Compare API token and GPU rental budgets using dated source rates and your own workload.
Price observations: 21 October–1 November 2026
Ads are currently disabled. Optional analytics loads only with your consent and excludes your calculator entries; use Cookie settings to change your choice. Selected reviewed catalog: 10 generative models, 1 text-embedding model and 3 single-GPU configurations. Observations span 21 October–1 November 2026 (UTC); each row shows its own collection date. This is a dated snapshot, not a live feed or a complete market survey. No provider account or prompt is required.
01 / API usage
API token budget
Include all billable input and output tokens, including history, tools and reasoning.
PROCESSING MODES
Compare batch with standard processing
Use the selected API model and token/call assumptions above. This comparison replaces the API budget’s batch percentage with the share below. Cache reads, writes and storage must be zero unless their exact batch compatibility is verified. Batch eligibility, completion time and operational suitability require provider confirmation.
Batch method and example
Standard cost uses the model’s standard input/output rates. The alternative mixes standard and separately sourced batch rates according to your eligible-call share, then adds the entered monthly workflow cost. Saving = standard cost minus alternative total; percentage saving = saving ÷ standard cost × 100. If standard cost is zero, the percentage is unavailable. Cost-parity monthly calls = fixed workflow cost ÷ per-call processing saving, rounded up; unavailable when processing provides no saving. This assumes constant token counts, rates and eligible share as volume changes.
For 1,000 input/200 output tokens, 100 calls/day and 30 days, USD 2/10 standard rates give USD 12/month. Separately verified USD 1/5 batch rates give USD 6. Add USD 1.50 workflow cost: total USD 7.50, saving USD 4.50 (37.5%), with cost parity at 750 calls/month. This is an arithmetic example, not a universal batch discount or recommendation.
02 / Pricing observations
Compare model rates
USD per million tokens at standard context length; taxes, residency modifiers and tool charges excluded. Verified model context/output limits are enforced. For OpenAI prompts above 272,000 input tokens, the calculator applies the documented full-request long-context premium. Equal token costs do not imply equal capabilities.
Rates below are per million tokens in the displayed currency. Embedding vectors have no generated-text output price. Context sizes have different input/combined scopes; sorting them does not establish equal capabilities. Pricing-mode filters exclude observations past their estimator cutoff, but do not verify your account or workload eligibility.
Manually verified standard pricing observations
Model
Provider
Input
Output
Cache read
Batch rates
Context / modalities
Source & freshness
gpt-6.1-sol Text generation
OpenAI
USD 2
USD 10
USD 0.1
USD 1 input; 5 output
1,050,000 combined tokens; 128,000 maximum output text input/output; image input Model specifications
Embedding vectors support retrieval/search; they are not generated text. Filter Model task to Text embeddings to inspect the reviewed text-only rates. The estimator excludes image/audio/video/PDF prices, vector storage, retrieval and free-tier quotas. Select the Embeddings workload below to calculate these rates.
03 / Compute
Estimate a GPU rental budget
Select a documented single-GPU offering or enter a custom quote. Different GPUs and interconnects are different configurations; this is a budget comparison, not a performance ranking. Region and live capacity require provider confirmation.
Enter two to ten quotes you obtained from AWS, Azure, Google Cloud or another provider. This tool does not supply or verify their prices. It uses the instance count, runtime, shared storage and other costs in the GPU budget above. Shared charges use its quote-currency selection; each quote’s own extra charges use that quote’s currency.
Keep pricing modes explicit. Spot prices may change and instances may be interrupted. For committed/ reserved quotes, enter an hourly equivalent and include allocated upfront/minimum-commitment costs as quote-specific monthly charges; otherwise the estimate is incomplete. Extra fees are not inferred. Hardware differences remain visible; equal runtime is not equal performance. Unknown hardware fields may stay blank.
