← All themes and ideas
Agent reliability confidence: medium score 6.2/10 2026-11-12

Silent regressions: Fable 5.2 context window reportedly cut from 1M to 200k

The need

Users discover the advertised 1M context shrinking to ~200k in practice and can only find out by crowd-sourcing the Facebook feed; there is no independent visibility into real per-model limits.

Original post — verbatim Thai

ขนาด Context window ลดจาก 1M เหลือแค่ 200k

Group member · 2026-11-12  ·  open on Facebook ↗
Three business opportunities
OPPORTUNITY 01

Client-side model probe monitor

What
CLI/library that periodically probes live models for context ceiling, truncation behavior and latency, alerting on regressions via Thai/EN channels.
Who pays
Teams shipping context-heavy products pay per team to catch vendor-side regressions before customers do.
Price
THB199/mo/team (hypothesis)
Why now
2026's model churn and silent spec changes make this batch's bug report a recurring class of incident.
Effort
S
Moat
Historical probe dataset per model version that turns each alert into evidence.
Risk
Probe API costs and vendor pushback on benchmarking traffic.
OPPORTUNITY 02

Degradation-triggered fallback router

What
Gateway that reroutes to a healthy model (Opus, GPT) the moment probes detect context shrinkage, with quality checks on swapped runs.
Who pays
Agencies with SLAs pay per-token margin for automatic failover they can't build themselves.
Price
Token pass-through margin (hypothesis)
Why now
Multi-model availability in 2026 makes quality-aware routing a weekend project with compounding data.
Effort
M
Moat
Routing decision logs improve failover heuristics per task type.
Risk
Price wars collapse pass-through margins to zero.
OPPORTUNITY 03

Verified model-behavior status board

What
Thai/EN board of real-world per-version model specs, each report accompanied by a reproducible test snippet; enterprise API access monetizes the data.
Who pays
AI vendors and agencies consuming status data pay for the API; users use it free.
Price
Free + THB990/mo API (hypothesis)
Why now
Posts 22 and 39 show people deciding model choices on vibes; evidence infrastructure is missing.
Effort
S
Moat
Reproducible-test corpus makes reports more trustworthy than anecdotes.
Risk
Moderation load and low willingness to pay for 'just' information.