price and performance, six models, one honest comparison · from the hedgehog test · see the Claude gallery, the OpenAI gallery, or the hedgehog side-by-side
context windows tied at the top: both lineups now ship a 1M-token context window with a 128K max output — Sonnet 5, Opus 4.8 and Fable 5 on one side, Luna, Terra and Sol on the other. the fight moved to price and tokens-per-answer instead.
—
input $ / M tokens
output $ / M tokens
Sonnet 5 and Terra land almost on top of each other on output price. Fable 5 is the most expensive token on the board by a wide margin — it's Anthropic's Mythos-class flagship, not a like-for-like match for anything OpenAI currently ships.
the one benchmark both vendors have published for every model that has a public score.
Fable 5 leads the field outright. Sol trails Opus 4.8 by about 4.6 points despite output pricing $5/M higher. Luna and Terra don't have published SWE-bench Pro scores — OpenAI is positioning them for everyday and business work, not head-to-head coding evals, and only frames them against Fable 5 on cost-adjusted general benchmarks like Agents' Last Exam.
set a token budget for one request and watch the six models re-sort by real cost.
—
total cost, cheapest → priciest
pricing and specs pulled July 2026: Anthropic's Claude pricing docs and OpenAI's GPT-5.6 announcement. SWE-bench Pro figures via Simon Willison's GPT-5.6 writeup and Anthropic/OpenAI's own release posts. list prices only — no cache discounts, batch pricing, or volume deals. numbers move fast; check the source links before you budget against them.
$ tail -f stew