tokensitterA conversation prior
The relay asks an external classifier once, on the redacted first message. A timeout or a missing answer keeps the requested model. Only premium Anthropic requests are ever rerouted.
Features/Smart routing
Smart routing
The right model.
For the right work.
Choose a model at each step of API-key work. Keep the requested model when the request needs it, and account for cache costs before switching.
Add to Claude CodeOpenAI Codex: coming soon
API relay · opt-in per keyEvidence snapshot: 27 Oct 2026 · per-lever live deployment not verified
Inside the workflow
Classify the conversationA redacted first message sets a light, mid or strong ceiling.
Choose each stepStep rules select Haiku, Sonnet or the requested premium model.
Escalate when neededErrors send the request back to the premium model.
Transparent by designEvery claim needs a baseline
The technique
What changes.
And what does not.
Availability differs between the API relay and the local plugin. Whether each lever is deployed on the live relay has not been independently checked.
Rules at each step
New human instructions favor the strong tier; read-only tool results favor light; command and edit results favor mid. Balanced mode can retain mid on edit-task reads. The conversation prior caps the final tier.
Cache-aware escalation
A downgrade away from a warm model is taken only when the projected saving repays a cold cache write. A failed tool result, or an explicit escalation, returns to the requested model; repeated escalations lock it.
The boundary. API keys through the relay only. Subscription logins are never proxied. Direct-API lab results do not establish savings on Claude Code Pro or Max sessions.
Quality checks
Fewer tokens is only
half the question.
The bench checks answers against exact values, path-and-line answers, document cells and restored pristine test files; no judge model decides pass or fail. Short-task failures included wrong file and line answers. Every long run passed in both arms, but with so few runs the lower end of the pass-rate interval is wide, so this is not proof of equal quality. The pass checks were hardened after the runs, and the stored run records were not re-graded under the hardened checks, so the pass counts are as graded at the time.
Checks that would strengthen this evidence. They are not yet reported for this feature:
Task correctness
Check each routed task against an objective test, not against how plausible its answer looks.
Escalation cost
Count steps that returned to the requested model and the cost they added.
Cache-switch cost
Include cold cache writes caused by switching models.
Boundary cases
Include ambiguous tasks and inputs where routing should stay on the requested model.
How we compare
Same task.
A stated baseline.
Short tasks: Opus 5.5 on every step with no prompt cache. Long runs: prompt-cached Opus 5.5. The harness is a synthetic tool-using agent calling the API directly, with the relay's routing logic imported rather than run through the live relay. The classifier's own cost is excluded and the cache-switching clock was simulated.
Total baseline task cost − total routed task cost
We aim to report alongside: net API cost per completed task · task pass rate
We aim to report alongside: net API cost per completed task · task pass rate
Figures and evidence
The result.
With its boundaries.
Results are copied from our lab reports (snapshot 27 Oct 2026) and were not re-run for this website.
Lab-76.6%Short tasks (9 tasks, no cache read): total billed cost per task, routing each step vs Opus for every step (CI -87.3 to -59.9, 27 pairs)
Lab · No proven saving-3.7%Long agent runs (21.8 turns with cache, 24.1 routed; one tool call per turn), routing each step plus cache vs a prompt-cached Opus agent: total billed cost per run (CI -18.7 to +9.5, includes 0, 15 pairs, passed 15/15 both ways)
Lab26/27Tasks passed with routing per step (Opus every step: 27/27)
Read this alongside the figures. The long-run cost intervals include zero. Do not generalize the short-task headline to long edit-heavy work. No end-to-end Claude Code relay routing benefit is established; no competitor comparison was run.
