claude-models-value-analysis

What does each recent Claude model actually cost, at each effort level — and is paying for more effort worth it?

A normalized cost × model × effort matrix for the current Claude family, fused from independent public measurements, plus a self-contained interactive report with a live value-scoring model.

Open index.html in a browser — fully self-contained (no server, no external assets), light/dark aware, with zoomable charts and live tier tuning.

Live page: https://claude-models.agensse.com/. It is served from main: the server checks main every 5 minutes and copies a whitelist of files: index.html, claude-models-head-to-head.html, raw-data.csv, robots.txt, sitemap.xml, llms.txt, favicon.svg, favicon.png, og-image.png and the IndexNow key file (b3573dbc1da690e66e9ef05b081b7abe.txt). Nothing is built on the server, so index.html must be committed built, and a new file served at the root must be added to that whitelist. Exploratory blocks (value score, window tuner, full method) are collapsed by default; click to expand.

Publishing a fork: in site/config.py, set SITE_URL to the fork's address (every absolute URL, the sitemap, robots.txt, llms.txt and the domain shown in the share image derive from it) and REPO_URL to its repository; set BING_SITE_VERIFICATION and INDEXNOW_KEY to "" (they prove this site's ownership to Bing and IndexNow; an empty value leaves out the tag and the key file) or to your own values. Delete b3573dbc1da690e66e9ef05b081b7abe.txt, run python3 site/build.py, and redraw the share image with the new domain (site/assets/og_image.py, needs Pillow: uv run --no-project --with pillow python site/assets/og_image.py). The author credit and the GitHub links in gen/body.html point to this repository (attribution, CC BY 4.0). Then e.g. GitHub Pages: Settings → Pages → Source: Deploy from a branch → main / root.

Models covered: Fable 5.2, Fable 5.1, Fable 5, Opus 5.5, Opus 5, Opus 4.8, Opus 4.7, Sonnet 5.5, Sonnet 5, Sonnet 4.6, Haiku 5.5, Haiku 4.5. No reference couple since 30 Sep 2026: cost is a multiple of the cheapest couple, quality the expected score on the benchmark panel (the page was relative to Opus 5 @high up to tag v2026.09.29b, to Opus 4.8 @medium up to v2026.09.23; the git tags reproduce each earlier scale). Fable 5, Opus 4.8, Opus 4.7 and Sonnet 4.6 stay in the data but are hidden on the page by default (an Older models switch brings them back).

---

What the report shows

1. Consolidated landscape — one curve per model, one point per configuration (an effort level from minimal to max, thinking off, thinking on), cost (a multiple of the cheapest couple) × quality (the latent scale, labelled in expected panel score). Optional tier bands shade the four usage tiers on this chart and on the Pareto view. Optional ovals draw each couple's 16–84 % interval. A Pareto view isolates the non-dominated couples, fits a price trend over every couple (what a given quality typically costs), and says how many times cheaper or dearer than the trend each couple is. A tier picker (with live target/σ sliders) turns the frontier into a decision: the best-value (model, effort) for four task-complexity levels, plus a crowned overall pick (the couple furthest below the trend).
2. Normalized matrix — cost per model × configuration (effort level, thinking off) as a multiple of the cheapest couple, sorted by expected score, each cell with its interval.
3. Sources — every source that measured ≥2 couples on the same task, with its verified configuration and the couples it links (names are clickable).
4. Method — how the numbers and their intervals are built; the full methodology is in docs/METHODOLOGY.md.

How the numbers are built

Scores from different benchmarks do not share a scale, and raw dollars from different tasks are not comparable, so the report fuses the measurements with a latent-quality model (model/lqm.py). In short:

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