How we compute the cost comparison
The inputs behind the calculator and the home-page comparison, published so you can reproduce them instead of trusting them.
Every figure on this page is computed when the page is built, from the same modules the calculator uses: the model catalog, the mix model and the published price list. Nothing here is typed into a translation file.
What this page establishes
At the default preset the mixed-catalogue cost is 16% of an all-frontier baseline — about one in 6 — while the capability score retains 97.3% of the baseline. Both numbers are outputs of the formulas below, not marketing copy.
The default scenario
The calculator opens on one scenario: 30B tokens a month, 10% of them output, 90% of execution on the open tier, and a cache hit rate of 80% on input. The planner is Claude Opus 5.5; the executor is DeepSeek V4.1 Flash. These are illustrative enterprise volumes, not a measurement of any customer.
Default preset
| Input | Value | Source |
|---|---|---|
| DEFAULT_MIX_INPUT.monthlyTokens | 30B | shared/mix-economics.ts |
| DEFAULT_MIX_INPUT.outputRatio | 10% | shared/mix-economics.ts |
| DEFAULT_MIX_INPUT.openShare | 90% | shared/mix-economics.ts |
| DEFAULT_MIX_INPUT.cacheHitRate | 80% | shared/mix-economics.ts |
| DEFAULT_MIX_INPUT.plannerId | Claude Opus 5.5 | shared/model-catalog.ts |
| DEFAULT_MIX_INPUT.executorId | DeepSeek V4.1 Flash | shared/model-catalog.ts |
| INTRINSIC_PLANNING_FRAC | 10% | shared/mix-economics.ts |
The prices used
List prices are taken from vendor pricing pages and stamped as of 2026-11-02. Under BYOK you pay them directly and GateLLM adds no markup. Where a vendor publishes several prices for the same model, the catalog takes the higher one — the direction that understates our savings.
Catalog prices (USD per million tokens)
| Model | Vendor | Input | Output | Cached input | Source |
|---|---|---|---|---|---|
| Claude Sonnet 5.5 | Anthropic | $2 | $10 | $0.10 | https://platform.claude.com/docs/en/about-claude/pricing |
| Claude Opus 5.5 | Anthropic | $4 | $20 | $0.20 | https://platform.claude.com/docs/en/about-claude/pricing |
| Claude Fable 5.1 | Anthropic | $10 | $50 | $0.25 | https://platform.claude.com/docs/en/about-claude/pricing |
| GPT-6 Astra | OpenAI | $10 | $50 | $1 | https://developers.openai.com/api/docs/pricing |
| Claude Haiku 5.5 | Anthropic | $0.10 | $0.50 | $0.01 | https://platform.claude.com/docs/en/about-claude/pricing |
| GLM-5.3-Flash | Zhipu | $0.15 | $0.50 | $0.03 | https://docs.z.ai/guides/overview/pricing |
| DeepSeek V4.1 Flash | DeepSeek | $0.30 | $1.20 | $0.006 | https://api-docs.deepseek.com/quick_start/pricing/ |
| Qwen3.8-Max | Alibaba | $2 | $6 | $0.25 | https://www.alibabacloud.com/help/en/model-studio/qwen3-8-max |
How the cost ratio is computed
The baseline sends all traffic to the planner and applies the same cache assumption. The mix sends 90% of the traffic to the executor. The ratio is mix divided by baseline: 16%, or about one in 6, at the default preset. In absolute terms that is $86K baseline, $13K mixed and $73K saved per month.
Monthly cost at the default preset
| Line | USD per month | Note |
|---|---|---|
| planner input | $3K | cache discount applied |
| planner output | $6K | |
| executor input | $2K | cache discount applied |
| executor output | $3K | |
| baseline (frontier-only) | $86K | all traffic on the planner |
| mix | $13K | sum of the four rows above |
| saved | $73K | baseline − mix |
| cache saved | $14K | information only — already inside input rows |
How capability retention is computed
Each model carries two scores, planning and routine. 10% of the work is treated as planning and the rest as routine execution. The baseline uses the planner for both; the mix gives the planner at most the planning share and sends the remainder to the executor. Retention at the default preset is 97.3%. The scores are illustrative ranges derived from public leaderboards, not a benchmark we ran.
Capability scores
| Role | Planning | Routine | Source |
|---|---|---|---|
| planner | 100 | 100 | https://artificialanalysis.ai/models/claude-opus-5-5 |
| executor | 88 | 97 | https://lmarena.ai/leaderboard |
| baseline score | 100 | INTRINSIC_PLANNING_FRAC weighted | |
| mix score | 97 | INTRINSIC_PLANNING_FRAC weighted |
What moves the ratio
Two things dominate: which model takes the executor role, and how much traffic it takes. The tables below re-run the same calculation for every executor in the catalog, and across a range of shares around the sweet spot of 70% to 92%. Both directions are shown — a cheaper executor saves more, a stronger one saves less.
Executor sensitivity
| Executor | Cost ratio | One in N | Retention |
|---|---|---|---|
| Claude Haiku 5.5 | 12% | 8.1 | 95.5% |
| GLM-5.3-Flash | 13% | 7.6 | 94.6% |
| DeepSeek V4.1 Flash | 16% | 6.4 | 97.3% |
| Qwen3.8-Max | 46% | 2.2 | 97.3% |
| Claude Sonnet 5.5 | 55% | 1.8 | 98.2% |
Open-share sweep
| Open share | Cost ratio | One in N | Retention |
|---|---|---|---|
| 0% | 100% | 1 | 100.0% |
| 70% | 34% | 2.9 | 97.9% |
| 90% | 16% | 6.4 | 97.3% |
| 92% | 14% | 7.3 | 97.1% |
| 100% | 6% | 16 | 96.1% |
What the ratio does not include
The ratio compares model API spend only. It excludes the GateLLM licence ($200 a month on the Pro tier, which at these volumes is 1.5% of the mixed figure above), your infrastructure, and engineering time. It also does not model discounts you negotiate with providers separately.
Boundaries worth knowing
Read these before quoting any number from this page.
- •The preset is one plausible enterprise shape; your own mix will differ.
- •Capability scores are illustrative, and the cost side does not depend on them.
- •Price changes move the ratio: the catalog is stamped and refreshed, not frozen.
- •The ratio is an architecture comparison, not a commitment about your bill.