Pricing
Your platform subscription includes general Lemon assistance and Lemon Elicit, including preparation, interviews, review learning, and proposed standards. Lemon Weave, Code, and Goal use metered capacity, alongside Trialground execution and rubric evaluation.
Editions
Both editions provide the same alignment capabilities. Enterprise Critical adds an agreed operating environment for sensitive AI development.
Teammately Credits
The complete platform for bringing enterprise expertise into AI development.
$2.00 / TC
USD per Teammately Credit
Includes
Teammately Credits
For AI development with mandatory restrictions on data processing, network access, and operational access.
$3.00 / TC
USD per Teammately Credit
Everything in Enterprise, plus
AI calls use only approved providers and processing locations.
Execution connects only to approved destinations.
Teammately access follows agreed approval and logging requirements.
Critical’s configuration and controls are confirmed before activation.
The same work consumes the same Credits in both editions. Platform subscriptions cover the agreed deployment and operating commitments. Dedicated infrastructure and custom implementation are quoted separately.
Proposed pricing · September 2026. Rates and profiles take effect only when incorporated into a signed Order Form. Platform fees and authorized external inference are separate. Availability depends on the contracted configuration.
Included Lemon and Elicit
Core workflows and their supporting model, speech, transcription and execution resources are included within your subscribed deployment.
No native Credit consumptionWhen paid Credits run out, Included Lemon and Elicit continues within your contracted scope.
Usage rates
Credits are the shared purchasing unit. Each service keeps its own clearly defined meter, with rates and scope agreed before work starts.
One newly generated case version, including its requested inputs, metadata and artifact bundle. Text, image, document, workbook, presentation, codebase and mixed-artifact cases each have a scoped profile.
Construction included. Planning, internal drafts, asset generation, ordinary repairs, rendering and construction checks are part of the case price. Included artifacts are not billed again as world assets.
| Case profile | Profile defines |
|---|---|
| Text | Context and output bounds, schema, constraints and generation mode |
| Image | Image count, resolution, reference inputs and text context |
| Document | Document and page bounds, layout and cross-file requirements |
| Workbook | Sheets, populated cells, formulas and validation |
| Presentation | Deck and slide bounds, layout and included images |
| Codebase | Runtime, repository scope, dependencies and required build/tests |
| Mixed artifact | Explicit manifest, asset allowances and cross-artifact checks |
Managed selection uses a fixed profile rate within the approved model policy. A customer-pinned model has its own fixed profile. Customer-funded inference, where supported, uses a separately quoted software-only case rate.
Count conforming delivered cases, not internal attempts. A changed brief or a requested alternative after conforming delivery may create a new paid case. Delivery does not guarantee expert approval or benchmark performance.
02 / Lemon Weave
Content-only worlds are priced by a manifest of generated logical assets. Executable worlds add a scoped environment build for integrated state, tools, behavior, reset and validation.
Each component counted once. Assets included in a build or bundle are not charged again. Reusing, reloading or resetting an existing world has no new synthesis charge.
| Component | Billing unit |
|---|---|
| Text content | Delivered logical item or bounded content pack |
| Images | Delivered image |
| Audio | Delivered duration or quoted clip bundle |
| Documents | Delivered bounded document or pack |
| Workbooks and presentations | Delivered bounded workbook or deck |
| Structured state | Delivered approved record pack or state snapshot |
| Code | Delivered functional module or scoped repository |
| Executable environment | One assembled world version under an approved build profile |
Total Credits cover the scoped build, when needed, plus new assets not already included. A manifest can instead have one fixed world quote. Consistency checks and internal build validation are included; later customer harness execution uses Trialground runtime.
Storing assets together does not create an environment-build fee. New variants charge only newly synthesized components; behavior or schema changes may require a quoted build revision. Complex unsupported worlds require a fixed-scope discovery/build quote.
One output or trajectory, evaluated against one rubric version, for one requested repetition. Pairwise evaluations use one output pair against one rubric version per requested repetition.
Judging included. Approved judge inference, internal voting and retries, context preparation and judging-service execution are included. There is no extra judge-token or judge-worker charge.
Passing and failing completed judgments both count. Skipped, invalid, incomplete or duplicate judgments do not. Ten rubric judgments in one model call count as ten units; several internal calls for one judgment still count as one unit. Customer-requested reruns count again.
Judge model or model policy, context and media bounds, rubric format and judgment protocol are agreed in the profile. Long-context, multimodal and pairwise work can require different rates. Target-model inference remains separate.
04 / Trialground
Pay for the time ready workers are allocated to authorized harness work, including agreed setup. Time is summed across workers, with no hourly minimum.
Clear allocation boundaries. Queueing and Teammately provisioning are excluded. Allocated workers waiting for an external API remain billable. Runtime stops on pause, release or termination.
| Profile | vCPU / GiB | TC / worker-hour | TC / worker-second |
|---|---|---|---|
| Small | 2 / 4 | 0.1500 | 0.15 ÷ 3,600 |
| Medium | 4 / 8 | 0.3000 | 0.30 ÷ 3,600 |
| Large | 8 / 16 | 0.6000 | 0.60 ÷ 3,600 |
GPU and dedicated configurations require an accepted resource profile. Allocated time rounds up once per allocation to whole seconds. Target inference is customer-funded or separately reimbursed at authorized allocable provider cost, outside Credits and edition premiums.
*Proposed Small shared CPU rate: 2 vCPU / 4 GiB. Availability and rates require an activated profile.
