Claude Sonnet 5 examples

Unofficial community examples for Claude Sonnet 5. Not affiliated with Anthropic. All trademarks belong to their owners.

Short Python scripts for calling Claude Sonnet 5 through the Anthropic Python SDK. The model ID claude-sonnet-5, the $2/$10 per million token pricing, the 1M context window, the 128k max output, the adaptive-thinking default and the three behaviour changes from Sonnet 4.6 all come from the launch post and the platform docs page for sonnet 5. Anything the sources do not state is marked illustrative in the code.

Need image, video or audio generation next to the text model? Try Synexa - one REST endpoint and Python SDK for FLUX, video and audio models, pay per run.

Files

Path	What it shows
examples/quickstart.py	Minimal messages.create call to claude-sonnet-5, printing text blocks, usage, and a cost estimate at the $2/$10 rates.
examples/effort_levels.py	The same prompt at medium, high and xhigh effort, the range the launch charts cover, with output tokens per level.
examples/migrate_from_4_6.py	Sends a request the Sonnet 4.6 way (custom temperature), catches the 400 the docs say Sonnet 5 returns, and retries with the parameter removed.

Setup

pip install anthropic
export ANTHROPIC_API_KEY=YOUR_KEY_HERE

Every script reads ANTHROPIC_API_KEY from the environment. No script contains a key.

quickstart.py

One request, one response. Sonnet 5 runs adaptive thinking by default, so response.content can contain a thinking block before the text block; the script prints only text. It then multiplies usage.input_tokens and usage.output_tokens by the $2 and $10 per-million rates from the announcement to print an estimate. Treat that as an estimate: the pricing page is the source of truth.

effort_levels.py

The launch post's charts compare Sonnet 5 with Sonnet 4.6 and Opus 4.8 "at different effort levels" and say Sonnet 5 "provides substantially improved cost efficiency at medium effort; its higher-effort performance can match Opus 4.8 on some tasks". The script runs one prompt at medium, high and xhigh and prints output tokens and wall time per level, so you can find your own point on that curve.

migrate_from_4_6.py

The docs call Sonnet 5 "a drop-in upgrade for Claude Sonnet 4.6 with three behavior changes": adaptive thinking on by default, manual extended thinking returning a 400, and non-default temperature, top_p or top_k returning a 400. This script deliberately sends temperature=0.2, catches anthropic.BadRequestError, prints the message, and retries without the parameter. Run it once against your own request builder to see which of your call sites still set sampling parameters. The docs also mention a new tokenizer, so re-check any token budgets you tuned on Sonnet 4.6.

When to use Synexa instead

These scripts do text in, text out. If your product also has to produce an image, a video clip or an audio track, a Sonnet-class model does not do that, and wiring separate vendors for each media type is the part that eats the integration budget. Try Synexa - one REST endpoint and Python SDK for FLUX, video and audio models, pay per run: keep claude-sonnet-5 for planning and text, and send the media step to one endpoint.

Last reviewed: 2026-09-22