New models ship, old ones don't just get switched off. claude-sonnet-4.6 is still available at $0.3 per million tokens, input and output, the same price as Sonnet 5. Here's who should stay on this revision and why.
What "stable revision" actually means
4.6 has run through months of real traffic. Its behavior on typical prompts is predictable: if a system worked six months ago and still works today, it'll keep working the same way tomorrow. This isn't about raw capability, it's about the absence of surprises.
Who pins the version, and why
Teams with tuned prompts. If your system prompt and few-shot examples are dialed in against a specific model's behavior, moving to a new revision means recalibrating all of it. Sometimes the result improves, sometimes it's just different, and either way that's a production risk you didn't ask for.
Regulated industries. Where you need reproducible answers for an audit or a compliance review, swapping the model mid-reporting-period raises questions you'd rather not answer after the fact.
Projects with a locked eval suite. If you have an eval set calibrated against 4.6, and the metrics confirm the error rate you need, don't change the model without rerunning that suite first.
Plain "don't fix what works." A legitimate reason on its own. Upgrading for the sake of upgrading costs testing time, and testing time is always scarce.
When it's worth moving anyway
If you're starting a new project, there's no reason to pin the old version, start with the current one. If 4.6 is hitting a quality ceiling on specific tasks, run the same prompts against Sonnet 5 and compare before committing to a full migration.
A working example
from anthropic import Anthropic
client = Anthropic(
api_key="sk_cdx_your_key_here",
base_url="https://api.claudexia.tech",
)
message = client.messages.create(
model="claude-sonnet-4.6",
max_tokens=1024,
messages=[{"role": "user", "content": "Check this SQL query for N+1 queries."}],
)
print(message.content[0].text)
The OpenAI Chat Completions format works on the same key and the same endpoint — swapping model to claude-sonnet-4.6 is all it takes.
Switching without the risk
Keep the model version in config, not hardcoded. That way an A/B test between 4.6 and 5 is a one-line change, not a redeploy. Run both in parallel on shadow traffic first, and judge by your own metric, not a gut feeling about which output "reads better."
Running both versions at once doesn't cost extra to manage: one key covers the whole lineup, and the dashboard breaks usage down by model, so you see exactly what went to 4.6 versus the 5 without stitching logs together by hand.
About the price
Sonnet 4.6 and Sonnet 5 are priced identically: $0.3 per million tokens on each side. Price isn't an argument for either one here. The decision is purely behavioral consistency versus a newer model.
In short
Sonnet 4.6 stays in the lineup not as a leftover, but as a real option for anyone who values predictability over novelty. The price doesn't change when you move to the 5, so test on your own data and migrate when you're ready, not because a version number went up.