Cost per successful task
Runtime cost is incomplete unless failures, retries, human review and exception handling are included. The useful unit is work that reaches an acceptable outcome.
Cybatar AI Economics
Measure the economics of intelligence—not just the price of tokens.
Cybatar connects model pricing, workload cost, enterprise outcomes and value-engineering methods so institutions can compare the full cost of machine work with the value it actually creates.
What is AI economics?
AI economics is the discipline of connecting the full cost of AI activity—including model usage, infrastructure, human review, failure, governance and implementation—to measurable outcomes, validated value and investment decisions.
Methodology
From model price to institutional economics.
A system-level method for comparing the full cost of machine work with measurable institutional value.
Cybatar Economic Efficiency Score
Where comparable capability evidence exists, CEES relates capability to a standardised workload basket and normalises the strongest observed capability-per-dollar result to 100. It is a research comparator—not a forecast of every production workload.
Cybatar AI Value Engineering 1.0
Methodology status: active · effective 23 Aug 2026
Model economic efficiency
Capability and cost on a common workload.
Use a standardised reference basket to compare current model economics without pretending the basket predicts your own production workload.
Knowledge Assistant. A bounded enterprise knowledge interaction used for price comparison across models. Basket: 8,000 input + 2,000 output tokens; 25% cached-input assumption.
| Provider | Model | Cost / task | Cost / 1,000 tasks | Capability evidence | CEES |
|---|---|---|---|---|---|
| DeepSeek | DeepSeek V4 Flash | USD 0.0014 | USD 1.41 |
52.0 AA Intelligence Index |
100.0 / 100 |
| OpenAI | GPT-5.6 Luna | USD 0.0036 | USD 3.64 |
52.0 AA Intelligence Index |
38.6 / 100 |
| DeepSeek | DeepSeek V4 Pro | USD 0.0044 | USD 4.36 |
53.0 AA Intelligence Index |
32.9 / 100 |
| Gemini 3.7 Flash | USD 0.0122 | USD 12.15 |
56.0 AA Intelligence Index |
12.5 / 100 | |
| OpenAI | GPT-5.6 Terra | USD 0.0364 | USD 36.40 |
57.0 AA Intelligence Index |
4.2 / 100 |
| xAI | Grok 4.6 | USD 0.0500 | USD 50.00 |
61.0 AA Intelligence Index |
3.3 / 100 |
| OpenAI | GPT-5.6 Sol | USD 0.0720 | USD 72.00 |
61.0 AA Intelligence Index |
2.3 / 100 |
| Anthropic | Claude Opus 5 | USD 0.0900 | USD 90.00 |
63.0 AA Intelligence Index |
1.9 / 100 |
| Anthropic | Claude Sonnet 5 | USD 0.0360 | USD 36.00 | No common capability observation | — |
| xAI | Grok 4.5 | USD 0.0492 | USD 49.20 | No common capability observation | — |
| Cohere | Command A | USD 0.0400 | USD 40.00 | No common capability observation | — |
| Gemini 3.6 Flash | USD 0.0122 | USD 12.15 | No common capability observation | — | |
| Gemini 3.5 Flash | USD 0.0273 | USD 27.30 | No common capability observation | — | |
| Gemini 3.5 Flash-Lite | USD 0.0069 | USD 6.86 | No common capability observation | — | |
| Mistral AI | Mistral Large 3 | USD 0.0061 | USD 6.10 | No common capability observation | — |
| Mistral AI | Mistral Medium 3.5 | USD 0.0243 | USD 24.30 | No common capability observation | — |
| Mistral AI | Mistral Small 4 | USD 0.0021 | USD 2.13 | No common capability observation | — |
Research signals
What the market is learning about AI economics.
Published observations remain connected to their source, date and confidence so they can inform decisions without being mistaken for universal truths.
Organisations using AI in at least one function
88percent
88% of surveyed organisations reported using AI in at least one business function in 2025.
Stanford HAI
Open sourceOrganisations regularly using generative AI
79percent
79% of surveyed organisations reported regular generative AI use in at least one function in 2025.
Stanford HAI
Open sourceOrganisations beyond AI experimentation
62percent
62% of respondents in McKinsey’s Enterprise AI FinOps survey had moved beyond experimentation into active AI deployment.
