Pipeline cost model — verified pricing, September 2026¶

Date: 2026-09-07 · Source: RunPod pricing page (runpod.io/pricing), Modal pricing (spheron.network comparison), GitHub Actions docs Decision status: Provider unit-price inputs remain useful. ADR-011 makes these infrastructure costs inputs to accepted-outcome economics, not the product's final cost metric. The fixed Phase 1b and Phase 3 matrices below are superseded by evidence-gap scheduling and retained only as historical sensitivity scenarios; they are not execution commitments. Rule: every number below has a source. No estimates without labels.

Outcome-economics contract (ADR-011)¶

Provider price, GPU-hour, per-token, per-image, per-audio/video-second, per-request and cost per attempted task answer different questions. The canonical records retain the provider's raw billed quantity, unit, tier and price snapshot and report them separately:

• attempt_cost: endpoint/API usage, judge usage, tool/provider charges and attributable infrastructure for one attempt;
• observed_cost_per_accepted_outcome: total attributable cost of all included attempts, including failures and retries, divided by accepted outcomes;
• observed_time_per_accepted_outcome: the protocol's declared elapsed-time boundary divided by accepted outcomes, with parallel execution disclosed;
• expected_cost_or_time_to_acceptance: a prediction with its probability/retry model, calibration scope and uncertainty, never relabeled as an observation;
• steady_state_solution_cost: startup/transfer, replicas, idle capacity, utilization, interruption, storage and availability costs over the accepted serving horizon.
• market_equivalent_cost: reproducible public/list or contracted price for the same execution boundary before project-only credits;
• gross_attributable_execution_cost: the full cost assigned to the attempt before subsidy;
• subsidy_or_credit_applied: provider/model-lab contribution applied to this attempt, with contributor and terms relevant to interpretation;
• project_out_of_pocket_cost: what the Apron project actually paid after that contribution.

Managed and self-hosted candidates are comparison-eligible only when task/application acceptance, serving horizon, availability target, included cost components and currency/time snapshot match. Provider-opaque charges or infrastructure fields remain unknown rather than being normalized away. Artificial Analysis Optima's cost/time per task is useful evaluation evidence, but Apron additionally retains rejected outcomes and reports accepted-outcome economics when the protocol defines acceptance. A project-only grant never enters solution ranking as a user discount unless the accepted user has the same durable entitlement. Sources reviewed: https://artificialanalysis.ai/optima and https://artificialanalysis.ai/methodology/coding-agents-benchmarking/.

MaintainerBaselineAllocation and ContributedResourcePool are distinct. Before Phase 1a execution, the accepted run plan is priced against current provider quotes and must fit the owner-controlled baseline with the contributed pool set to zero. Contributed credits, scoped credentials, quota or capacity can reduce project spend or expand evidence breadth; they cannot reduce evidence obligations, change candidate ranking or become assumed future funding (ADR-010, INV-29).

GPU rental pricing (per-second billing, both providers)¶

RunPod — current as of September 2026 [W: runpod.io/pricing]¶

Billing: per-second. A 16-minute job on a $1.99/hr GPU costs $0.53, not $1.99.

Community Cloud is cheaper but unvetted datacenters. Secure Cloud is vetted. The concept uses Secure Cloud pricing for cost estimates (conservative).

Modal — current as of 2026 [W: spheron.network/blog/modal-gpu-pricing-2026]¶

Modal: per-millisecond billing, no idle charges. More expensive than RunPod for sustained work. Better for short bursts (<5 min). No RTX consumer GPUs available.

Which provider for what¶

Storage pricing [W: runpod.io/pricing]¶

Weight pre-staging: upload via RunPod S3 API (no pod needed, $0 compute). Models are downloaded once to the volume and reused across boots.

Estimated storage for 20 models: ~500 GB = $35/month on RunPod standard.

Free infrastructure [W: verified sources]¶

Superseded Phase 1b fixed-matrix sensitivity scenario (20 models × 3 GPUs)¶

Assumptions: - 60 model×GPU combinations. Planning assumption: ~40% may be rejected as provably infeasible by the free static prediction, leaving 36 GPU candidates. A passing prediction is not vLLM validation. - Weights pre-staged to RunPod network volume via S3 API before GPU instances start. - Boot + benchmark: ~15 min per feasible combination (includes vLLM startup, profile_run, short benchmark). - Failure injection: 6 classes × 4 attempts average = 24 additional boots. - Diagnosis convergence: ~10 extra boots for agent-proposed corrections on hard failure classes. - Using RunPod Secure Cloud pricing (conservative).

With Community Cloud pricing (4090 at $0.34/hr instead of \(0.74/hr): total drops to ~\)50.

Phase 1b fits within $100 comfortably. The $150-200 revised budget has room for additional GPU types (L4, L40, A6000, 3090) and more benchmark presets.

Superseded Phase 3 fixed-matrix monthly sensitivity scenario¶

Assumptions: - 2 vLLM releases per month (biweekly cadence). - Top 20 models × 5 GPU types per release = 100 boots per release. ~50% infeasible = 50 feasible boots per release. - 100 feasible boots/month total. - Storage stable at ~500 GB (models accumulate slowly, old ones pruned).

At $200/month budget: covers the full matrix PLUS additional GPUs (L4, L40, 3090, A6000) and extended benchmark presets.

At $100/month budget: covers top 15 models × 3 GPUs per release. Storage + reasoning = $41 fixed, leaving $59 for ~80 boots.

At $50/month budget: covers top 5 models × 2 GPUs per release. Storage + reasoning = $41 fixed, leaving $9 for ~12 boots. Tight but functional.

At \(0/month budget: Tier 0 still runs (\)0). Knowledge base still serves (\(0). Calculated verdicts still work (\)0). Issue scanning still runs ($2-4 LLM cost — the only nonzero item). No new measured records until budget resumes.

Historical fixed-matrix budget sensitivity¶

This table describes the superseded matrix assumptions above. Under D10, a budget constrains the evidence-gap scheduler; it does not promise a fixed number of models, boots, or GPU types. Operative coverage is reported by claim scope and exact execution fingerprint.

Sources¶

• RunPod pricing: runpod.io/pricing — fetched 2026-09-07
• RunPod storage and S3 API: docs.runpod.io/storage/network-volumes
• Modal pricing: spheron.network/blog/modal-gpu-pricing-2026 — fetched 2026-09-07
• GitHub Actions free tier: cicdcalculator.com/github-actions-free-tier — public repos unlimited
• RunPod per-second billing: hivenet.com/post/runpod-pricing
• Provider-resource mechanisms, used only to establish that contributed capacity exists and never assumed in the baseline: Baseten startup program, Modal startup program, Together AI startup accelerator, RunPod startup program