This is our rubric, not an industry benchmark. We do not have a peer cohort and we will not pretend to place you in one. What we can do instead is publish the whole model — every scoring weight, every coefficient, every threshold and every formula — so you can argue with it before you trust it.
The page is generated from the same modules the four instruments run on. A number cannot appear on a tool and be absent here, and it cannot say one thing here and another there, because there is only one copy of it.
Every number in the toolkit carries exactly one of two labels, and there is no third state. A figure we took from someone else’s published document is marked PUBLISHED and carries the source, the date we checked it and a live link. A figure we chose is marked ASSUMPTION and is yours to change.
An assumption is never described as a benchmark, and a published figure is never quietly adjusted to suit an argument. Where a figure is secondary reporting rather than a primary source, the label says so in the same sentence.
PUBLISHED — VelocityMind published pricing, /pricing, 26 Jul 2026
Source, date and link are all mandatory on this branch. A citation without a link is not a citation.
ASSUMPTION — yours to set
An assumption has no source by definition. The tools render every one of these as a control, because a number that materially moves the answer and cannot be seen or changed does not ship.
| Concern | Owned by | Consumed by |
|---|---|---|
| Published prices, bands and the delivery window | src/lib/pricing-tracks.ts | /pricing, the necessity build band, the payback bill of materials |
| Reliability arithmetic | src/lib/tools/reliability.ts | the necessity test and the configurator |
| Vendor model prices and cache mechanics | src/lib/tools/model-prices.ts | the configurator |
| Tokens, orchestration and run cost | src/lib/tools/architecture.ts | the configurator, and the payback model via its cost-per-item input |
| Build decomposition and cash flow | src/lib/tools/payback.ts | the payback model |
| Artefact gates and verdict rules | src/lib/tools/readiness.ts | the readiness ledger |
Every tool that touches money or time is bounded by the price list on /pricing. These constants live in one module and are imported by both places, so a price change propagates rather than drifting. A modelled payback shorter than the build is structurally impossible, because the build is spent across the published delivery window before any benefit is allowed to start.
| Track | Published range | Duration | Endpoints used in the model |
|---|---|---|---|
| Agent strategy and consulting | $5K - $15K | 2-4 weeks | $5,000 – $15,000 · 2–4 wk |
| Multi-agent architecture design | $10K - $30K | 3-6 weeks | $10,000 – $30,000 · 3–6 wk |
| Agent development and deployment | $25K - $75K | 6-12 weeks | $25,000 – $75,000 · 6–12 wk |
| Enterprise integration | $15K - $50K | 4-8 weeks | $15,000 – $50,000 · 4–8 wk |
$25,000 – $150,000
12–20 weeks from kickoff to production. The build is spent across the first 4 months of every cash flow — the mid-point of that window expressed in whole months.
$5,000 – $15,000
2-4 weeks. A diagnostic, not a build.
This scenario models above our published implementation band. Engagements of this size are scoped as an Enterprise Partnership — a programme across several workflows, not a single build.
This models below our published $25,000 minimum engagement — smaller than we take on. This is your own team's fortnight, or a short advisory.
The band is used for display framing only. Returns are always computed from the unclamped modelled build. Clamping the cost while letting the benefit scale without limit is exactly how the old calculator printed a four-figure percentage return.
The tiers at the top of this scale are our most expensive engagements. The full scoring table is below so you can check we are not pushing you up it.
Ten questions feed two meters. Autonomy decides the tier; control never raises or lowers it, it only prices it. An unanswered question contributes zero, which biases the verdict down — a half-finished test never over-recommends.
A = clamp( round( max(rawA, 0) / 168 × 100 ), 0, 100 ) → picks the tier C = clamp( round( rawC / 68 × 100 ), 0, 100 ) → prices the tier
| Option | Answer | Points | Also sets |
|---|---|---|---|
| a | Structured records — database fields, API payloads, form submissions | 0 | |
| b | Semi-structured documents with a known layout — invoices, claim forms | 8 | |
| c | Mixed formats, some scanned or photographed | 16 | |
| d | Free-form language — email, tickets, clinical notes, call transcripts | 24 | |
| e | Multi-modal — text plus image, audio or documents together | 30 |
| Option | Answer | Points | Also sets |
|---|---|---|---|
| a | One path, always identical | 0 | |
| b | A handful of known branches, under ten, all documented | 6 | |
| c | Dozens of branches — enumerable in principle, nobody has | 16 | |
| d | The path depends on what you find as you go | 28 | |
| e | Nobody in the business can enumerate the steps in advance | 36 |
Out of 100 items, how many need a human to THINK, not just to check?
| Option | Answer | Points | Also sets |
|---|---|---|---|
| a | Under 2 | 0 | |
| b | 2 to 10 | 5 | |
| c | 10 to 25 | 12 | |
| d | 25 to 50 | 20 | |
| e | Over 50 | 26 | judgementDominant |
For invoice processing, best-in-class exception rate is 9% against 22% typical (Ardent Partners, ePayables 2024). Shown as context, never as our coefficient.
