Claude Opus 4.8

SEEDED, calibration protocol (100% personal-evaluation weight). Anthropic frames 4.8 as 'a modest but tangible improvement' but the deltas matter where we work. Key benchmarks: SWE-bench Verified 88.6% (+1.0pp vs 4.7 87.6% — saturated), SWE-bench Pro 69.2% (+4.9pp vs 64.3% — the real coding gain), Terminal-Bench 2.1 74.6% (+8.5pp vs 66.1% — finally competitive, still trails GPT-5.5 82.7%), USAMO 2026 96.7% (+27pp vs 69.3% — huge math jump), GPQA Diamond 93.6%, GDPval-AA 1890 Elo (#1, +121 vs GPT-5.5, +137 vs 4.7), GraphWalks 1M-token F1 68.1% (+27.8pp vs 40.3%). Operational: 4x less likely to let its own code flaws pass unremarked (drives Self-Correction 17, Candor 17). ~15% fewer turns & ~35% fewer output tokens per task vs 4.7. Fixed the 4.7 skipped-tool-call regression (Tool Use 17). 'Most honest model yet' per Anthropic alignment assessment (Safety Alignment 18). Same $5/$25 pricing; Fast mode 3x cheaper ($10/$50, 2.5x speed). Effort enum low/medium/high/xhigh/max now on claude.ai. Aggressive recalibration: scored CLEARLY above 4.6 & 4.7 on coding + agentic + honesty per John's directive & Theo's stance. Sources: Anthropic model card, VentureBeat, Vellum, llm-stats, Artificial Analysis (GDPval-AA), digitalapplied. Cost Efficiency still low (5) — frontier pricing; the value is quality/turn-count, not $/token. --- 2026-06-09 ADDENDUM (Fable/Mythos 5 launch + LLM-training re-run): Opus 4.8 is now the DESIGNATED FALLBACK model for Claude Fable 5 on cyber/bio/chem/distillation queries. Fable/Mythos 5 System Card corroborates the GDPval-AA #1 (1890) and the honesty framing. ONE notable negative signal: on the fixed-CPU LLM-training-speedup re-run (System Card §2.3.7.1), Opus 4.8 scored 32.64x — BELOW its predecessor Opus 4.7 (50.67x re-run) and Mythos Preview (60.81x). Anthropic's read: Opus 4.8 'typically stops after a single round of follow-up optimization and judges the result sufficient'; its strongest attempts approach 4.7's strongest, so the capability is present but not as consistently exercised. This is the SAME 'stops early / declares done' pattern flagged in the honesty evals (false-completion, missing-context fabrication) — a known 4.8 trait, not a re-score trigger. Scores held; documented for traceability per. No 1-20 dimension changed.