artificial-analysis-xhigh
Auto-updating links to the Artificial Analysis Intelligence
Index chart at xhigh, high or medium effort instead of its default max, plus a summary of
the top models at each effort level.
Links
Bookmark these. Each redirects to the latest chart.
A model with no variant at that level stays as it is, and models with no effort setting are
untouched. Add #info to a link to see which models were swapped instead of redirecting. See
Add it to your bookmarks for the easy way to save them.
Why I made this
Artificial Analysis charts Claude, GPT and other models at their max effort by default, and
xhigh is cheaper and faster. Artificial Analysis's own per-task measurements show it for every model
in the default chart that has both variants (snapshot from 2026-10-24, max → xhigh):
| Model | Cost per task (API, then on a subscription) | Time per task | Output tokens per task | Intelligence Index |
|---|
| Claude Opus 5.5 | $5.98 → $3.46 (−42%)on Max 20x: $0.132 → $0.076 | −38% | −45% | 57.6 → 56.0 |
| Claude Sonnet 5.5 | $5.46 → $2.01 (−63%)on Max 20x: $0.121 → $0.044 | −55% | −62% | 56.0 → 51.9 |
| Claude Fable 5.1 | $7.63 → $5.98 (−22%)on Max 20x: $0.168 → $0.132 | −18% | −22% | 53.4 → 53.2 |
| Claude Haiku 5.5 | $0.213 → $0.124 (−42%)on Max 20x: $0.0047 → $0.0027 | −31% | −45% | 43.4 → 41.2 |
| GPT-6 Astra | $3.26 → $2.31 (−29%)on ChatGPT: $0.318 → $0.225 | −35% | −38% | 52.7 → 52.4 |
| GPT-6.1 Sol | $0.724 → $0.393 (−46%)on ChatGPT: $0.071 → $0.038 | −50% | −54% | 51.8 → 51.0 |
| GPT-6 Luna | $0.068 → $0.042 (−38%)on ChatGPT: $0.0066 → $0.0041 | −38% | −45% | 38.1 → 34.6 |
The second line in each cost cell is the same cost paid from a subscription: Artificial Analysis's
cost divided by how much more usage the plan gives than API list prices. The multipliers, 45.3x for
Claude Max 20x ($200 a month), 10.25x for any ChatGPT plan and 80x for SuperGrok plus X Premium+
(used by the chart bars below), were measured by
Paweł Huryn by burning real plans. He also
re-priced a full run of the whole index this way: Opus 5.5 at max costs $8,708 at API prices but
about $192 on Claude Max 20x, nearly a full $200 month, and about $90 at xhigh ($4,057 at API prices).
These are one person's estimates, measured on one model per vendor. SemiAnalysis
found that what the same $200 Claude plan
buys depends on the model and workload, from about 12x for Fable 5.1 to about 63x for Sonnet 5.5,
so the Fable 5.1 line in particular is probably too low by a factor of three or four.
xhigh uses 22% to 62% fewer output tokens, which is where the savings come from. The trade-off is a
slightly lower Intelligence Index score, 0.2 to 4.1 points lower in this snapshot. Anthropic's docs
describe max as "absolute maximum capability with no constraints on token spending".
The site does show xhigh scores, but you have to pick every model by hand in its model selector.
This repo builds the link that does that for me and keeps it current as the site's default models
change. The high and medium links do the same for lower effort levels.
Top models at each effort level
The ten highest-scoring models on the Intelligence Index at each effort level, regenerated every
3 hours. Each chart shows a model at the best setting it has at or below that level:
- A model with a variant at the level is shown at that variant.
- Otherwise it is shown at its highest setting below the level, with that setting in brackets. A
model that only has High, like Gemini 4 Argon, is on the max, xhigh and high charts.
- A model with no labelled effort setting, like Kimi K3.1 Pro, could be anything up to max, so it
appears only on the max chart, marked with a †.
- A model whose every setting is above the level is left off that chart.
Every column uses the color Artificial Analysis gives its lab (Anthropic's terracotta, Meta's blue
and so on), read from the same page as the scores. Because some lab colors are close, the lab is
also named in each column's tooltip and in the table views on the
summary page.
The thin dark bar beside a column is what one full run of the index costs on that lab's
subscription, using the multipliers described above, read off the dollar axis on the right. It is a
different measure from the score, so it has its own scale, and that scale is shared by all four
charts so you can see cost fall as effort drops: Opus 5.5 on Claude Max 20x is $192 at max, $90 at
xhigh, $48 at high and $36 at medium (2026-10-24). Only labs with a measured multiplier get a bar
(Anthropic, OpenAI and xAI), so Google, Meta, Kimi and the rest have none. The exact figures are in
the table views on the summary page.
Meta and Google are waiting on a measurement. The only Muse multipliers were taken on Muse Spark 1.3
before Muse Wonder launched, they range from about 11x to 51x for the $50 plan depending on the
source, and they do not carry over to Muse Wonder Full, which users report burns a plan's quota far
faster and which has no API price to divide. Nobody has published a Gemini plan multiplier yet.
