A software factory.
Bootstrapped. Agent-first. Targeting niche arbitrage opportunities with B2B and B2C SaaS and agent services.
Each product has its own market, pricing, and reason to exist.
The economics
have changed.
Intelligence costs less to run. Getting a product into someone’s hands still takes work. The opening is a specific job, a reachable buyer, and less software between them.
A different cost base.
A GPT-3.5-level benchmark score went from $20 to $0.07 per million tokens. That changes which small, repeated jobs can make sense as products.
FACTORY THESIS / Keep the scope narrow. Price the completed job.GPT-3.5-level inference price
USD per million tokensTwo reported endpoints at a fixed benchmark threshold. Different models; inference price only.
Built for agents.
Usable by people.
Agent-first means the product exposes actions an agent can actually use. People get an interface. Agents get a direct route to the same job. Both need permissions, clear results, and a way to recover when something fails.
Software-as-a-Service gives access to the tool.
Agents-as-a-Service takes on the task.
monthly MCP SDK downloads
Python + TypeScript
What intelligence costs.
Follow the published prices. Compare the companies. Inspect the evidence behind a forecast.
The market, in motion.
Swipe to explore the chart →
A model becomes a building block.
GPT-3 reached the OpenAI API in June 2020. The first dated price sheet plotted here is from August 2022: Davinci at $60 per million tokens.
6 dates · 4 retrospective splits · weighted model ensemble
Model range: $0.73–$2.00. Not a confidence interval. Compute-stress value: $1.53.
The fixed-capability result still matters: Stanford tracked a GPT-3.5-level MMLU score from $20 to $0.07 per million tokens, Nov 2022 to Oct 2024.
Stanford AI Index 2025 ↗Google’s published Gemini 3.8 Flash rates rise from $0.75 / $3.75 to $1.50 / $7.50 when introductory pricing ends on 1 Jan 2027. That announced change is separate from the fitted forecasts.
Google’s published schedule ↗Prices, sources & comparison rules +
A curated history of major text-model families, not every model or hosting offer. Rates are USD per million tokens, standard uncached short-context pricing. Model tokenizers, quality, context tiers, availability, and reasoning-token use differ. Low token prices do not establish a lower cost per successful job. Lines connect selected releases within families; they do not imply continuous availability or an unchanged model. Current catalog snapshots use the observation date, not a guessed release date.
How the forecasts are calculated
Forecasts combine an unchanged-price baseline, an exponentially damped log-linear trend and a robust median-slope trend. Trends need three distinct dates spanning 18 months. Weights use chronological, rolling-origin absolute log errors and shrink toward unchanged pricing when validation is sparse. The damping time is an explicit 18-month assumption. The range shows model disagreement, not a confidence interval. Sparse series use persistence with no invented uncertainty. These curated historical comparisons have selection bias; they do not establish future accuracy or equal model capability.
News and compute covariates are recorded with first-seen timestamps. A regularized cost/news regression is enabled only with 18 complete observations across a year, and is evaluated using past features carried forward, never future observed features. The new collector has insufficient history for that model today. The news row links to published source events; it does not assert a causal effect on prices. Supply, demand, energy and efficiency controls add a separate structural stress overlay; their elasticities are assumptions, not learned facts.
The chart draws one family per company, prioritizing an eligible family with the most dated history, then the longest span. A newer cheap model does not displace a better-supported forecast. A different family can be selected. Direct vendor and OpenRouter routes remain separate series. A source receipt alone does not mean its numeric rows were verified; the inspector and source-health section distinguish current row matches, parsed historical archives, dated announcements, manual observations, failures and stale sources.
For context, Epoch AI’s 2026 research ↗ measures cost at a fixed level of benchmark performance. Its market-wide results are not used as a company’s future token-price discount.
