Context Stability

Context stability is a model's ability to maintain consistent application of every design rule, brand guideline, component name, and voice constraint you set at the start of a session all the way to the final output instead of gradually forgetting or mutating them. Claude Opus 4.8 finally delivers usable levels of it for design work. The model can now run through a 50-screen audit or full system migration and still respect the exact token scale you defined in the first message instead of quietly switching to its own made-up conventions by message 12. This is not about raw context window size. The 1M token window has been around since Opus 4.7. Stability is the quieter training advance that keeps the model from drifting when the task stretches across hours and thousands of tokens.

Context stability is not the same as a big context window. You can cram an entire design system into a 200K model and still watch it forget your no-exclamation-points rule by screen 18. It is not prompt repetition either. Pasting your brand book at the top of every message is a 2023 crutch that burns tokens and still collapses at real scale. It is also not session resets or the prompt-pinning gymnastics designers used with Claude 3.5 Sonnet to fight drift. Those were workarounds for unstable models. Real stability means the model treats your 2024 Figma token spec as permanent law across a four-hour session instead of treating it as friendly advice that fades with time. It does not make the model tasteful or creatively brilliant. It only stops it from sabotaging itself.

Take the 2024 Primer redesign at GitHub. The team loaded their complete legacy 2022 component library, the new 2024 token primitives, 11 months of design critique threads from Linear, and the brand voice guide that forbids passive language. Older models would nail the first 30 components then start reintroducing deprecated colors and inventing component variants that never existed. Opus 4.8 held every convention from first token to last export. The migration that once required three designers and two weeks of reconciliation spreadsheets finished in one focused day with output clean enough to merge after a single review. The same pattern appeared at Linear in Q1 2025 during their checkout redesign. They fed the model their full Radix component library from 2023, the exact spacing scale, and 62 screen designs. Previous runs with GPT-4o always invented non-existent props like Button size="jumbo" by the halfway mark. Stable context kept the generated code inside the actual library 93 percent of the time. Another case hit at Miro in late 2024. Their researchers dropped 41 interview transcripts, three prior study reports, and 2023 persona docs totaling 240K tokens into one session. The model tracked a specific pain point about real-time cursors across 34 separate calls without mixing up participant quotes or inventing themes. Earlier tools sampling five transcripts at a time missed the pattern entirely.

Reach for context stability when the work lives or dies by consistency at scale. Use it for design system migrations like Airbnb's 2023 overhaul from their old design language to the new one or brand voice rollouts across 80 screens for a product launch like Figma's 2025 dev mode release. It delivers for complete research synthesis where every transcript must inform the insights instead of a cherry-picked handful and for design-to-code handoffs where your exact component taxonomy has to survive translation. These tasks expose any drift instantly and turn it into hours of painful cleanup. The stability makes the million-token window actually usable instead of a party trick you try once then abandon.

Avoid it for early ideation or pure craft judgment. If you are exploring five radically different visual directions for a 2026 identity project at Adobe do not load your current brand constraints. The stability will drag old rules into places they do not belong and kill the exploration. Start fresh sessions for each direction. Skip it for pixel-level layout calls or taste critiques. Asking the model to decide whether 4px or 8px of gap feels better on a particular card then hold that decision across 40 screens is pointless. It has no eye. It can only preserve decisions you already made with clarity. Context stability also becomes a straightjacket when you are deliberately breaking your current system. In those cases fresh context is the feature you want.

Context stability turns the million-token window from a spec sheet flex into the reliable backbone of actual design system work.