When a human does one of these, what are they actually doing?
| Option | Answer | Points | Also sets |
|---|---|---|---|
| a | Following a documented rule | 0 | |
| b | Applying a rule nobody ever wrote down | 10 | |
| c | Weighing several sources and trade-offs | 18 | |
| d | Making a contested call two of your own experts could disagree on | 26 | noGroundTruth |
How many systems must one item touch to finish?
| Option | Answer | Points | Also sets |
|---|---|---|---|
| a | One, read-only | 0 | |
| b | One, read and write | 4 | |
| c | Two to four systems | 10 | |
| d | Five or more, or a system with no API | 16 |
How many discrete decisions or actions does one item take, end to end?
| Option | Answer | Points | Also sets |
|---|---|---|---|
| a | 1 to 2 | 0 | S = 2 |
| b | 3 to 5 | 2 | S = 4 |
| c | 6 to 12 | 5 | S = 9 |
| d | 13 to 25 | 9 | S = 18 |
| e | 25 or more | 12 | S = 30 |
| Option | Answer | Points | Also sets |
|---|---|---|---|
| a | None — internal convenience | 0 | |
| b | Internal policy or audit only | 12 | |
| c | Contractual or financial obligation — SOX, a customer SLA | 22 | |
| d | Sector-regulated — HIPAA, GxP, FDA/MDR, financial conduct | 32 | mandatoryAuditTracedeterministicPostCheck |
| e | Safety-critical — an error injures someone or stops a line | 40 | mandatoryAuditTracedeterministicPostCheckrecommendOnlyforcedHitl |
| Option | Answer | Points | Also sets |
|---|---|---|---|
| a | Trivial — someone notices and redoes it | 0 | |
| b | Rework — hours of someone's time | 8 | |
| c | Money moves — a payment, price or shipment goes wrong | 18 | |
| d | Legal, clinical or safety harm | 28 |
| Option | Answer | Points | Also sets |
|---|---|---|---|
| a | Sub-second — inside a keystroke or a machine control loop | -20 | |
| b | A few seconds, interactive | -6 | |
| c | Minutes | 0 | |
| d | Hours, or an overnight batch | 4 |
How often do the rules of this work change?
| Option | Answer | Points | Also sets |
|---|---|---|---|
| a | Once a year or less | 0 | |
| b | A few times a year | 5 | |
| c | Monthly | 11 | |
| d | Constantly — per customer, per jurisdiction | 18 |
| Coefficient | Default | Adjustable range | Label and basis |
|---|---|---|---|
| Monthly volume | 5,000 items/month | 10 items/month – 1,000,000 items/month | Contributes nothing to the verdict. Volume does not change what this should be — it only changes how much the answer costs you. |
| Tier | Autonomy band | Verdict | Build band at C = 0 | Build band at C = 100 | Cost of the tier above |
|---|---|---|---|---|---|
| T0Rules or a script | 0–14 | This is a job for code, not for a model. | below $25,000 | below $25,000 | +$25,000 → T1 |
| T1Deterministic workflow / RPA | 15–29 | Automate the path, not the thinking. | $25,000 – $45,000 | $25,000 – $60,750 | +$5,000 → T2 |
| T2One LLM call inside a deterministic pipeline | 30–47 | One model call, tightly fenced, inside code you control. | $30,000 – $60,000 | $30,000 – $81,000 | +$25,000 → T3 |
| T3Retrieval-grounded single agent | 48–69 | One agent, grounded in your corpus, choosing its own tools. | $55,000 – $105,000 | $55,000 – $141,750 | +$35,000 → T4 |
| T4Supervised multi-agent with human-in-the-loop | 70–100 | Several specialists under a supervisor, with a human on the gate. | $90,000 – $150,000 | $90,000 – $150,000 | — |
upliftMultiplier = 1 + (C / 100) × 0.35 bandTop = min( publishedTop × upliftMultiplier, $150,000 ) overEngineering = bandBottom(tier + 1) − bandBottom(tier)
| Coefficient | Default | Adjustable range | Label and basis |
|---|---|---|---|
| Control uplift on the build band top | 0.35× at C=100 | A modelling assumption. At full control demand it lifts the top of the band by 35%, covering the evaluation harness, the human-in-the-loop gate and the audit trace that regulated work requires. |
Gartner forecasts that over 40% of agentic AI projects will be cancelled by the end of 2027 on escalating cost, unclear business value or inadequate risk controls (poll of 3,412).
T0
Nothing in your answers asks for language understanding, open-ended planning or judgement under ambiguity. A model here adds latency, cost and a failure mode you did not previously have, in exchange for capability the workflow never uses. Write the rules down and run them.
T1
The work is repeatable and the path is known, so the value is in wiring the systems together reliably rather than in reasoning about them. A workflow engine gives you the same outcome with a fraction of the run cost and an audit trail that does not require an evaluation programme to defend.
T2
There is real language or format ambiguity here, but the path around it is fixed. That is the cheapest place a model ever sits: one call with a constrained output schema, wrapped in deterministic code that decides what happens next. You get the language capability without inheriting an agent's failure surface.
T3
The path genuinely depends on what the system finds as it goes, so a fixed pipeline would need branches nobody can enumerate. A single agent with retrieval and a bounded tool set handles that — but every step it takes multiplies down your end-to-end reliability, so the evaluation harness and the tool boundary are the build, not the prompt.
