Routing intelligence

Every turn has a tradeoff.

One surface for cache behavior, route decisions, model quality, tokens, latency, spend, savings and benchmark evidence — so “auto” never means opaque.

Cache awareModel scoringCost attributionDecision traceBenchmark evidence

Routed turns

14,286

2,041 / day

Cache hit

38.4%

+4.7 pts

Model score

92.1

quality floor held

Actual cost

$18.42

$1.29 / 1K turns

Saved

$31.67

63.2% vs baseline

P95 latency

1.18s

−22.8%

Turns over cost.

The line that matters: cumulative spend as work compounds. Compare the routed path against an explicit “always frontier” baseline.

Cumulative cost / turns

7 day sample

Arcana routedFrontier baseline

Why that model?

Expose the decision path instead of returning a model name without context. Cache, task fit, candidates, score and fallback policy remain inspectable.

Decision trace

sample

Cache lookupsemantic + exact request key

MISS

Task profilecoding · repo context · tool use

0.87 HARD

Candidate set4 healthy models satisfy policy

4 / 9

Score + selectquality floor first, then cost

OPUS 5 · 94.7

Fallback planGemini 3.7 Flash → GPT-5.6 Sol

ARMED

Candidate

Score

Quality

Task fit

Reliability

Latency

Est. cost

Claude Opus 5Anthropic · selected

94.7

96strong

98coding

99.1%7d

1.21sp50

$0.021turn

GPT-5.6 SolOpenAI · frontier

93.8

97strong

96coding

98.7%7d

1.34sp50

$0.027turn

Gemini 3.7 FlashGoogle · fast alternate

91.5

91good

94coding

99.4%7d

0.62sp50

$0.006turn

DeepSeek V4 ProDeepSeek · cost alternate

88.9

89good

90coding

98.9%7d

0.71sp50

$0.004turn

Cache is a first-class metric.

Separate hits, misses, writes and cached tokens. Savings should be attributable, not buried inside a blended cost number.

Cache effectiveness

all routes

38.4%hit rate

Hits / misses5,486 / 8,800

Cached tokens4.20M

Cache write tokens812K

Cost avoided$8.91

Tokens by model

input · output · cached · reasoning

Claude Opus 5

4.82M

Gemini 3.7

3.94M

GPT-5.6 Sol

2.71M

DeepSeek V4

1.86M

Qwen 3.8

1.22M

InputOutputCachedReasoning

Model economics, side by side.

Route share alone is not enough. Put score, tokens, cache, latency, actual spend and counterfactual savings on the same row.

Model

Score

Route share

Tokens

Cache

Cost

Saved

Claude Opus 5Anthropic

94.7

31.8%4,542 turns

4.82M33.1%

41.2%1.99M tok

$8.7447.4%

$9.0650.9%

Gemini 3.7 FlashGoogle

91.5

28.4%4,057 turns

3.94M27.1%

35.4%1.39M tok

$3.1216.9%

$8.7773.8%

GPT-5.6 SolOpenAI

93.8

18.7%2,671 turns

2.71M18.6%

37.9%1.03M tok

$4.6225.1%

$5.1152.5%

DeepSeek V4 ProDeepSeek

88.9

13.9%1,986 turns

1.86M12.8%

29.8%554K tok

$1.216.6%

$6.5284.3%

Qwen 3.8 MaxAlibaba

89.6

7.2%1,030 turns

1.22M8.4%

20.7%252K tok

$0.734.0%

$2.2175.2%

Benchmarks belong next to routing.

Quality and cost are a joint decision. Keep external evidence and production evidence visible beside the model score instead of hiding methodology elsewhere.

Evidence surface

Opus 5

GPT-5.6

Gemini 3.7

DeepSeek V4

Qwen 3.8

Coding indexnormalized evidence

Reasoning indexnormalized evidence

Agentic indextools + long horizon

Production successexample 7d telemetry

99.1%

98.7%

99.4%

98.9%

98.8%

Blended cost efficiencyquality-adjusted $/turn

Data status: benchmark and production values on this initial static page are illustrative UI fixtures, not verified Arcana benchmark claims. Replace them only from versioned benchmark artifacts or production telemetry.

The routing ledger.

A compact decision history for debugging, cost review and governance: what happened, why, how much it cost, and what the counterfactual would have been.

Time

Task

Cache

Model

Score

Cost

Saved

13:42:08

Refactor auth middleware

MISS

Opus 5

94.7

$0.021

$0.046

13:41:54

Explain test failure

HIT

cache

—

$0.000

$0.012

13:41:31

Generate migration query

MISS

Gemini 3.7

91.2

$0.004

$0.019

13:40:58

Review security boundary

MISS

GPT-5.6

95.1

$0.025

$0.031

13:40:12

Summarize build output

HIT

cache

—

$0.000

$0.006

13:39:44

Classify issue labels

MISS

DeepSeek V4

88.8

$0.002

$0.011

Make “auto” inspectable.

Connect real telemetry later without redesigning the page: the visual contract is already there.

Explore models Launch Arcana →