price per token. We will keep measuring finished work rather than tokens. If you measure this too, we would like to compare notes.
How our two $200 plans get spent
Every agent is a TMT member in a squad. Here is who used what, per calendar week, across both providers.
Share of active agent sessions per week, including subagents. The first week is almost all Claude because older Codex logs were pruned. Sessions we could not tie to a squad are "TMT, unattributed".
Core uses a fifth to a third of the dollars with a small share of the sessions: few agents, long contexts.
Night has many short sessions: lots of agents, a small share of the spend.
We ran 38 to 124 agent sessions a week on one machine.
How we measured
Sources: Claude Code transcripts and Codex session logs on the one machine the team runs on, 27 September to 30 October. Messages deduplicated by ID; automatic code review excluded.
Prices: public API list prices (Claude, OpenAI), fast mode at 2×. It is a ruler, not a bill.
Percent: read from each tool's status line, whole numbers only, so short windows carry about 5–8% error.
Gaps: Codex logs before 11 October were pruned; usage on web apps or other machines is not counted; squad labels are inferred from identities and worktrees.
Why we track it
TMT's digest watches how fast each squad burns its window and paces the team so it doesn't hit a wall on day three. That only works if we know what 1% is really worth. When a plan changes, we want to see it in our own numbers first.