Adding either is one line in PLANS in aa_costs.py once there is a number.
max (the Artificial Analysis default)
xhigh
high
medium
Best model one level down vs the next model
For each lab, its best model one effort level below its next model (Opus 5.5 at high against
Sonnet 5.5 at xhigh, and so on), to show whether the bigger model at a lower setting beats the
smaller one at a higher setting. The next model is drawn in a lighter shade of the lab's color. A
lab appears when it has two models with matching effort levels, which today means Anthropic and
OpenAI. Besides the subscription-cost bar, each column has a slate bar for staying in the flow:
Artificial Analysis's end-to-end response time (time to the first answer token, thinking included,
plus a 500-token answer at the model's speed), labelled in seconds. The table view on the summary
page gives the first-token time and output speed separately.
Coding Agent Index
A separate table on the summary page for real coding work, not just intelligence: Artificial Analysis's Coding Agent Index v2.0, a composite of FrontierCode v1.1, Terminal-Bench 4.0 and SWE-Atlas-QnA-AA. It shows the highlighted defaults plus Muse Spark variants, so the cheap-vs-best read is possible even when Spark is not highlighted: Muse Spark 1.3 (max) is the cheap Meta API fallback ($3.91/task), Muse Wonder Full / Wonder are the best Meta scores. Regenerated with the same 3-hour run from a third page; a failure there warns but never holds back the Intelligence charts.
How it works
The site picks its default 27 models server-side and accepts an override as
/models?models=slug1,slug2,.... Each effort level is its own slug
(claude-opus-5-5 is Max, claude-opus-5-5-xhigh is Xhigh).
aa_xhigh.py fetches the page and reads the data the site embeds in it:
- The default selection (
chartDefaultSelected), in the site's order.
- The full variant list, each entry with a
releaseSlug and an effort label in its name.
- For every default model labelled above the target level, it finds the sibling with the same
releaseSlug, the same qualifier (e.g. "Default Fallback") and the target label, and
substitutes it. It never raises a model to a higher level.
A model above the target with no sibling at it (for example GLM-5.3, which only has Max and Low
at the time of writing) stays as it is. Use --unmatched drop to remove such models from the
chart instead.
The charts use a second page: the site's /leaderboards/models, which lists an Intelligence Index,
the lab and the lab's color for each variant. aa_charts.py groups the variants of each model by
release, applies the rules above to pick one per model for each chart, and draws the SVGs.
The cost bars need the cost of a full index run for each model variant, which only the comparison
pages carry, two models per page. aa_costs.py fetches the variants it needs (every column of a
lab with a plan multiplier), caches them in docs/costs.json and refreshes an entry once it is a
week old, so most runs make no extra requests and a failed refresh keeps the old value.
The coding table needs a third page: the site's /agents/coding-agents, which carries the Coding
Agent Index, mean API cost per task and mean wall time per task for each harness. aa_coding.py
reads the data the site embeds in it, select_for_summary() keeps the highlighted defaults plus
Muse Spark variants, and the summary page shows Index, cost/task and time/task with its own
coding_changed_at in docs/charts.json.
I looked at placing unlabelled models on the lower charts by speed and decided against it: time per
task does not line up with effort across labs. In the 2026-10-24 data, Claude Haiku 5.5 at max
takes 430 s per task while Qwen3.8 27B at xhigh takes 1,227 s, and Kimi K3.1 Pro's 618 s sits
inside the max, xhigh and high ranges at once.
The scheduled workflow re-runs everything every 3 hours and commits docs/ only when something
changes.
Setup
Needs a public repo (Pages on private repos needs a paid plan).
gh repo create artificial-analysis-xhigh --public --source=. --push
gh api -X POST repos/{owner}/artificial-analysis-xhigh/pages -f 'source[branch]=main' -f 'source[path]=/docs'
gh workflow run update-link
Or in the web UI: Settings, Pages, deploy from branch main, folder /docs.
The site then lives at https://<your-username>.github.io/artificial-analysis-xhigh/. Each link
folder also has link.txt (the raw link) and latest.json (details).
Add it to your bookmarks
The links redirect, so pressing your browser's "bookmark this page" shortcut on one saves the long
chart URL it lands on, which goes stale when the default models change. Instead:
- Drag (easiest): open the bookmark page
and drag the links onto your bookmarks bar.
- Paste the URL: copy a link from the table above, then in Chrome press Cmd/Ctrl+Shift+O, use
the menu at the top right, choose Add new bookmark and paste. In Firefox, Cmd/Ctrl+Shift+O,
Organize, New Bookmark. Safari has no add-by-URL, so use the drag links.
Use it locally
No dependencies. Developed on Python 3.12.
python3 aa_xhigh.py # print the xhigh link
python3 aa_xhigh.py --level high # print the high link (xhigh, high, medium, low or minimal)
python3 aa_xhigh.py --explain # also show what was swapped
python3 aa_xhigh.py --open # open it in your browser
python3 aa_xhigh.py --unmatched drop
python3 aa_xhigh.py --write docs # regenerate the whole Pages site: links and charts
Tests
python3 -m unittest discover -s tests -v
AA_LIVE=1 python3 -m unittest discover -s tests -k Live -v # also hit the real site
When it breaks
The scripts read data embedded in the site's pages, which is not a public API. If a page changes
shape the script exits with an error (and checks that every slug it emits exists in the site's own
model list) rather than publishing something wrong. The links and the charts come from different
pages, so one failing does not hold back the other. The workflow run turns red, GitHub emails
you, and docs/ keeps the last good output.
GitHub disables scheduled workflows after 60 days without repo activity, so the workflow adds an
empty keepalive commit when the last commit is older than 50 days.