T4
The work spans distinct domains, the path is discovered rather than declared, and the cost of a wrong answer is real. That combination is the one case where multi-agent orchestration earns its complexity — and it is also the one where the review loop, not the model, sets your bill. Budget for evaluation and oversight before you budget for capability.
| # | Override | Fired by | Effect | What it means |
|---|---|---|---|---|
| 01 | Latency cap | Q09 — sub-second budget | caps at T1 | A sub-second budget takes an LLM off the critical path. Anything model-driven runs alongside the loop, never inside it. |
| 02 | Structured and stable | Q01 structured input, Q02 a fixed or documented path, Q10 stable rules | caps at T1 | Structured input on a fixed path with stable rules is the textbook definition of a job for code — regardless of every other answer. |
| 03 | Interactive cap | Q09 — a few seconds, interactive | caps at T3 | An interactive budget rules out a supervisor's extra serial hops. One agent, or a router that dispatches once. |
| 04 | Build nothing yetblocking | Q04 — a contested call two of your own experts could disagree on | components, not tier | Two of your own experts would disagree on the answer. That means you have no ground truth, which means you cannot build an evaluation set, which means you cannot tell whether the system is working. Fix the ground truth first — that is a workshop, not a build. |
| 05 | Decision support, not automation | Q03 — over half the items need a human to think | components, not tier | When most items need judgement, the system's job is to prepare the decision, not to make it. The verdict stands, but the shape of it changes. |
| 06 | Recommend only | Q07 — safety-critical | components, not tier | The system may recommend. It never actuates. Human-in-the-loop is not a design option here, it is the design. |
| 07 | Regulation prices it, it does not forbid it | Q07 — sector-regulated or safety-critical | components, not tier | A deterministic post-check and a reproducible audit trace become non-optional line items. Regulation does not forbid an agent. It prices one. |
One module owns this arithmetic and both the necessity test and the configurator import it. Serial steps multiply down: nothing about the model changes when you add a step, but the end-to-end number falls anyway. Production reliability is an architecture problem before it is a model-selection problem.
end-to-end P = r^S required per step r = target^(1 / S) rescue queue Q = volume × (1 − P) evidence cases N = ceil( ln(1 − confidence) / ln(1 − tolerance) ) [rule of three ≈ 3 / tolerance]
| Coefficient | Default | Adjustable range | Label and basis |
|---|---|---|---|
| Per-step reliability | 95.0% | 80.0% – 99.5% · step 0.5% | An assumption, not a benchmark. If you have error logs from the humans doing this work today, use their number. |
| End-to-end target used for the required-per-step figure | 95.0% | The target we solve the required per-step rate against. Change the target, change the requirement. | |
| Confidence level for the zero-failure bound | 95.0% | The 95% convention. Raise it and the sample size rises with it. |
| Serial steps (S) | End-to-end P | Items a month landing on a human | Per-step rate the target would need |
|---|---|---|---|
| 1 | 95.0% | 250 | 95.00% |
| 3 | 85.7% | 713 | 98.30% |
| 5 | 77.4% | 1,131 | 98.98% |
| 9 | 63.0% | 1,849 | 99.43% |
| 15 | 46.3% | 2,684 | 99.66% |
| 20 | 35.8% | 3,208 | 99.74% |
| Error ceiling claimed | Clean cases needed | Rule-of-three approximation |
|---|---|---|
| 5% | 59 | 60 |
| 1% | 299 | 300 |
| 0.5% | 598 | 600 |
| 0.1% | 2,995 | 3,000 |
| 0% | There is no sample size that proves a 0% error rate. Statistics can only give you an upper bound at a confidence level. Pick the rate you can live with and we will hold you to it. | — |
On tau-bench, state-of-the-art function-calling agents succeed on under 50% of tasks, and pass^8 falls below 25% in the retail domain. Shown as published context beside the panel. Never used as a coefficient.
9% best-in-class against 22% typical. Shown as context, never as our coefficient.
The configurator picks an orchestration pattern from an ordered rule tree, decides which components are in and which are out, assigns a model tier per role, then costs the whole thing bottom-up from tokens. The first rule in the tree refuses the sale.
NOT AN AGENT
One model call inside a deterministic workflow.
Fires when: steps ≤ 2 AND one domain AND consequence is not regulated AND the order of steps is fixed
Two steps, one domain, a fixed order and no regulatory exposure. There is nothing here for an agent to decide. Put one model call inside code you control and spend the difference on the validator.
You may not need an agent for this — run the necessity check
| Rule | Selects | Condition | What it prints |
|---|---|---|---|
| rule-0 | Not an agent | steps ≤ 2 AND one domain AND consequence is not regulated AND the order of steps is fixed | Two steps, one domain, a fixed order and no regulatory exposure. There is nothing here for an agent to decide. Put one model call inside code you control and spend the difference on the validator. |
| rule-1 | Single agent | steps ≤ 2 AND one domain | Two steps in one domain is a single tool-calling loop. Anything more elaborate is scaffolding you will maintain forever. |
| rule-2 | Sequential pipeline | fixed order AND steps ≤ 6 | A known order is worth money: each stage can be narrower than the last, and nothing carries the transcript forward. |
| rule-3 | Router + specialists | interactive latency AND steps ≤ 6 | An interactive budget cannot afford a supervisor's extra serial hop. Classify once, dispatch to one specialist. |
| rule-4 | Supervisor + specialists | domains ≥ 2 AND steps ≥ 4 | Several domains and a discovered path is the one case where a supervisor earns its complexity — and its token bill. |
| rule-5 | Parallel fan-out + evaluator | (regulated OR a diagnosis or drafting workload) AND latency is not interactive AND runs ≤ 50,000/month | Where being wrong is expensive and the volume is moderate, three parallel attempts plus an evaluator buys accuracy with tokens rather than with time. |
| rule-6 | Supervisor + specialists | the selected pattern costs more than 3x the supervisor alternative (evaluated after costing) | The chosen pattern costs more than three times the simpler alternative on your own volume. We downgraded it and are telling you so. |
Fan-out stops being economic above 50,000 runs a month. Rule 6 downgrades a pattern that costs more than 3× the supervisor alternative on your own volume, and says so on screen rather than silently.
