A / PRICING
Replace the dated shortlist
The October 2025 table cited no sources. This revision uses named models and official prices checked on 22 October 2026. The original rates are not carried forward.
Decision brief / 22 October 2026
A smarter model
for AI economics.
Can ESIN CORE deliver 100 useful interactions for $1 per user per month? Test the assumptions behind the OUKITEL device proposition.
Explore the economics ↗09models · 5 providers
100 tasksthe original monthly target
Self-containedworks offline; sources open online
01 / IMAGEFIRM INTELLIGENCE
What needed to change
The source page’s arithmetic was broadly consistent with its own rates. The business conclusion was less defensible.
B / MEASUREMENT
Count the whole interaction
200 input tokens cover only a very short request. System instructions, chat history, retrieved content and extra model calls can dominate billed usage.
C / ECONOMICS
Separate headroom from profit
The unspent part of a $1 allowance is budget headroom. It is not profit unless revenue, hosting, support, taxes, commercial fees and other costs are reconciled.
Scope preserved: ESIN CORE on OUKITEL devices; education, productivity and entertainment in Paraguay, with Spanish and Guaraní. These are project assumptions from the original file, not independently verified partnership or deployment claims.
Arithmetic clarification: the original estimate of $0.003 per prompt gives $0.30 for 100 prompts. Adding 10–20% produces $0.33–$0.36, not $0.60. A higher figure needs an explicit change in workload or additional charges.
02 / IMAGEFIRM INTELLIGENCE
Compare the price of a task
Standard paid API text rates. This is a price comparison, not a quality ranking. Equal token counts do not imply equal language coverage or task success.
Price basis: 22 October 2026 · standard paid text API · uncached
| Model / provider | Input / 1M | Output / 1M | USD / task | Tasks / $1 | Evidence |
|---|---|---|---|---|---|
| GPT-6 LunaOpenAIStandard | $0.10 | $0.50 | $0.000170 | 5,882 | [1] ↗ |
| Claude Haiku 5.5AnthropicInput ≤100k tokens | $0.10 | $0.50 | $0.000170 | 5,882 | [2] ↗ |
| Mistral Large 3MistralEstablished GA version · v25.12 | $0.50 | $1.50 | $0.000550 | 1,818 | [4] ↗ |
| Gemini 3.5 Flash-LiteGoogleStandard paid | $0.30 | $2.50 | $0.000810 | 1,234 | [3] ↗ |
| Gemini 3.8 FlashGoogleIntroductory through 31 Dec 2026 | $0.75 | $3.75 | $0.001275 | 784 | [3] ↗ |
| GPT-6.1 SolOpenAIStandard | $2.00 | $10.00 | $0.003400 | 294 | [1] ↗ |
| Claude Sonnet 5.5AnthropicStandard | $2.00 | $10.00 | $0.003400 | 294 | [2] ↗ |
| Command ACoherecommand-a-03-2025 | $2.50 | $10.00 | $0.003500 | 285 | [5] ↗ |
| GPT-6 AstraOpenAIStandard | $10.00 | $50.00 | $0.017000 | 58 | [1] ↗ |
Reference task = 200 uncached input + 300 total billed output tokens; one API call. No retries, tools, reasoning beyond those 300 tokens, operating costs or reserve. Tasks per $1 are rounded down. All prices in USD; 1M = 1,000,000 tokens.
Known future change: Gemini 3.8 Flash moves from $0.75 / $3.75 to $1.50 / $7.50 per million input / output tokens on 1 January 2027 [3]. The future switch applies only that published change; all other prices stay fixed as an assumption. It is not a market forecast.
03 / IMAGEFIRM INTELLIGENCE
Build a defensible budget
All defaults below are planning assumptions, not measured ESIN CORE usage. Choose a preset, then replace them with pilot data.
Controlled pilot · standard uncached pricing checked 22 October 2026.
Sensitivity to monthly usage
Same current inputs and routing; only completed tasks per active user change.
| Tasks / month | User spend | Headroom after reserve | Fleet spend incl. fixed |
|---|
Show the calculation method
I = query + system + history + retrieved content
O = visible output + other billed output
Rin = Σ(route share × input rate); Rout = Σ(route share × output rate)
Ctask = [(I × Rin + O × Rout) ÷ 1,000,000] × calls × (1 + overhead/100) + toolCalls × toolPrice
Cuser = tasks × Ctask + operating allowance
Reserve = budget × reserve%
Headroom = budget − reserve − Cuser
Capacity = max(0, floor((budget − reserve − operating allowance) ÷ Ctask))
Active = floor(devices × active% / 100)
Fleet = Active × Cuser + fixed cost
Fleet headroom = Active × (budget − reserve) − Fleet
Fixed allocation = fixed cost ÷ Active (only when Active > 0)
Fleet capacity/user = max(0, floor((budget − reserve − operating allowance − fixed allocation) ÷ Ctask))
O = visible output + other billed output
Rin = Σ(route share × input rate); Rout = Σ(route share × output rate)
Ctask = [(I × Rin + O × Rout) ÷ 1,000,000] × calls × (1 + overhead/100) + toolCalls × toolPrice
Cuser = tasks × Ctask + operating allowance
Reserve = budget × reserve%
Headroom = budget − reserve − Cuser
Capacity = max(0, floor((budget − reserve − operating allowance) ÷ Ctask))
Active = floor(devices × active% / 100)
Fleet = Active × Cuser + fixed cost
Fleet headroom = Active × (budget − reserve) − Fleet
Fixed allocation = fixed cost ÷ Active (only when Active > 0)
Fleet capacity/user = max(0, floor((budget − reserve − operating allowance − fixed allocation) ÷ Ctask))
This is an average-cost model with identical token loads across routes and a proportional cost uplift for extra billed model work. Averages cannot describe the full range of costs for individual users. “Task” means one user interaction; it can require multiple model calls. The 100,000 total-input and 8,000 total-output limits keep the calculator in the scope of this short-task comparison. Larger workloads need model-specific context, output and tariff checks. Report actual costs from provider billing records.
