Stable lanes explained: what stable-* buys and what it costs
A stable-* id is a dedicated lane to the same upstream model, served on our own key instead of the shared pool. Four times the multiplier on Claude, and no pool failover.
Same model, different path to it
A stable-* id names the same upstream model as its pool sibling. stable-claude-opus-5 and claude-opus-5 both carry a 1,000,000-token context, both emit up to 128,000 tokens, both support tools and extended thinking, and both answer on /v1/messages. The weights are the same; the id prefix describes how the request gets there.
The difference is the credential. A pool request is served from a shared account pool contributed by many accounts; a stable request is served on a key we hold directly with the supplier. That is the whole mechanism — the prefix is not a quality tier, a fine-tune, or a different context window.
Some ids exist only as a stable lane. stable-claude-fable-5-1, stable-claude-fable-5, stable-claude-haiku-4-5 and stable-gpt-6-astra have no pool sibling to compare against, so for those the lane is not a choice between two routes — it is the only route.
# Pool lane — 0.1x
curl https://token-share.app/v1/messages \
-H "x-api-key: $TOKEN_SHARE_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "content-type: application/json" \
-d '{
"model": "claude-opus-5",
"max_tokens": 512,
"messages": [{"role": "user", "content": "Summarize this changelog."}]
}'
# Dedicated lane — 0.4x, same weights, same endpoint
curl https://token-share.app/v1/messages \
-H "x-api-key: $TOKEN_SHARE_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "content-type: application/json" \
-d '{
"model": "stable-claude-opus-5",
"max_tokens": 512,
"messages": [{"role": "user", "content": "Summarize this changelog."}]
}'What it costs: four times the multiplier on Claude
Every Claude stable lane is billed at 0.4×, against 0.1× for the pool ids. claude-opus-5 and stable-claude-opus-5 share the same official rates — $5 per million input, $25 output — so the stable lane costs four times as much for the same tokens. stable-gpt-6-astra is 0.2×, and its pool sibling does not exist to compare against.
One stable lane is priced on lower official rates than its pool sibling. stable-claude-sonnet-5 lists $2 input and $10 output against claude-sonnet-5's $3 and $15, so at 0.4× versus 0.1× the effective gap is smaller than four times — but still a gap. Multiply the official rate by the multiplier for each id rather than assuming the ratio.
Stable lanes cannot be paid for with earned credit. Earned credit exists because someone contributed an account to the shared pool, and this lane is bought from an outside supplier, so a request that would draw on earned balance is refused with 402 earned_credit_not_accepted before anything is spent.
What it gives up: no pool failover
A pool request that fails on one machine can be retried against another cluster with a different set of accounts. A stable request does not enter that walk at all. The point of the lane is one known upstream on our own key, so quietly serving it from the shared pool would hand you the thing you paid to avoid.
The practical consequence is that a stable failure is returned as a failure. A 429 on this lane means the channel is at its own rate ceiling — not a signal for the pool to absorb the overflow — and comes back as stable_channel_rate_limited or stable_channel_busy. Availability problems return 503 stable_channel_unavailable, timeouts 504 stable_channel_timeout.
That makes retry handling your side's job in a way it is not on the pool lanes. Deciding between the two is a question about your workload: what a 429 costs you when it lands, what four times the token price costs you when it does not, and whether a pool sibling exists for the id at all. Run both against the same traffic and compare the bill and the error rate you actually observe.
The reverse direction does exist, and it is worth not confusing with this one. A pool request that fails is re-served on its stable sibling where one is live, billed at the stable rate, with the trace recording the id you asked for and why the pool could not serve it. That runs from the cheaper lane to the dearer one, which is why it can be done without taking away anything you paid for; going the other way would.