McKinsey & Company · 20 Jul 2026
Open sourceAgentic task token consumption compared with simpler interactions
1,000x approximate
McKinsey cites research indicating that context-heavy agentic tasks can consume roughly 1,000× more tokens than simpler code-reasoning or chat tasks. Exact ratios vary by workload.
McKinsey & Company · 08 Jul 2026
Open sourceAgentic cost associated with refinement and verification
60percent approximate
McKinsey cites production coding-workflow research indicating about 60% of agentic task cost can be tied to checking, repairing and re-verifying outputs. Exact shares vary by workload.
McKinsey & Company · 08 Jul 2026
Open sourceOrganisations exceeding AI budgets
93percent
93% of qualified respondents in McKinsey’s May 2026 Enterprise AI FinOps survey reported exceeding their AI budgets.
McKinsey & Company · 20 Jul 2026
Open sourceAI spend increase from isolated use cases to enterprise-wide adoption
4x approximate
McKinsey reports AI spend increasing nearly fourfold as organisations move from isolated use cases to enterprise-wide adoption.
McKinsey & Company · 20 Jul 2026
Open sourceAI spend often unaccounted for
McKinsey reports that 20–30% of AI spend is often unaccounted for because investments are fragmented across vendors, tools and commercial models.
McKinsey & Company · 20 Jul 2026
Open sourceCompanies with mature AI FinOps practices
McKinsey estimates only about 20–25% of companies have mature AI FinOps practices.
McKinsey & Company · 20 Jul 2026
Open sourceAI spend savings associated with high forecasting maturity
10percent more
McKinsey reports organisations with high forecasting maturity saving 10% more on AI spend than peers on average.
McKinsey & Company · 20 Jul 2026
Open sourceSavings from active AI-spend optimisation
About one-third of surveyed organisations had achieved 20–30% savings through active AI-spend optimisation actions.
McKinsey & Company · 20 Jul 2026
Open sourceProductivity gains in structured measurable work
The 2026 AI Index summarizes studies reporting roughly 14–15% gains in customer support, 26% in software development and 50% in marketing output, while noting smaller gains in deeper-reasoning work.
Stanford HAI
Open sourceSame-task token usage variance
30x up to
McKinsey cites evidence that token usage can vary by up to 30× when an agent executes the same task, reinforcing the need to model cost as a distribution rather than a fixed unit price.
McKinsey & Company · 20 Jul 2026
Open sourceWage premium for workers with AI skills
62percent
PwC’s 2026 Global AI Jobs Barometer reports a 62% average wage premium for workers with AI skills compared with comparable roles without those skills.
PwC
Open source
Enterprise value evidence
Reported outcomes connected to underlying deployment evidence.
Cooper University Health Care
4minutes
documentation time saved per patient
documentation time saved per patient: 4 minutes
NHS England
505,000people
planned users
planned users: 505,000 people
NHS England
30,000people
trial users
trial users: 30,000 people
NHS England
43minutes
administrative time saved per user daily
administrative time saved per user daily: 43 minutes
Virtua Health
80percent
sepsis identification improvement
sepsis identification improvement: 80 percent
Virtua Health
1day
average patient stay reduction
average patient stay reduction: 1 day
City of Johannesburg
95percent
false positive reduction
false positive reduction: 95 percent
City of Johannesburg
46.7percent
security operations productivity gain
security operations productivity gain: 46.7 percent
Public Service Commission of Canada
90percent
travel mailbox email reduction
travel mailbox email reduction: 90 percent
Public Service Commission of Canada
5days
complex query time before
complex query time before: 5 days
Public Service Commission of Canada
30seconds
complex query time after
complex query time after: 30 seconds
Public Service Commission of Canada
120,000CAD
annual salary cost avoidance
annual salary cost avoidance: 120,000 CAD
Genentech
5years
manual biomarker validation effort expected to be saved
manual biomarker validation effort expected to be saved: 5 years
ASAPP
91percent
first-call resolution
first-call resolution: 91 percent
ASAPP
77percent
cost per chat reduction
cost per chat reduction: 77 percent
ASAPP
3x
complex interactions handled concurrently
complex interactions handled concurrently: 3 x
ASAPP
49percent
self-service engagement growth
self-service engagement growth: 49 percent
Sun Life
10,000queries
queries answered weekly
queries answered weekly: 10,000 queries
From intelligence to institutional decision
Build the economic case before committing capital.
Model task volume, human baseline cost, runtime consumption, review, failure economics, implementation investment, revenue uplift and risk avoidance before the institution scales AI spend.