04 / IMAGEFIRM INTELLIGENCE
Where the cost really moves
01 / CONTEXT
History compounds
A 200-token question may carry thousands of earlier tokens. Keep summaries and retrieved passages concise. Measure the complete billed input, including any carried reasoning context.
02 / OUTPUT
The visible answer is only part
Some models bill hidden reasoning as output [7]. Set task-specific output limits, then check answer completion and usefulness. A shorter answer is a saving only when the task still succeeds.
03 / CACHING
Measure reads, writes and storage
A cache discount is not automatic savings on every token. Minimum prefix lengths, reuse, cache-write fees and storage charges matter [2][3][6]. The calculator intentionally uses uncached rates.
04 / TOOLS
Budget beyond the model
Search, retrieval, voice, images and orchestration may add charges. This calculator covers text and your tool allowance; voice and image generation require a separate workload estimate.
Why not use maximum context as a value score?
Context capacity is a technical limit, not the number of tokens every call consumes. Compare cost per accepted task, response time and language quality at the workload you actually serve.
What happened to the original 30 / 50 / 20 mix?
The source did not specify shares within each model tier, so its weighted average was not reproducible. The original-workload preset applies 20% routine, 50% general and 30% escalation to explicitly named current models. It removes additional history, output, tools and service allowances for a token-only comparison; fleet assumptions remain unchanged. It does not reconstruct the historical bill.
Can free access continue after the budget is exhausted?
Only if someone funds that usage. Define a hard monthly allowance, a lower-cost fallback and clear limits. A free tier or rate limit is not a production funding plan.
Are Spanish and Guaraní equally economical?
Do not assume equal tokenization or quality. Test natural Spanish, Guaraní and mixed-language prompts with native reviewers. Cohere’s Command A documentation lists Spanish among its trained languages but does not list Guaraní [5]. That is not proof that other models meet the requirement.
05 / IMAGEFIRM INTELLIGENCE
Turn the estimate into evidence
Suggested pilot design, not a claim about existing performance. Agree the quality and spend gates before broader rollout.
WEEK 01
Instrument the complete request
Log model version, route, billed input/output, tool fees, latency and task outcome. Use aggregated telemetry; avoid storing raw personal prompts by default.
WEEK 02
Evaluate local usefulness
Start with 100 reviewed tasks per language as a practical pilot sample, not a statistical guarantee. Cover schoolwork, writing, translation and everyday questions. Measure errors and successful completion, not just fluency.
WEEKS 03–04
Validate the usage tail
Pilot with a limited cohort. Compare median, p95 and p99 user spend and daily peaks. Re-run the January price scenario. Scale only after quotas, fallback behavior and unit costs hold.
Launch readiness checklist
Decision metrics that matter
| Metric | Why it matters |
|---|---|
| Cost / accepted task | Captures failures and retries |
| p95 spend / active user | Exposes expensive usage tails |
| Language acceptance rate | Prevents cheap but unusable answers |
| Escalation + fallback rate | Tests whether routing assumptions hold |
| p95 response time | Connects cost to user experience |
Set numeric gates after the first measured baseline; the source file provides no validated quality, latency or usage distribution. Keep the $1 allowance separate from any selling price.
06 / IMAGEFIRM INTELLIGENCE
Evidence, assumptions & scope
Official sources checked on 22 October 2026. Recheck the linked tariffs before procurement or renewal.
[1] OpenAI API pricing ↗GPT-6 family · standard input / output[2] Anthropic pricing ↗Haiku 5.5 threshold · Sonnet 5.5 · cache economics[3] Google Gemini API pricing ↗Flash-Lite 3.5 · Flash 3.8 · scheduled price change[4] Mistral Large 3 model card ↗Named GA version and API rates[5] Cohere Command A ↗Model ID · rates · listed languages[6] OpenAI prompt caching ↗Eligibility, write costs and reuse[7] OpenAI reasoning models ↗Billed reasoning tokens and output limits
Reading this analysis: published tariffs are sourced facts; workload, routing, operating allowances and the pilot design are ImageFirm planning assumptions; calculator outputs are derived estimates. No model-quality benchmark or API execution test was performed; this is a researched planning tool. The shortlist is selective, not exhaustive. API access, regional terms and rate limits must be checked for the actual deployment.
Rebuilt from Token_Cost_Breakdown.html, originally dated 18 October 2025. Consumer chat subscriptions are not used as API pricing. Tax, currency conversion, payment/partner fees and unentered expenses are excluded. Valid calculator settings and checklist progress are saved only in this browser when supported. No data is sent to an API. Export CSV for a portable record with assumptions, formulas, model prices and source URLs. The table uses an archived price snapshot; it does not fetch live tariffs.