| Pattern | Calls / run | Serial hops | Context scale | Carries transcript | Extra context copies | Branches |
|---|---|---|---|---|---|---|
| Not an agent | 1 | 1 | 1× | no | 0 | 1 |
| Single agent | 8 | 8 | 1× | yes | 0 | 1 |
| Sequential pipeline | 8 | 8 | 0.65× | no | 0 | 1 |
| Router + specialists | 5 | 5 | 1× | yes | 0 | 1 |
| Supervisor + specialists | 8 | 8 | 1× | yes | 1 | 1 |
| Parallel fan-out + evaluator | 12 | 4 | 1× | yes | 0 | 3 |
prefix = 900 (system) + 220 × tools + 600 (few-shot) + 350 (tool-use overhead) tools = systems + (retrieval ? 1 : 0) + 1 context/call = base × contextScale × memoryUplift × (1 + extraContextCopies) base = retrieval ? 3,500 : min(pages × tokensPerPage, 120,000) snowball = branches × (O + K) × N(N−1)/2 ← the accumulated transcript input/run = calls × (prefix + context/call) + snowball + longTermMemory output/run = calls × (O + Th) cached input = uncached + cacheable × ( hit × 0.1 + (1 − hit) × 1.25 ) cost/run = Σ roles [ input × inputRate + output × outputRate ] / 1e6 × retry × deployment × batch × priceIndex
| Steps per run | Total tokens per run | Tokens per step |
|---|---|---|
| 1 | 11K | 11K |
| 4 | 30K | 7,578 |
| 8 | 195K | 24K |
| 12 | 323K | 27K |
| 16 | 476K | 30K |
| 20 | 657K | 33K |
| Coefficient | Default | Adjustable range | Label and basis |
|---|---|---|---|
| Base system prompt | 900 tokens | A modelling assumption for a production system prompt, re-sent on every call. | |
| Tool schema, per declared tool | 220 tokens | A modelling assumption for one typed tool definition. | |
| Few-shot examples | 600 tokens | A modelling assumption. Remove the examples and this goes to zero. | |
| Tool-use system-prompt overhead | 350 tokens | ||
| Retrieved context per call | 3,500 tokens | A modelling assumption for a top-k retrieval result set. | |
| Tokens per page | 750 tokens | 300 tokens – 1,200 tokens · step 25 tokens | Derived from the published ~0.75 words-per-token ratio, not a measured figure. Dense legal or academic text runs higher; double-spaced runs lower. |
| Working-memory uplift on per-call context | 0.12× | A modelling assumption: a summarised state blob carried forward once a run exceeds three steps. Separate from the accumulated transcript, which is modelled explicitly. | |
| Retrieval hop latency | 0.35s | A modelling assumption for one vector-store round trip. | |
| Tool call latency | 1.00s | A modelling assumption for one system round trip. | |
| p95 / p50 latency ratio | 2.10× | Measured per-call tails run 1.6–3.2x with a ~2.1x average. Applied to a whole chain this is a conservative upper bound, because independent-ish steps make the aggregate tail tighter than the per-call tail. |
| Constant | Value |
|---|---|
| Tool round trips per system touched | 2 |
| Planner output tokens | 600 tokens |
| Router input / output tokens | 1,200 / 80 tokens |
| Guardrail calls per run, and their shape | 2 calls · 900 in / 40 out |
| Long-term memory injected per run | 200 tokens |
| Raw document context ceiling | 120K tokens |
| On-prem or air-gapped penalties | price ×1.6 · throughput ×0.6 |
Modelling assumptions for self-hosted capacity: amortised hardware and operations against list API pricing, and lower sustained throughput than a managed endpoint.
| Answer length | Output tokens |
|---|---|
| Short | 250 |
| Standard | 450 |
| Long, drafted prose | 900 |
| Latency tolerance | Budget |
|---|---|
| Interactive (under 5s) | 5.00s |
| Near-real-time (under 30s) | 30.0s |
| Batch (minutes) | 300s |
| Overnight | no budget — switches to batch pricing |
| Workflow class | Domains | Fixed order | Needs a corpus | Systems | State between runs |
|---|---|---|---|---|---|
| Extract and validate structured data | 1 | yes | no | 1 | no |
| Answer questions over an internal corpus | 1 | no | yes | 1 | no |
| Triage and route an inbound queue | 2 | no | no | 2 | no |
| Diagnose an anomaly from telemetry and history | 3 | no | yes | 3 | yes |
| Draft a regulated or clinical document | 2 | no | yes | 2 | no |
| Execute a multi-system transaction | 2 | yes | no | 4 | no |
regulated → advisory (100% reviewed) costly → assisted (15% reviewed) annoying AND P(end-to-end) < 90% → autonomous with an audit sample (3%) annoying AND P(end-to-end) ≥ 90% → autonomous (0%)
| Posture | Review rate | Why |
|---|---|---|
| Advisory — every output reviewed | 100% | Regulated or safety-critical work is reviewed work. The system recommends; a person decides. |
| Assisted — exceptions only | 15% | A wrong answer moves money, so the exceptions get a human and the rest do not. |
| Autonomous with an audit sample | 3% | Low consequence, but end-to-end reliability is soft enough that a sample has to be checked. |
| Autonomous | 0% | Low consequence and reliable enough end to end that routine review would cost more than the errors. |
Inputs: Extract and validate structured data, 20,000 runs a month, 8 steps, 8 pages, Standard answers, Near-real-time (under 30s) latency, Costly consequence, Vendor cloud API deployment. Selected by rule-4: Several domains and a discovered path is the one case where a supervisor earns its complexity — and its token bill.