| Experiment | Workload and date | Result | Limits | Report |
|---|---|---|---|---|
| Short-task run | 162 runs: 9 tasks x 3 repetitions x 6 arms, no run errored 26 Oct 2026 | -76.6% Short tasks (9 tasks, no cache read): total billed cost per task, routing each step vs Opus for every step (CI -87.3 to -59.9, 27 pairs) 26/27 Tasks passed with routing per step (Opus every step: 27/27) $0.0160 Total billed cost per task with routing per step (Opus every step: $0.0685) | Route: 9 tasks, 27 runs per arm, direct API. Routing per step passed 26/27 against 27/27 for Opus every step and that difference is not significant, so 27 runs cannot rule out a few points of quality loss. The first run's cost saving was larger than the rerun's (-76.6% against -69.5%, intervals overlap), so read it as a range. Route: the one miss with routing per step was the long file audit, where the answer named the wrong file. The rerun passed 27/27, but its arm with cache, compress and read less missed the same task once. |
|
| Short-task rerun | 81 runs: 9 tasks x 3 repetitions x 3 arms 26 Oct 2026 | -69.5% Rerun, same tasks: total billed cost per task, routing each step vs Opus (CI -83.9 to -46.7, passed 27/27 vs 27/27) | The rerun passed every task; read the two runs as a range. |
|
| Other routing arms, same short tasks | 162 runs: 9 tasks x 3 repetitions x 6 arms, no run errored 26 Oct 2026 | -40.3% Routing the whole task to one model (routing-savings) vs Opus, total billed cost per task (CI -49.0 to -29.2, passed 26/27) -47.1% Routing read-only tasks to Haiku (routing-readonly) vs Opus, total billed cost per task (CI -70.4 to -8.7, passed 26/27) | Compared with Opus on every step; the pass counts are in each label. |
|
| Long runs, one tool call per turn | 15 pairs per comparison: 5 tasks x 3 repetitions x 2 arms (cache baseline, routing per step plus cache), run twice (one tool call per turn, and parallel tool calls allowed); all passed in both arms 27 Oct 2026 | -3.7% Long agent runs (21.8 turns with cache, 24.1 routed; one tool call per turn), routing each step plus cache vs a prompt-cached Opus agent: total billed cost per run (CI -18.7 to +9.5, includes 0, 15 pairs, passed 15/15 both ways) No proven saving +5.1% Long runs, one tool call per turn, without the read-only audit task (CI -3.9 to +17.0, 12 pairs): no saving No proven saving | Route, long runs: only 5 synthetic coding tasks (15 pairs per comparison, one date). The intervals are wide, so a small saving or a small extra cost cannot be ruled out. Route, long runs: contexts reached only about 19k tokens, so contexts of 100k tokens and more were not tested. |
|
| Long runs, parallel tool calls | 15 pairs per comparison: 5 tasks x 3 repetitions x 2 arms (cache baseline, routing per step plus cache), run twice (one tool call per turn, and parallel tool calls allowed); all passed in both arms 27 Oct 2026 | -7.0% Long agent runs with parallel tool calls allowed (about 6 turns), same comparison (CI -18.1 to +0.7, includes 0, 15 pairs, passed 15/15 both ways) No proven saving -0.4% Long runs, parallel tool calls, without the read-only audit task (CI -4.5 to +3.4, 12 pairs): no saving No proven saving -96.6% Long runs, the read-only audit task alone: the only large saving (3 pairs per run set) | Route, long runs: only the read-only audit task saved a lot (-96.6%, 3 pairs per run set). Without it the long runs show no saving. |
|
Lab results at Anthropic API list prices on fixed tasks show what a lever did on those tasks, not what it will do on your bill. For plugin levers on Pro or Max, list-price values describe usage economics, not the subscription invoice. Relay levers apply to API keys only.
Read the measurement method and evidence limitsQuestions, answered
Does routing work with Pro or Max logins?
No. Relay routing uses your API key. Local plugin features are separate; Tokensitter never proxies a subscription login.
What did longer agent runs show?
Against prompt-cached Opus, the long-run cost intervals include zero. These experiments do not demonstrate a saving for edit-heavy long agent work.
Keep exploring
A little more care.
A clearer picture.
Add Tokensitter to Claude Code and see where your tokens go.
Add to Claude CodeOpenAI Codex: coming soon