Quote comparison method
For each entered quote, round one daily session to its billing interval after applying its minimum. Multiply by active days and instance count, then by the original hourly quote. Per-GPU quotes also multiply by GPUs per instance. Add shared storage/other costs only when their currency matches, plus quote-specific monthly charges. For two instances at USD 2/instance-hour, 8 hours/day and 30 days, compute is USD 960. Add USD 10 storage and USD 5 shared costs: USD 975. A USD 1/GPU-hour quote with two GPUs per instance has the same arithmetic cost; that does not establish hardware or performance equivalence.
All quote inputs stay in this browser. They are not sent to a provider, saved to an account or added to the public price database. Download a CSV to retain your own scenario.
04 / Financial scenarios
Compare API spending with self-hosting
Use the current API workload and GPU quote entered above, even if you have not calculated them separately. Both budgets must use the same currency and represent the same modeled month. Different active-day counts are allowed for each schedule. Enter remaining operating costs once, excluding charges already entered in the GPU budget.
Comparison method and example
Self-hosted cost = rounded GPU compute + persistent storage + other infrastructure + additional operating costs. Difference = API estimate minus self-hosted cost. Productive instance-hours = scheduled hours × days × instances × productive share. Idle-cost allocation includes paid idle time and billing-rounding overhead; changing utilization never discounts the provider bill. With a USD 12 API budget and two USD 2/hour instances running 8 hours/day for 30 days, compute is USD 960. Add USD 10 storage, USD 5 other infrastructure and USD 100 operations: self-hosting totals USD 1,075, a USD 1,063 higher monthly cost. At 50% utilization, 240 productive instance-hours remain and USD 480 of compute is allocated to idle time. These figures establish no throughput or model-quality equivalence.
How the calculations work
For standard calls, cost = uncached input tokens × input rate + cache-read tokens × cache-read rate + cache-write tokens × cache-write rate + output tokens × output rate, divided by 1,000,000. Input shares are disjoint and cannot exceed 100%. Batch uses its separately verified rates for the eligible share of calls. For the OpenAI models listed here, more than 272,000 input tokens triggers 2× input/cache rates and 1.5× output rates for the entire request, including batch estimates.
Mixed batch and cache are rejected until that exact combination is verified. Cache population has a separate cost; a cache hit assumption does not mean building the cache was free. A negative saving means the selected pricing mode costs more.
Cache storage and dated rates
Cache-storage token-hours are separate from cached reads. One million tokens retained for 24 hours means 24 million token-hours. Use zero for implicit caching; for explicit caching also include cache-creation tokens in uncached call usage. Daily and per-call figures cover model calls; monthly totals and savings include any entered storage. Unsupported storage and cache modes are rejected.
Annual figures multiply the current modeled month by twelve. They are annualized estimates, not forecasts across announced rate changes. Gemini promotional rates show an estimator cutoff and stop calculating after that deadline until reviewed. Cutoffs use 00:00 UTC after the published final promotional day; the source does not specify its billing timezone, so confirm the transition with Google. Gemini 3.x limits apply to input separately; Argon and the other listed models use a combined context limit. Verified output limits also apply; an unknown output cap does not establish that a request is supported. Mistral implicit caching is modeled in complete 64-token blocks per call, rounded down from the requested cached-input share. The remainder uses the standard input rate; actual cache hits are not guaranteed.
Worked example
At USD 2 per million input tokens and USD 10 per million output tokens, 1,000 input plus 200 output tokens costs USD 0.004 per call. At 100 calls per day for 30 days, the monthly estimate is USD 12 and twelve identical months cost USD 144. These are arithmetic examples, not offers.
GPU billing
Each daily session is rounded up to the quoted billing interval after applying the minimum session. Multiply billable hours by instances and the price per instance; for a per-GPU quote, multiply by GPUs per instance too. Storage and additional monthly charges are then added.
Assumptions and limitations
Annual estimates mean twelve modeled months, not a 365-day calendar. No exchange-rate conversion is performed. Currency selection labels your manually entered rates. Availability, latency, quality, quotas, minimum invoice charges and special contracts require separate verification. Effective dates, taxes and regional eligibility are not inferred. Stale observations remain visibly dated and should be rechecked before budgeting. Browser checks refresh age labels at least once a minute and on return to this page, and check estimator cutoffs before each budget calculation. Changed freshness invalidates displayed budgets; it does not refresh source prices.