Every rate below comes from the vendor’s own pricing page, not from a third-party aggregator. Aggregator figures for the same models already disagree with vendor pages by up to a quarter, which is why they are not cited here at all. Prices move; the date stamp is the point.
| Model | Tier | Input $/MTok | Output $/MTok | Cache read | Cache write 5m | Context | Min cacheable prefix |
|---|---|---|---|---|---|---|---|
| Claude Fable 5 | frontier | $10.00 | $50.00 | $1.00 | $12.50 | 1.00M | 512 tok |
| Claude Opus 5 | frontier | $5.00 | $25.00 | $0.50 | $6.25 | 1.00M | 512 tok |
| Claude Sonnet 5 | mid | $3.00 | $15.00 | $0.30 | $3.75 | 1.00M | 1,024 tok |
| Claude Sonnet 4.6 | mid | $3.00 | $15.00 | $0.30 | $3.75 | 1.00M | 1,024 tok |
| Claude Haiku 4.5 | small | $1.00 | $5.00 | $0.10 | $1.25 | 200K | 4,096 tok |
| Model | Tier | Input $/MTok | Output $/MTok | Cache read | Cache write 5m | Context | Min cacheable prefix |
|---|---|---|---|---|---|---|---|
| gpt-5.6-sol | frontier | $5.00 | $30.00 | $0.50 | not published | 1.05M | — |
| gpt-5.6-terra | mid | $2.50 | $15.00 | $0.25 | not published | 1.05M | — |
| gpt-5.6-luna | small | $1.00 | $6.00 | $0.10 | not published | 1.05M | — |
| gpt-5.5 | frontier | $5.00 | $30.00 | $0.50 | not published | — | — |
| gpt-5.4 | mid | $2.50 | $15.00 | $0.25 | not published | — | — |
| gpt-5.4-mini | small | $0.75 | $4.50 | $0.08 | not published | — | — |
| gpt-5.4-nano | small | $0.20 | $1.25 | $0.02 | not published | — | — |
| Model | Tier | Input $/MTok | Output $/MTok | Cache read | Cache write 5m | Context | Min cacheable prefix |
|---|---|---|---|---|---|---|---|
| Gemini 3.1 Pro Preview | frontier | $2.00 | $12.00 | $0.20 | not published | 1.05M | — |
| Gemini 3.6 Flash | mid | $1.50 | $7.50 | $0.15 | not published | 1.05M | — |
| Gemini 3.5 Flash | mid | $1.50 | $9.00 | not published | not published | — | — |
| Gemini 3.5 Flash-Lite | small | $0.30 | $2.50 | $0.03 | not published | — | — |
| Gemini 2.5 Flash | small | $0.30 | $2.50 | not published | not published | — | — |
| Gemini 2.5 Flash-Lite | small | $0.10 | $0.40 | not published | not published | — | — |
| Model | Tier | Input $/MTok | Output $/MTok | Cache read | Cache write 5m | Context | Min cacheable prefix |
|---|---|---|---|---|---|---|---|
| gpt-oss-120B | open | $0.15 | $0.60 | not published | not published | — | — |
| gpt-oss-20B | open | $0.05 | $0.20 | not published | not published | — | — |
| Llama 3.3 70B | open | $1.04 | $1.04 | not published | not published | — | — |
| Qwen3.5-397B-A17B | open | $0.60 | $3.60 | not published | not published | — | — |
| Qwen3.5 9B | open | $0.17 | $0.25 | not published | not published | — | — |
| DeepSeek V4 Pro | open | $1.74 | $3.48 | $0.20 | not published | — | — |
| Kimi K2.6 | open | $1.20 | $4.50 | not published | not published | — | — |
| Model | Long-context input $/MTok | Long-context output $/MTok |
|---|---|---|
| gpt-5.6-sol | $10.00 | $45.00 |
| gpt-5.6-terra | $5.00 | $22.50 |
| gpt-5.6-luna | $2.00 | $9.00 |
read ×0.1 · write 5m ×1.25 · write 1h ×2
At a 1.25x write and a 0.10x read, a cached prefix pays for itself after one read.
×0.5
Does not apply to interactive requests or to managed-agent sessions.
×1.1 on every token category
Below the threshold, caching does nothing at all — and you still pay the write premium for trying. Common default 1,024 tokens; Claude Opus 5 512 · Claude Fable 5 512 · Claude Sonnet 5 1,024 · Claude Sonnet 4.6 1,024 · Claude Haiku 4.5 4,096.
| Model | auto / none | any / tool |
|---|---|---|
| Claude Opus 5 | 286 | 406 |
| Claude Sonnet 5 | 354 | 474 |
| claude-opus-4-7 | 675 | 804 |
Server-tool extras: bash 325 · text editor 700 · computer use 735 tokens per call. The configurator uses a single rounded 350-token figure and publishes this table as its basis.
| SKU | Rate |
|---|---|
| Web search, per 1,000 searches | $10.00 |
| Web fetch | no surcharge |
| Code execution, per container-hour | $0.05 after 1,550 free hours a month |
| Managed agent runtime, per session-hour | $0.08 — metered to the millisecond |
| Tier | Reference model | Input / output $/MTok | Throughput | TTFT | Reasoning first answer |
|---|---|---|---|---|---|
| Frontier | Claude Opus 5 | $5.00 / $25.00 | 78 tok/s | 0.90s | 28.0s |
| Mid | Claude Sonnet 5 | $3.00 / $15.00 | 85 tok/s | 0.90s | 28.0s |
| Small | Claude Haiku 4.5 | $1.00 / $5.00 | 160 tok/s | 0.40s | 12.0s |
Public-endpoint medians move constantly. Method follows Artificial Analysis; the numbers themselves are our assumptions and are yours to change.