Interactive harness development uses a separately agreed Lemon Code tariff for input, output, and applicable cache tokens. The tariff includes the Code software charge and identifies whether inference is managed or customer-funded.
The conversation and its experiments have distinct meters. Trialground execution and rubric evaluation, or newly synthesized Lemon Weave material, use their respective rates when authorized. General Lemon assistance and Lemon Elicit remain included in the platform subscription.
| Work | Pricing basis |
|---|---|
| Coding-agent input, output, and cache | Accepted Lemon Code model and configuration tariff |
| Harness trials and rubric evaluation | Applicable Trialground runtime and evaluation profiles |
| New cases or worlds | Applicable Lemon Weave profile |
A Lemon Goal optimizer tariff does not establish the price of Lemon Code. The accepted Code profile defines its numerical rates, model route, and included resources before use.
Authorized work can consume usage without producing an accepted change. Agree on the scope and spending limits before starting; defects and unauthorized work are handled under the service terms.
Optimizer planning, analysis, mutation and search reasoning consume input, output and cache tokens. The Lemon Goal software premium is already included in the tariff.
One meter for each kind of work. New cases and worlds use Lemon Weave rates. Benchmark and confirmation runs use Trialground evaluation and runtime rates. Included Lemon and Elicit assistance remains included.
| Model | Input | Output | Cache read | 5-min write | 1-hour write |
|---|---|---|---|---|---|
| Claude Haiku 4.5 | 1.000 | 5.000 | 0.100 | 1.250 | 2.000 |
| Claude Sonnet 5 | 2.000 | 10.000 | 0.200 | 2.500 | 4.000 |
| Claude Opus 5 | 5.000 | 25.000 | 0.500 | 6.250 | 10.000 |
These are proposed Teammately service tariffs, not provider prices or a promise of model availability. Rates require an activated exact model, provider and configuration profile. Supported open-weight and other routes require an accepted tariff.
Tokens are prorated, not rounded to millions. Input and cache categories are exclusive for a given token occurrence. Additional priority, region, tool or media categories need an accepted rate before use. The edition price applies once, with no additional optimizer percentage.
Customer-funded optimizer inference requires a quoted software-only token tariff that replaces the managed-inference tariff. Authorized unsuccessful exploration can consume Credits; improvement is not guaranteed.
Billing, explained
Choose Critical when your organization requires Teammately to enforce approved AI-processing locations and providers, network destinations, or operational-access procedures beyond the standard managed environment. We confirm the supported configuration, controls, and costs before activation.
Your platform subscription is quoted separately. Native usage charges are the sum of each billable quantity multiplied by its agreed Credits per unit, then multiplied by your contracted USD per Credit. The edition price applies once. Authorized target-model inference and separately contracted services are additional.
Case, world and rubric-evaluation profiles need numerical rates and scope limits agreed before work starts. Unsupported configurations require a quote; they do not automatically switch to token billing. Credit prices and profile rates are fixed for the agreed rate term, with no retroactive repricing. Listing a model or service does not establish its availability.
Paid services share one Credit pool. Project and contribution-plan budgets authorize work within that pool; reservations are not consumption. Paid dispatch stops when capacity is exhausted, while Included Lemon and Elicit assistance continues within the subscribed scope. Top-ups require authorization. Expiry and rollover apply only as expressly agreed.
General Lemon assistance and Lemon Elicit’s interviews, review questions, explanations, preparation, and learning are included. A newly synthesized case delivered during expert review is paid Lemon Weave work under an approved contribution-plan budget. Its metered status must be identified before generation. The same delivered-case profile has the same rate across Case Foundry, expert review and training augmentation.
Lemon Weave charges for conforming delivered cases and agreed world deliverables, not discarded drafts or failed construction attempts. Trialground counts completed applicable rubric judgments, including passing and failing results; internal retries, skipped rubrics and duplicate deliveries do not add evaluation units. Reuse has no new synthesis charge. Requested new variants, reruns or evaluations can be metered. Charges affected by Teammately defects, duplicate or unauthorized work are waived or reversed.
Where supported, customer-funded generation, judging or optimization requires a separately quoted software-only rate that replaces the corresponding managed-inference rate. You pay your provider separately. A target-model API key does not remove Trialground judging or runtime charges. Teammately budgets do not cap third-party provider invoices.
Yes. Authorized optimizer tokens, benchmark evaluations and allocated harness runtime can be consumed during valid unsuccessful exploration. Campaigns define spending limits and stopping rules in advance. There is no fee per candidate, success percentage or amount of improvement. Service defects and unauthorized execution are not treated as exploration.
Completed deliverables and authorized metered work already consumed settle under the agreed stopping rules. Unused reservations are released. Undelivered fixed-price cases do not become token charges. Authorized runtime consumed by a customer harness error may still be billable.
Human expert services, bespoke implementation, fine-tuning compute and custom development require separate agreement. Dedicated capacity, storage, egress, deployment requirements and service-level commitments are also defined in the contract. Target-model inference is customer-funded or reimbursed at explicitly authorized allocable provider cost, outside Credits and edition premiums.
Platform scope
Scope the expert work, coverage, and experimental capacity your alignment program needs.
Discuss the expert groups, benchmark programs, managed trials, and deployment requirements involved. We will work through that scope with you before proposing an engagement.
Pricing basis: TC-2026-09-R3 · September 2026 · All monetary amounts in USD. The signed Order Form defines subscription scope, rates, capacity and applicable terms.