LLM API Pricing Comparison
Compare LLM API prices only after matching routes, units, workloads, and acceptance criteria, while separating official rates from measured evidence.
Published: March 20, 2026 · By the QVeris Team
TL;DR
Fix prompt distribution, context, tools, modalities, latency and quality gates.
Map units for comparison but keep every provider-native billing field.
Store source URL, retrieval time, currency, region, model and route.
Calculate total delivered cost after retries, tools and quality failures.
Why simple price tables fail
Providers may separate input, output, cached input, cache writes or storage, batch, flex, priority, long context, media and tool charges. Access paths can add platform fees, commitments, taxes or regional differences.
The billing words are not always equivalent. OpenAI currently separates short- and long-context rates and multiple service tiers. Claude distinguishes base input, two cache-write durations, cache reads, Batch, inference geography, and server-side tools. Gemini prices models, modalities, processing tiers, grounding, and tools under different rules. DeepSeek separates cache-hit and cache-miss input. Normalize each native field, but do not erase what it means.
A model row is also not a quality-equivalence claim. Different models may require different prompt lengths, reasoning settings, output tokens, tool retries, or human correction to finish the same task. A candidate that needs more attempts or review can cost more per accepted outcome despite a lower listed token price.
Comparison layers
| Layer | Best fit | Verify before choosing |
|---|---|---|
| Official list price | Build a timestamped reference for exact models and routes. | Units, thresholds, currencies, tiers, regions and special modes. |
| Measured usage | Estimate each route with representative workload evidence. | Native usage, cache, tools, media, retries and failed calls. |
| Accepted-output cost | Compare delivered work after a shared quality gate. | Sample size, evaluator version, review cost and uncertainty. |
| Total commercial cost | Support procurement and long-term operating decisions. | Support, SLA, commitments, credits, taxes, data terms and engineering. |
Comparison worksheet fields
Provider, platform, model version, endpoint, region and processing mode.
Input, output, cache, batch, tools, modalities, storage and extras.
Latency, retry rate, valid schema, quality acceptance and review effort.
Official URL, timestamp, owner, currency, tax basis and expiry review.
Create a repeatable comparison
Use a workload manifest that records the task mix, prompt and context distribution, expected schema, permitted tools, latency target, factuality and safety rubric, and human-review rule. Freeze that manifest for one comparison run. Otherwise a prompt improvement or easier sample can be mistaken for a price advantage.
- Version the workload, acceptance rubric, prompt, tool schemas, SDK, and provider configuration.
- Fetch current official pricing and terms, retaining source URL, retrieval time, currency, region, route, model ID, context band, and service tier.
- Run candidates against the same sample and capture every native request ID, usage field, cache event, tool charge, retry, timeout, and failed attempt.
- Apply the shared quality gate before calculating cost per accepted result and per completed user workflow.
- Reconcile measured usage to invoices and report discounts, credits, taxes, platform fees, and uncertainty separately.
- Publish the date, scope, assumptions, excluded features, sample size, confidence, owner, and next review trigger.
Join price snapshots to workload runs
A source registry stores official pricing snapshots. A workload runner creates traceable native-usage events and acceptance results. The comparison layer normalizes units, adds retry and tool costs, applies the quality gate and reports cost per accepted output.
Production rule: never mix price data from different dates, model versions or commercial routes without labeling the difference.
Compare external capability costs on their own ledger
QVeris discovers and calls external APIs, tools, services and live data. Their fees may use requests, records, compute time or subscription units rather than tokens. Keep the native units, then join end-to-end cost at the workflow level.
A dated API price snapshot for workload tests
Verified against official pricing pages on July 21, 2026. Prices are USD per 1 million tokens and are not quality-equivalent recommendations. Use them to seed a workload calculator, then refresh before procurement or publication because providers can change models and rates.
| Model or route | Input / cost | Output | Scope |
|---|---|---|---|
| OpenAI GPT-5.6 Luna | $1.00 | $6.00 | Cached input $0.10/MTok |
| Claude Haiku 4.5 | $1.00 | $5.00 | Cache read $0.10/MTok |
| Gemini 3.1 Flash-Lite | $0.25 | $1.50 | Standard text/image/video input |
| DeepSeek V4 Flash | $0.14 | $0.28 | Cache-miss input; cache hit $0.0028 |
Always verify the linked official pricing source immediately before a buying decision. Model quality, output length, cacheability, retries, tools, service tier, region, taxes, and discounts can reverse a token-price comparison.
Continue with related implementation guides
Use these adjacent guides to compare the next architectural decision, implementation boundary, or operating control without starting the research again.
FAQ
Only as a dated snapshot with an owner and review process; prices change.
Cost per accepted workload result, supported by full billing and quality evidence.
Yes when they materially affect production feasibility or total cost.