AI Balling on a Budget
Drafted with Fable, before was kill — a three-lane strategy for the credits era
Editor's note: You're reading the pilot issue of Many Many Victories — an occasional newsletter on AI loadouts, product strategy, and whatever the Raspberry Pi is doing this week. If it earns a No. 2, that's on you. Comments are open at the bottom.
I'm back for year 2 of mpd² at Northwestern, and my AI stack got patched hard over the summer. Today, Anthropic moves Claude Fable 5 onto usage credits — $10 per million tokens in, $50 out. The model I leaned on all spring is now metered, GPT-5.6 is stuck in limited preview, and a Chinese lab is renting me a million-token context window for the price of a coffee. Time to re-draft the loadout.
Three lanes, about $67 a month at full spend: Claude Pro at $20 (Opus 4.8, Sonnet 5), ChatGPT Plus at $20 (GPT-5.5, plus GPT-5.4 Cyber for security work), and z.ai's GLM Coding Plan — GLM-5.2 with a 1M-token context — grandfathered at roughly $2 a month on legacy pricing. Current plans start at $18; if you're on legacy, don't let it lapse. On top: a $25 monthly cap on Fable 5 credits that I don't expect to hit.
The routing rule is one sentence: latency-sensitive work goes to the nearest subscription model, anything that can wait goes to GLM overnight, and Fable credits only come out when a task needs frontier quality and huge context and nothing cheaper delivered.
Here's the part my finance professors would appreciate: the chart needed a log axis (Fig. 1). GLM's Lite tier works out to roughly 75–450 million tokens a month. Claude Pro and ChatGPT Plus land in the tens of millions. My capped Fable budget buys one to two. In tokens per dollar, legacy GLM delivers about a thousand times what Fable credits do.
This is year 1 of mpd² sneaking into my hobby projects. My day job at Stryker involves Make vs. Buy business cases, and this stack is the same analysis in a hoodie: subscriptions are sunk cost, Fable is pure marginal cost. Route bulk work to the deepest flat-fee quota, keep the sunk lanes busy by day, and treat the metered lane like a consultant you bring in for one hard question — not a staff hire.
The daily cycle borrows from marathon training: quality work when I'm fresh, volume when it's dark. Days are for architecting a tight spec on Opus 4.8 or GPT-5.5. Overnight, GLM-5.2 gets the spec plus the whole repo in one 1M-token, off-peak pass (Fig. 3). Mornings are for reviewing the diff and spot-fixing. Only if GLM fell short do the Fable credits come out. Base miles are cheap; race shoes are for race day.
Stated plainly, because I fact-checked this harder than my capstone: July 7 is a billing cutover, not a removal — Anthropic says Fable "isn't permanently leaving subscriptions," which is why I'm not pre-loading credits. GPT-5.6 is preview-only, whatever the hype threads say (prediction markets near 97% for July; odds are not release notes). The cyber model is GPT-5.4, not 5.5. And nobody publishes exact token quotas — the chart bands are estimates, drawn as ranges instead of false precision.
There's a fourth lane that costs nothing and produces mostly humility: OpenClaw on the Raspberry Pi in my office — technically operational, perpetually underutilized, the first thing I show guests. The only quota it burns is my patience. I've been thinking about trying Hermes on for size as the local brain; if that experiment happens, it gets a full issue. Pilot newsletter, pilot hardware.
Year 1 taught me to interrogate the economics under a product decision; apparently it took, because I now run portfolio strategy on my chatbots. The meta will shift — Fable may return to subscriptions, GPT-5.6 will land, the legacy pricing will die like all good exploits. The framework holds: know your sunk costs, meter your marginal ones, never pay frontier prices for base miles.
See you at the starting line, cohort 2.
flowchart LR
REQ([Incoming request]) --> DEC{Latency-sensitive?}
DEC -- Yes --> FRONT{Frontier quality
needed?}
FRONT -- No --> SUB["Subscription daytime
Opus 4.8 / Sonnet 5
GPT-5.5 / GPT-5.4 Cyber"]
FRONT -- Yes --> FAB["Fable 5 credits
$10/$50 per Mtok
NICHE ONLY"]
DEC -- No / overnight --> GLM["GLM-5.2 (Lite, flat)
1M context
use spare weekly quota"]
SUB --> DONE([Done])
FAB --> DONE
GLM --> REVIEW([Morning review / diff])
REVIEW --> DONE
classDef claude fill:#b3cbe4,stroke:#185FA5,color:#0C447C
classDef fable fill:#e5aeae,stroke:#A32D2D,color:#791F1F
classDef glm fill:#e8c98f,stroke:#854F0B,color:#633806
classDef plain fill:#efe9d8,stroke:#8a8172,color:#1c1710
class SUB claude
class FAB fable
class GLM glm
class REQ,DEC,FRONT,REVIEW,DONE plain
flowchart LR
PLAN["Plan — day
Opus 4.8 / GPT-5.5
spec + task breakdown"] --> EXEC["Execute — night
GLM-5.2 whole-repo, 1M ctx
off-peak · flat fee"]
EXEC --> REV["Review — morning
diff review, spot-fix"]
REV -. only if GLM fell short .-> FAB["Fable 5 credits"]
REV --> SHIP([Ship])
classDef claude fill:#b3cbe4,stroke:#185FA5,color:#0C447C
classDef fable fill:#e5aeae,stroke:#A32D2D,color:#791F1F
classDef glm fill:#e8c98f,stroke:#854F0B,color:#633806
classDef plain fill:#efe9d8,stroke:#8a8172,color:#1c1710
class PLAN,REV claude
class EXEC glm
class FAB fable
class SHIP plain
What's your current meta — and what am I sleeping on? Should the Pi run Hermes, or does OpenClaw deserve one more patch cycle? Drop a comment below. Enough letters and the pilot becomes a series.
Setup: Claude Pro $20 · ChatGPT Plus $20 · GLM Coding Plan ~$2 (legacy) · Fable 5 credit cap $25. Estimates as of July 6–7, 2026; claims validated against a separate claim register with GPT-5.5 as the checker, which felt appropriately adversarial.
Fact desk: Claude usage limits · Codex rate card · Codex burn rates · GLM plan limits · Fable 5 "not permanently leaving subscriptions"
Many Many Victories is published whenever the meta shifts. Corrections to the Product Desk. Printed on 100% recycled tokens.