Finishes the long runs
The difference is not the benchmark score, it is whether the run finishes. In public evaluations, models that score alike diverge sharply past twenty steps — some time out and give up, some reach the end. This tier buys the latter.
1M
Context window
25
turns / 1,000 crystals (estimate)
20×
Cost vs the cheapest model
Good for
- Cross-module refactors
- Autonomous runs of dozens of steps
- Bugs that are hard to reproduce
Not for
- Simple edits — it wastes credits
Rates
Published price per million tokens — what your crystals are charged at. It is the upstream official list price; check it yourself.
Cache read$0.2600$0.2600
Input$1.4000$1.4000
Output$4.4000$4.4000
Turns per the same amount of crystals.
fast
pro
max
Turn counts are derived from the published rates, assuming a typical agent request of 30k input and 2k output.
How to use it. Put it in an environment variable to call it by name, bypassing the alias default.
Environment variables
export ANTHROPIC_BASE_URL="https://api.pixcode.ai/anthropic"
export ANTHROPIC_AUTH_TOKEN="sk-px-..."
export ANTHROPIC_MODEL="pixcode-max"Other tools are covered in the docs.
Which plans include it.
Pro $49/moMax $129/mo