Reasoning mode is a different latency category, not an increment — the reasoning figure already includes thinking time and must never be charted on the same axis as time-to-first-token.
| Anchor | Tokens |
|---|---|
| Characters per token | 4 |
| Tokens per word | 1.33 |
| Average web page (~10 kB) | 2,500 |
| Large documentation page (~100 kB) | 25K |
| Research-paper PDF (~500 kB) | 125K |
| Coefficient | Default | Adjustable range | Label and basis |
|---|---|---|---|
| Tokenizer multiplier for Claude 4.7 and later | 1.30× | ||
| Tokens per page | 750 tokens | 300 tokens – 1,200 tokens · step 25 tokens | Derived from the published ~0.75 words-per-token ratio, not a measured figure. Dense legal or academic text runs higher; double-spaced runs lower. |
Any tokens-per-page heuristic built on an older tokenizer baseline understates cost by about a third before any other modelling error. The anchors above are published; the tokens-per-page figure derived from them is still an assumption, and we label it as one.
10,000 conversations at 3,700 tokens each on a small-tier model: $37.
50K in / 15K out over 1 hour: $0.705, or $0.525 with caching. Serving 40k of the input from cache cut that session's cost by about 26% — caching, not model choice, was the dominant lever.
1x H100 $3.99/hr · 1x H200 $5.49/hr · 1x B200 $9.95/hr
Third-party aggregator reporting, not a vendor page. We do not cite aggregators as published prices — treat these as an order-of-magnitude marker and confirm with the vendor.
Industry-cited agentic token multipliers run 5×–30×. Secondary reporting, shown as context. Never used as a coefficient anywhere in the toolkit.
Anthropic pricing documentation, worked examples is the basis for the two worked examples above. A configurator whose output for a comparable workload lands orders of magnitude away from them is mis-specified, and we check against them rather than trusting the arithmetic.
Selecting a segment pre-fills every downstream default — people, hours, rate and volume, not just a coefficient — so nobody faces an empty slider. Every one of those pre-fills is an assumption and is labelled as one.
| Segment | Agent | Automatable share | Build multiplier | People | Hours / wk | Rate | Volume / mo |
|---|---|---|---|---|---|---|---|
| Document Intelligence | DocuMind | 55% | ×1 | 12 | 18 | $65 | 5,000 |
| Process Automation | FlowAgent | 50% | ×0.95 | 15 | 16 | $60 | 12,000 |
| Predictive Maintenance | MaintainAI | 42% | ×1.05 | 10 | 14 | $72 | 2,000 |
| Semiconductorevidence pack locked on | ChipSense | 38% | ×1.2 | 8 | 20 | $95 | 3,000 |
| Healthcareevidence pack locked on | MedAgent | 32% | ×1.35 | 14 | 15 | $78 | 6,000 |
Process and yield decisions carry a change-control and evidence burden. The evidence pack is not optional here, so we price it in rather than quietly leaving it out.
PHI handling, audit evidence and mandatory human review are regulatory requirements, not scope options. The evidence pack is priced in.
annual manual cost = people × hours/week × 52 × hourlyRate released capacity = annual manual cost × automatable realized savings = released capacity × realization inference/yr = volume × 12 × costPerItem review/yr = volume × 12 × exceptionRate × (reviewMinutes / 60) × hourlyRate run cost/yr = inference/yr + platform/yr + review/yr net annual benefit = realized savings − run cost/yr
| Track | Formula | Published range | Modelled |
|---|---|---|---|
| Agent development and deployment | 25,000 + 5,000 × (1 variant − 1) | $25K - $75K | $25,000 |
| Enterprise integration | 15,000 + 8,000 × (2 systems − 2) | $15K - $50K | $15,000 |
| Subtotal × domain multiplier × regulated multiplier | $40,000 × 1 × 1 | $25K - $150K | $40,000 |
Variants and systems are each capped at 5 in the segmented control — documented rather than hidden. The total is floored at our published $25,000 minimum engagement, which is what makes the entry price reachable in the model at all.
| Ramp | Months to full value | Factor by month after go-live |
|---|---|---|
| Fast (3 months to full) | 3 | 0.5 · 0.8 · 1 |
| Standard (7 months to full) | 7 | 0.4 · 0.4 · 0.4 · 0.7 · 0.7 · 0.7 · 1 |
| Slow (12 months to full) | 12 | 0.25 · 0.25 · 0.25 · 0.25 · 0.5 · 0.5 · 0.5 · 0.5 · 0.75 · 0.75 · 0.75 · 0.75 · 1 |
| Scenario | Benefit × | Cost × |
|---|---|---|
| conservative | 0.875 | 1.075 |
| expected | 1 | 1 |
| aggressive | 1.1 | 0.95 |
Forrester's TEI methodology risk-adjusts benefits down 10–15% and costs up 5–10%. The conservative column uses the mid-points of those two ranges. A calculator returning one to-the-dollar figure from seven guesses is not being more precise than this; it is being less honest.
| Order | Condition | Verdict |
|---|---|---|
| 01 | Net annual benefit ≤ 0 | negative — this workflow costs more to run than it saves |
| 02 | Payback is null (never crosses zero inside 36 months) or later than 24 months | slow — positive, but a portfolio decision rather than a pilot decision |
| 03 | First-year return above 400% | implausible — we show the threshold rather than the figure and ask you to re-check the scope |
| 04 | Everything else | clears |
The model also cross-checks your own inputs against each other: implied handling time outside 0.5–240 minutes per item means the headcount and the volume are probably describing different things, and the tool says so rather than quietly modelling it.