These estimates are informational and do not provide financial, procurement or investment advice. Confirm billable usage and current terms with your provider. No model calls are made.
Sources and updates
Published pricing observations are manually reviewed. The private monitor checks approved Google sources and queues candidate changes for review; other providers remain permission-gated. Checks never automatically update the prices shown here. Three Google generation-model prices and the Mistral Small 4 price were rechecked on 1 November 2026 (UTC); their older specification dates remain unchanged. Mistral Large 4 preview prices were separately reviewed from a new pinned snapshot. Large 3 and the other provider observations retain their original dates. These separate checks do not establish continuous price stability or a historical trend. OpenAI API pricing, Anthropic API pricing, Lambda list prices and Lambda billing rules. Model specification sources appear in the comparison table. Google pricing/specification facts are adapted from Google AI for Developers under CC BY 4.0; changes include normalization and calculation assumptions. No Google endorsement is implied. Last calculation review: 22 October 2026.
models.dev data attribution and MIT license
Mistral price facts use dated, pinned models.dev community snapshots. The earlier Large 3 observation remains historical; Large 4 is a separate public-preview entry. Prices were separately checked against the official USD Standard table. Small 4 and Large 3 context precision remains from the original community snapshot. Large 4 context and output caps are unknown here; unsupported release dates are omitted. Pinned original license. No endorsement.
MIT License
Copyright (c) 2025 models.dev
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
25 approved observations across 14 products, covering 2 collection dates. First observed 2026-10-21; latest observed 2026-11-01. These are the dates we checked sources, not necessarily when prices took effect. No earlier price history is inferred.
Each observation is checked against its original source record and approval before publication. A source-page checksum change alone is not a price change. Automated collection remains limited to separately approved operations. Display-only encoding repairs retain the original records and evidence hashes; CSV exports identify correction hashes separately.
Provider observation coverage
These are reviewed price observations, not the time of a background source check. A single collection date cannot establish a long-term trend.
Coverage of the approved observation ledger
Provider
Products
Reviewed records
Collection dates
Latest reviewed observation (UTC)
Anthropic
2
4
1
2026-10-21T04:39:10.000Z
Google Gemini
4
10
2
2026-11-01T08:19:49.235Z
Lambda
3
3
1
2026-10-21T04:39:10.000Z
Mistral
3
4
2
2026-11-01T20:22:18.951Z
OpenAI
2
4
1
2026-10-21T04:39:10.000Z
All 5 observed providers; 14 products.
Latest recorded comparable price changes
Only a change between reviewed prices under the same recorded terms appears here. New products, changed terms, source-page checksums and future announcements are not historical price changes. Latest 30 matching events are shown.
No comparable price changes recorded yet. This does not establish that provider prices stayed stable.
Latest reviewed changes in observed prices
Product / component
Collected at / effective date
New price
Change
Source
Selected product and component
Only recorded prices appear; gaps are not filled.
Claude Opus 5.5 · batch-input
Collected at (UTC)
Price and unit
Observation
Terms and source
2026-10-21T04:25:59.000Z
2 USD / million-tokens
First observation for these recorded terms
claude-opus-5-5-standard; Standard API; residency not specified; tax unknown Effective date: Unknown Reviewed: 2026-10-21T04:25:59.000Z Long-context rule not recorded; rate cutoff not recorded; minimum cached input not recorded; cache block tokens not recorded Official source
2026-10-21T04:39:10.000Z
2 USD / million-tokens
Price unchanged
claude-opus-5-5-standard; Standard API; residency not specified; tax unknown Effective date: Unknown Reviewed: 2026-10-21T04:39:10.000Z Long-context rule not recorded; rate cutoff not recorded; minimum cached input not recorded; cache block tokens not recorded Official source
No capacity, quality, procurement or investment recommendation is made. New versions, plans, currencies, regions, tax treatment, price units or recorded pricing rules start separate comparisons. Added rule details do not prove a provider changed its terms. A percentage change from a zero price is unavailable.