llm-price-watch

Tracks what changed in Anthropic and GitHub Copilot pricing, not just what it costs today.

→ zrrbite.github.io/llm-price-watch · Atom feed

Vendors publish what a model costs right now. They do not publish a changelog. The thing you actually miss is a change — a price cut, a promotional rate, a scheduled increase that quietly got cancelled.

The case that prompted this: on 2026-09-01 Anthropic made Claude Sonnet 5's $2/$10 introductory pricing permanent, and the increase to $3/$15 announced for that date did not happen. That fact is not in the pricing table. It is one sentence in a callout beside it. Read only the table and you learn nothing.

So this watches both the numbers and the prose around them.

Every number comes from the vendor's own published source. No aggregators, no estimates, and no model in the data path — the parsers are deterministic, because an extractor that occasionally invents a price is worse than no tool.

Promotions hide in a footnote marker on the model name — GPT-5.6 Sol[^gpt-56-sol-promo] — whose text lives on the content page and whose model names come from a Liquid variables file. That marker is both the only machine-readable "this is temporary, and it ends on this date" signal either vendor publishes, and a trap: it is part of the model name, so a key that keeps it reports the same model as removed-then-added every time a promotion starts or ends.

Copilot's published pages render those tables from Liquid templates and contain no numbers themselves. The YAML data files behind them are the real source — structured, and version-controlled in a public repo, which is also where the backfilled history comes from.

Three sizes, because for an agent the cost of reading is part of the answer. Measured on the live data:

models.tsv is one line per usable model — no repeated keys, no prose, retired models omitted — with a class column carrying the vendor's own capability tier so price can be judged against what you get rather than in the abstract:

model	vendor	class	in	out	blended	offer	ends	retires
Claude Sonnet 5	Anthropic	Versatile	2	10	4	none	-	-
GPT-5.6 Sol (≤ 272K)	Copilot	Powerful	2	10	4	LAPSED	2026-09-03	-

If your consumer runs a script, size does not matter — only what it prints reaches the context. Reach for the small endpoints when the file itself is going into a prompt.

skill/llm-price-check/ is an agent skill that warns you before you commit to a model — promotion about to end, price already past its end date, model retiring, dearer than its class — and stays quiet when there is nothing to say.

It is one folder in the portable SKILL.md format, which Claude Code and GitHub Copilot both read from the same personal location, so one install serves both.

macOS and Linux:

curl -sSL https://raw.githubusercontent.com/zrrbite/llm-price-watch/main/skill/install.sh | bash

Windows:

irm https://raw.githubusercontent.com/zrrbite/llm-price-watch/main/skill/install.ps1 | iex

Both default to ~/.claude/skills (%USERPROFILE%\.claude\skills), find a working Python, copy the two files, and run a real check so you know whether it works before you rely on it in a session. Pass a different skills root as the first argument — ~/.copilot/skills, ~/.agents/skills, or a project's .github/skills to commit it with the code.

They copy rather than symlink, so the skill keeps working on a machine with no checkout of this repo — which is the machine you usually want it on. The Windows script prefers the py launcher and skips the Microsoft Store python stub, which otherwise exits without running anything.

Or by hand — it is two files in a folder named llm-price-check:

~/.claude/skills/llm-price-check/
├── SKILL.md
└── check.py

The skill works two ways so it degrades rather than failing: it runs check.py when the agent can execute scripts, and otherwise tells the agent to fetch brief.txt (~180 tokens) and reason over that. Note that when a script runs, endpoint size is irrelevant — only what the script prints reaches the model's context.

Ordered so the actionable things come first and the raw data last.

Punching above its price — models costing less than the typical model of the tier below them. This is the question a price table cannot answer: not what is cheapest, but what is underpriced for what it is. Right now GPT-5.6 Sol is a Powerful-tier model at $4.00 blended, 11% under the median Versatile model, and it undercuts 9 of the 17 of them outright. Promotional prices are marked, because an anomaly with an expiry date is a different proposition from a permanent one.

On offer — promotions the vendors have written down, plus price cuts still in force, soonest deadline first. Each cut is re-checked against the current price, so a cut that was later reversed is not reported as a deal.

Retiring soon — models with a retirement date inside 45 days, with GitHub's suggested replacement.

Good for this, right now — hand-written picks from data/picks.yml. This is the one place opinion lives. Prices are not stored there; they are looked up on every build, and a pick whose cost has drifted more than 10% since it was written is flagged on the page rather than left to quietly mislead.

Cheapest per tier, then What changed, then the full tables.

A scheduled Action fetches each source, parses it, compares it against the last committed snapshot, and records what moved. Snapshots are committed, so the repo's own git history is the price history.

Three decisions do the real work:

Diff on values, not on text. A textual diff fires on every wording tweak and reflow. You would be ignoring the notifications within a month — at which point the tool is worse than nothing, because you believe something is watching while it isn't. Rows are compared on parsed numbers and set membership.

Diff the prose notes too, separately. This is what catches the Sonnet 5 case. Callouts are extracted and diffed as text, reported as advisories, and kept apart from price deltas so the price feed stays trustworthy.

A sanity gate. When a vendor restructures a page, a naive parser returns nothing and the differ faithfully reports every model as deleted. If a parse yields under half the previous row count, the run refuses to commit and raises a parser broken alert instead. That failure is guaranteed to happen eventually; it must not look like news.

Email stays inert until configured: with no SMTP_HOST the step logs a line and the run stays green. To switch it on, add repository secrets SMTP_HOST, SMTP_TO, and as needed SMTP_PORT (default 587), SMTP_USER, SMTP_PASSWORD, SMTP_FROM.

python3 -m venv .venv && .venv/bin/pip install pyyaml

.venv/bin/python -m unittest discover -s tests   # 94 tests, no network
.venv/bin/python src/update.py --dry-run         # report, write nothing
.venv/bin/python src/update.py --cache           # reuse fetches while iterating
GITHUB_TOKEN=$(gh auth token) .venv/bin/python src/backfill.py

Fixtures under tests/fixtures/ are the real vendor payloads as of 2026-09-02, so the parser tests fail if a vendor changes shape in a way the code does not handle. That is intentional — it is the early warning.

Design notes: docs/specs/2026-09-02-llm-price-watch-design.md.

• Anthropic publishes no historical archive, so its record starts when this tool first ran. Copilot's is backfilled from github/docs and reaches back to 2025.
• Two vendors only. OpenAI and Google were considered and cut: neither publishes a clean machine-readable source, and mixing verified numbers with aggregator-sourced ones under one roof would undermine both.
• List price is not job cost. A model at a fifth the price that needs three attempts and emits more output tokens is more expensive. This tracks prices, not value.