| Coefficient | Default | Adjustable range | Label and basis |
|---|---|---|---|
| Automatable share of in-scope hours | 55.0% | 10.0% – 80.0% · step 5.0% | The largest single lever in this model, which is exactly why it is a slider. Defaults come from the industry profile and reset when you change it. Bounded above by the published 60–70% technically-automatable ceiling, which is a ceiling and not our number. |
| Share of released hours converted to budget | 60.0% | 20.0% – 100.0% · step 5.0% | Released hours are capacity, not a guaranteed headcount reduction. This is your call on how much of that capacity turns into budget. Our 60% default is a modelling assumption with no published source — Forrester's TEI methodology recaptures 50% of hours saved in its own composite model, so ours is on the optimistic side of that. |
| Cost per item processed by the agent | $0.120 | $0.010 – $2.00 · step $0.010 | An assumption unless you bring it from the Architecture Configurator, which derives it bottom-up from tokens and orchestration pattern. |
| Exception rate | 15.0% | 2.0% – 50.0% · step 1.0% | For invoice processing, best-in-class is 9% against 22% typical (Ardent Partners, ePayables 2024). That is shown as context. Our 15% default is an assumption sitting between the two, and it is yours to set. |
| Review minutes per exception | 4 min | 1 min – 20 min · step 1 min | A modelling assumption. If you time it, use your number. |
| Platform and observability, annual | $18,000/yr | $6,000/yr – $60,000/yr · step $1,000/yr | Gateway, vector store, tracing, evaluation runs and on-call. A modelling assumption. |
| Discount rate for NPV | 10.0% | 0.0% – 20.0% · step 1.0% | The de facto convention, matching Forrester's TEI 10% and three-year NPV treatment. |
| Regulated evidence pack, build multiplier | 1.25× | A modelling assumption for validation documentation, change control and the reproducible audit trace a regulated deployment requires. |
McKinsey Global Institute puts 60–70% of work hours in scope as technically automatable with generative AI. We use that as a ceiling on this slider, never as the number itself.
Net $176,726 a year against a modelled $40,000 build, crossing zero at month 10 from kickoff — including the 12-20 weeks build itself.
The ledger asks about artefacts that either exist or do not — not about confidence, ambition or maturity. Its three column headings are Gartner’s three named causes of agentic project cancellation, used verbatim. “Don’t know” counts as No for the verdict but is reported separately and prominently, because four don’t-knows is itself the finding and it is the most common one.
Nobody owns the number, the released time or the system a year from now.
Nothing to measure against, so nothing can be proven.
No access, no authority, no way back.
| Gate | Group | The artefact | What that actually means | Typical effort |
|---|---|---|---|---|
| G01blocking | Unclear business value | A written definition of what counts as a wrong answer for this workflow, agreed by the people who own the process. | A page naming the failure modes and, for each, what the correct outcome would have been — signed off by the process owner, not written by the project team alone. | typically a day, once the right two people are in a room |
| G02blocking | Inadequate risk controls | Read access to the system of record in a non-production environment, available today. | Working credentials against a non-production instance with representative data — not a promise that access can be requested. | typically a fortnight, and it is almost always the long pole |
| G03blocking | Inadequate risk controls | A named individual with authority to let the agent act without review, and the authority to stop it. | One person, named, who can both switch it on and switch it off. A committee is not an answer to this question. | typically a day, and the difficulty is political rather than technical |
| G04blocking | Inadequate risk controls | A rollback that does not require a vendor to execute. | A documented, rehearsed way for your own team to reverse what the agent did, without opening a support ticket. | typically a few days to write down, longer to rehearse |
| G05the one gate | Unclear business value | A measured baseline error rate for the humans doing this work today. | A number, from a sample somebody actually checked, for how often the current process gets it wrong. Not an estimate from memory. | typically a fortnight of sampling, and it is the single most valuable fortnight in the project |
| G06 | Unclear business value | At least 50 real historical items, with their correct outcomes, that could be used as an evaluation set. | Fifty real cases with the right answer attached — the minimum from which an evaluation harness can be built. | typically a week, assuming the outcomes were recorded at the time |
| G07 | Escalating cost | A number the CFO already believes for what this workflow costs today. | A cost figure that has survived finance once already. A number invented for the business case will be challenged exactly when you need it. | typically a week |
| G08 | Inadequate risk controls | A written escalation path for the case the agent gets wrong — who is told, how fast, what they can do. | A named recipient, a stated time, and a stated remedy. One page. | typically a day |
| G09 | Inadequate risk controls | Write access to the system of record, or a documented decision that the agent will only recommend. | Either the write path exists, or somebody has decided in writing that it will not — both are fine, ambiguity is not. | typically a fortnight if write access is needed |
| G10 | Inadequate risk controls | A named owner for the data the agent will read, who can grant access without a committee. | One person who owns the data and can say yes. | typically a day to identify, longer if the answer is nobody |
| G11 | Inadequate risk controls | An agreed retention and logging policy that covers model inputs and outputs. | A written policy stating what is logged, where it lives and how long it is kept — covering prompts and completions, not only application logs. | typically a week with legal |
| G12 | Escalating cost | A decision on what happens to the time the agent frees up, made by someone who owns that budget. | A written answer to 'and then what?' from the person whose budget the released hours sit in. Without it, the saving is a slide. | typically a day, and it is usually the conversation nobody wants to have |
| G13 | Escalating cost | Somebody who will still own this system in twelve months. | A named owner with a role that survives the project. An unowned agent is a liability with a monthly bill. | typically a day |
| G14 | Unclear business value | A test environment where a failure costs nothing. | Somewhere the agent can be wrong repeatedly without touching a customer, a payment or a patient. | typically a fortnight |
| G11aregulated swap | Inadequate risk controls | A reproducible reasoning trace, sufficient to answer a regulator's question of why the system decided that. | A stored, replayable record of the inputs, retrieved evidence and intermediate decisions behind any single output — reproducible months later, not a log line. | typically designed in, not added later |
| G11bregulated swap | Inadequate risk controls | A validation and change-control procedure the agent will be brought under. | The existing procedure your regulated systems already follow, with a written decision on how a model-driven system enters it. | typically a fortnight with quality and regulatory affairs |
You cannot state the error rate of the humans doing this work today. That means you will have no way to prove the agent is better, and no way to defend it the first time it is wrong.
