Claude Opus 5 vs Fable 5: What It Really Costs

Anthropic says Claude Opus 5 "comes close to the frontier intelligence of Claude Fable 5 at half the price." That is exactly true. It is also exactly what Opus 4.8 already cost. Half the price of what? We read the launch post, the pricing docs, the migration notes and all 193 pages of the system card, then checked the lot against the only independent lab that has actually measured this model. What it really costs once thinking is on by default, the chart on page 151 that no launch-day article mentioned, and the reason the "very verbose, slow, expensive" number everyone is about to quote is not the number you get. Verdict at 9:27. —— LINKS —— Claude Opus 5 (official announcement): https://www.anthropic.com/news/claude... Claude pricing: https://www.claude.com/pricing No affiliate link on this one. We know of no public affiliate programme for Claude subscriptions, and nothing in this video is sponsored. Standing rule either way: verdicts are never influenced by commissions, and tools that score badly get no affiliate link. The verdict here was written before we checked whether a programme existed. —— THE NUMBER THIS VIDEO TURNS ON —— Artificial Analysis measured Opus 5 at every effort setting. Their headline entry — the one that reads "very verbose", "slower than average" and $3,835.51 — is the MAX EFFORT configuration. The API default is "high". Check it yourself: effort index output tokens cost to run the AA index low 50.61 12.1M $556 medium 56.28 28.8M $1,115 high 58.86 52.1M $1,974 xhigh 60.07 76.4M $2,910 max 60.69 101.2M $3,836 Fable 5 59.86 86.9M $5,631 At the default, Opus 5 used 17% FEWER output tokens than the 63M peer average, and cost 65% less to run than Fable 5 for a score one point behind. At xhigh it beat Fable 5 outright for 48% less. Anthropic's "half the price" claim survives contact with an independent lab's actual bill. —— EVERY CLAIM SOURCED —— • Claude Opus 5 announcement — the "half the price" wording, the CursorBench "at max effort, within 0.5%" line, Sualeh Asif's "just under Fable 5" quote, the cybersecurity limitation, the 85% cyber-classifier figure, and the "no data retention requirements for general access" line (Jul 24, 2026): https://www.anthropic.com/news/claude... • THE SYSTEM CARD, 193 pages. Page 151 is figure 8.4.A, the effort-by-score chart (medium 53.4, high 48.0, xhigh 43.6, max 48.0 — FrontierCode v1.1, run and scored by Cognition). Page 82 is the pilot-user "overthinking" note AND Anthropic's own hedge on it. Section 8.5 has FrontierBench, where xhigh wins — which is why we say the medium peak is one benchmark, not a law of the model — and the measured classifier refusal rates (5% vs Fable 5's 42%). Check our work: https://www.anthropic.com/claude-opus... • Official API pricing — $5 / $25 per Mtok, cache hits $0.50, batch 50% off, fast mode 2x (research preview, ~2.5x speed), US data residency 1.1x and stacking, web search $10/1,000 (as of Jul 24, 2026): https://platform.claude.com/docs/en/a... • What's new in Opus 5 — "unchanged from Claude Opus 4.8", "default user-facing responses and written deliverables run longer", thinking on by default, the 400 error above high effort, the 512-token cache minimum: https://platform.claude.com/docs/en/a... • Effort documentation — the five-rung ladder, the API default of "high", "start with xhigh for coding and agentic work", AND the very next sentence: "use low and medium liberally… run a fresh effort sweep on your evals": https://platform.claude.com/docs/en/b... • Model comparison — Opus 5 $5/$25 against Fable 5 $10/$50 and Sonnet 5: https://platform.claude.com/docs/en/a... • Artificial Analysis — the independent measurement. Index 60.69 (1st of 191), 100M output tokens at max effort against a 63M average, 60 tokens/sec against an average of 74, $3,835.51 to run the index, plus the full per-effort table above (measured Jul 24, 2026): https://artificialanalysis.ai/models/... • Consumer plans — Free / Pro $20 (or $17 billed annually) / Max from $100: https://www.claude.com/pricing • Fable 5 on your plan — the promotion that gave Pro subscribers 50% of their weekly limit on Fable 5 ended July 19, 2026; from July 20 it needs paid usage credits, with a one-time credit for eligible seats: https://support.claude.com/en/article... • Hacker News launch thread — both quoted comments, attributed on screen by handle: https://news.ycombinator.com/item?id=... —— CHAPTERS —— 0:00 Half the price of what? 1:00 What it is (and what we did not do) 2:07 The independent measurement 3:23 Their own words 4:55 What the pricing page doesn't tell you 6:23 The real price: page 151 8:12 Who it's actually for 9:27 Verdict What should we test next? Tell me in the comments.