While Opus 5 holds the flagship spot and Opus 4.8 covers the fallback role, two earlier revisions, 4.6 and 4.7, keep running unchanged in the Claudexia lineup. We're not in a hurry to switch them off, some traffic is deliberately tied to them.
Why keep older revisions at all
The reason is simple: some pipelines were built and validated against a specific revision, and any silent update under the hood is a potential regression. If a prompt, tool schema, and response parsing are tuned to Opus 4.6's behavior, swapping in a newer revision without testing isn't an upgrade, it's an experiment running in production.
Claudexia doesn't force a migration. Both revisions are reachable by their own ids exactly like the flagship.
Who pins to a specific revision, and why
- Teams with a frozen eval dataset. If you have a golden case set and a pass-rate metric tied to a specific revision, switching models breaks comparability of the numbers over time. Pinning fixes the baseline you're comparing against.
- Regulatory and audit requirements. Some processes need to document exactly which model version handled a given request, especially when the output feeds into reporting or a legally significant document.
- Long-running agentic scenarios. If an agent is built around a specific reasoning style and output format, a silent mid-flight revision swap can change behavior partway through a scenario that spans hours or days.
- Gradual rollout. Some traffic is intentionally kept on an older revision as a control group while a newer one gets exercised on the rest.
The difference between 4.6 and 4.7
Anthropic doesn't publish a line-by-line changelog between minor Opus revisions, so specific behavioral differences are better checked against your own cases than taken on faith from any blog post, including this one. The practical approach: if you already have an eval dataset, run it against both revisions and compare the numbers, that's more reliable than any writeup.
How to call a specific revision
Only model changes, everything else works the same as any model on Claudexia:
curl https://api.claudexia.tech/v1/messages \
-H "x-api-key: sk_cdx_..." \
-H "anthropic-version: 2023-06-01" \
-H "content-type: application/json" \
-d '{
"model": "claude-opus-4.6",
"max_tokens": 1024,
"messages": [{"role": "user", "content": "Run this case from the golden dataset"}]
}'
The same call against revision 4.7:
from openai import OpenAI
client = OpenAI(api_key="sk_cdx_...", base_url="https://api.claudexia.tech/v1")
resp = client.chat.completions.create(
model="claude-opus-4.7",
messages=[{"role": "user", "content": "Run this case from the golden dataset"}],
)
print(resp.choices[0].message.content)
Both ids live on the same endpoint, both API formats work identically, one key covers either.
The risk of pinning long-term
Pinning removes the risk of a silent regression today, but adds a different kind of risk: sooner or later any revision can be marked deprecated by Anthropic itself. If you've hardcoded a model id, keep a standing task to check revision status periodically and have a migration plan ready, rather than finding out about a shutdown from production errors.
The price doesn't depend on the revision
Through Claudexia, 4.6, 4.7, 4.8, and 5 all cost the same: $0.4 per million input tokens and $0.4 per million output tokens. Pinning to an older revision doesn't save money or cost extra, the difference is in behavior and stability, not the bill. Compare numbers for your own load with the calculator on the site, and per-key spend is broken out separately in the dashboard.
In short
We keep Opus 4.6 and 4.7 running for teams that need reproducibility: frozen evals, audits, long agentic scenarios, gradual rollouts. Check behavioral differences between revisions against your own cases. Every Opus revision costs the same, so pinning is a decision about stability, not about saving money.