| Order | Rule | Verdict |
|---|---|---|
| 01 | Any blocking gate (G01–G04) answered No or Don't know. | no |
| 02 | Three or more non-blocking gates answered No or Don't know. | not-yet |
| 03 | Everything else. | go |
The threshold on rule 2 is 3 non-blocking gates. The blocking set is G01, G02, G03, G04.
Stating the instrument’s own blind spots is the part that makes the rest read as candour rather than technique. None of these tools can see the things that most often decide whether a build survives.
MIT NANDA (2025) found roughly 95% of enterprise GenAI pilots showed no measurable P&L return, with barriers that were organisational rather than model-related.
The tiers at the top of this scale are our most expensive engagements. The full scoring table is below so you can check we are not pushing you up it.
We sell agent implementations. The tiers at the top of the necessity scale, the patterns at the expensive end of the configurator and the verticals with the highest build multipliers are all our largest engagements. That is precisely why every weight that could push you toward them is printed on this page rather than described in a paragraph.
Deduplicated across all four instruments. Where a figure is secondary reporting rather than a primary source, the entry says so. We never cite a third-party price aggregator.
Anthropic pricing documentation
Per-MTok input and output rates, 0.10x cache read, 1.25x five-minute cache write, 2.0x one-hour cache write, 50% batch discount, minimum cacheable prefix by model, 286–406 token tool-use system-prompt overhead, metered server-tool and managed-agent SKUs.
Anthropic, Building Effective Agents
Workflows are predefined code paths; agents direct their own process. Avoid agents when latency or cost matter, or when the path is predictable.
Ardent Partners, ePayables 2024
9% best-in-class versus 22% typical invoice exception rate, shown as context beside the exception-rate slider and explicitly not as our default.
Artificial Analysis benchmarking methodology
Cited as the method behind the throughput and time-to-first-token figures, which are themselves labelled assumptions and are adjustable.
Cisco AI Readiness Index methodology
Cited as the reason we do NOT claim benchmarking: their cohort is 7,985 double-blind respondents across 30 markets. Ours does not exist.
Forrester Total Economic Impact methodology
Risk-adjusting benefits down 10–15% and costs up 5–10%; 50% recapture of hours saved; the 10% discount rate and three-year NPV convention. Cited by name as the basis for the conservative band.
Gartner press release, June 2025
Over 40% of agentic AI projects forecast cancelled by end-2027 on escalating cost, unclear business value and inadequate risk controls (poll of 3,412). The three named causes are the three column headings of this ledger.
Gartner, March 2026, as reported by Spheron and Forbes
Agentic workloads reported at 5–30x the tokens of a standard chatbot exchange. Secondary reporting, shown as context only.
Google Cloud, Choose a design pattern for your agentic AI system
The pattern cost/latency/complexity trade-off table underpinning the pattern tree.
Per-MTok rates, the 200k-token input price step, and the hourly context-cache storage fee the other vendors do not charge.
Hanley & Lippman-Hand, JAMA 1983
The rule of three for zero-failure confidence bounds — the basis for the evidence card's sample-size approximation.
McKinsey Global Institute, 2023
60–70% of work hours technically automatable with generative AI. Used strictly as a CEILING on the automatable-share assumption, never as the assumption itself.
MIT NANDA, 2025
Roughly 95% of enterprise GenAI pilots show no measurable P&L return, with barriers that are organisational rather than model-related.
Per-MTok rates including the short-context and long-context tier structure that most public calculators omit.
1 token ≈ 4 characters ≈ 0.75 words — the published basis for the tokens-per-page assumption, which remains an assumption.
OpenAI, A Practical Guide to Building AI Agents
Three positive criteria for an agent; otherwise a deterministic solution will do.
Open-weights serverless hosting rates per MTok.
VelocityMind published pricing
Agent strategy $5K–$15K, multi-agent architecture design $10K–$30K, agent development and deployment $25K–$75K, enterprise integration $15K–$50K, and the published $25,000–$150,000 / 12–20 week implementation band.
Yao et al., tau-bench (arXiv:2406.12045)
State-of-the-art function-calling agents succeed on under 50% of tasks; pass^8 falls below 25% in the retail domain.
This page is built to print. Every table, formula and coefficient renders on paper with its provenance intact, and every link resolves to its full URL. The market norm is a gated PDF; this is the ungated, auditable version you can forward to whoever has to sign the invoice.
If a weight is wrong, a coefficient is off or a source is stale, tell us which one and why. We would rather change the model than defend it.