| Family / basis | Dates / splits | Method log errors | 2028 estimate | Model range |
|---|---|---|---|---|
| OpenAI GPT flagship linehistorical ensemble · Anchor: GPT-6.1 Sol | 6 / 4Aug 22, 2022 – Nov 9, 2026 | persistence: 0.811damped-trend: 8.575robust-trend: 0.744 | $1.53 | $0.73 – $2.00 |
| OpenAI GPT small-model linehistorical ensemble · Anchor: GPT-6 Luna | 6 / 4Mar 1, 2023 – Nov 9, 2026 | persistence: 0.922damped-trend: 1.069robust-trend: 1.076 | $0.0707 | $0.0316 – $0.10 |
| OpenAI GPT-5.4 nanopersistence baseline · Anchor: GPT-5.4 nano | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $0.20 | Not calibrated |
| OpenAI GPT-5.6 Lunapersistence baseline · Anchor: GPT-5.6 Luna | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $0.20 | Not calibrated |
| OpenAI GPT-5.4 minipersistence baseline · Anchor: GPT-5.4 mini | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $0.75 | Not calibrated |
| OpenAI GPT-6.1 Terrapersistence baseline · Anchor: GPT-6.1 Terra | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $1.00 | Not calibrated |
| OpenAI GPT-5.1persistence baseline · Anchor: GPT-5.1 | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $1.25 | Not calibrated |
| OpenAI GPT-5.2persistence baseline · Anchor: GPT-5.2 | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $1.75 | Not calibrated |
| OpenAI GPT-5.6 Terrapersistence baseline · Anchor: GPT-5.6 Terra | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $2.00 | Not calibrated |
| OpenAI GPT-6 Solpersistence baseline · Anchor: GPT-6 Sol | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $2.00 | Not calibrated |
| OpenAI o3persistence baseline · Anchor: o3 | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $2.00 | Not calibrated |
| OpenAI GPT-4opersistence baseline · Anchor: GPT-4o | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $2.50 | Not calibrated |
| OpenAI GPT-5.4persistence baseline · Anchor: GPT-5.4 | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $2.50 | Not calibrated |
| OpenAI GPT-5.6 Solpersistence baseline · Anchor: GPT-5.6 Sol | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $4.00 | Not calibrated |
| OpenAI GPT-5.5persistence baseline · Anchor: GPT-5.5 | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $5.00 | Not calibrated |
| OpenAI GPT-6 Astrapersistence baseline · Anchor: GPT-6 Astra | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $10.00 | Not calibrated |
| Anthropic Claude Opus linehistorical ensemble · Anchor: Claude Opus 5.5 | 5 / 3Jul 2023 – Nov 9, 2026 | persistence: 0.441damped-trend: 0.530robust-trend: 0.552 | $3.67 | $2.71 – $4.00 |
| Anthropic Claude Sonnet linehistorical ensemble · Anchor: Claude Sonnet 5.5 | 4 / 2Mar 4, 2024 – Nov 9, 2026 | persistence: 0.203damped-trend: 0.203robust-trend: 0.203 | $1.96 | $1.69 – $2.00 |
| Anthropic Claude Haiku linehistorical ensemble · Anchor: Claude Haiku 5.5 | 3 / 1Jul 2023 – Nov 9, 2026 | persistence: 0.916damped-trend: 2.414robust-trend: 2.414 | $0.0954 | $0.0399 – $0.10 |
| Anthropic Claude Haiku 4.5persistence baseline · Anchor: Claude Haiku 4.5 | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $1.00 | Not calibrated |
| Anthropic Claude Sonnet 4.6persistence baseline · Anchor: Claude Sonnet 4.6 | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $3.00 | Not calibrated |
| Anthropic Claude Opus 4.6persistence baseline · Anchor: Claude Opus 4.6 | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $5.00 | Not calibrated |
| Anthropic Claude Opus 4.7persistence baseline · Anchor: Claude Opus 4.7 | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $5.00 | Not calibrated |
| Anthropic Claude Opus 4.8persistence baseline · Anchor: Claude Opus 4.8 | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $5.00 | Not calibrated |
| Anthropic Claude Opus 5persistence baseline · Anchor: Claude Opus 5 | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $5.00 | Not calibrated |
| Anthropic Claude Fable 5.1persistence baseline · Anchor: Claude Fable 5.1 | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $10.00 | Not calibrated |
| Google Gemini Flash linehistorical ensemble · Anchor: Gemini 2.5 Flash-Lite | 3 / 1Aug 12, 2024 – Nov 9, 2026 | persistence: 0.981damped-trend: 5.492robust-trend: 0.981 | $0.10 | $0.10 – $0.14 |
| Google Gemini 3 Flash Previewpersistence baseline · Anchor: Gemini 3 Flash Preview | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $0.50 | Not calibrated |
| Google Gemini 3.6 Flashpersistence baseline · Anchor: Gemini 3.6 Flash | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $0.75 | Not calibrated |
| Google Gemini 3.8 Flashpersistence baseline · Anchor: Gemini 3.8 Flash | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $0.75 | Not calibrated |
| Google Gemini 2.5 Propersistence baseline · Anchor: Gemini 2.5 Pro | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $1.25 | Not calibrated |
| Google Gemini 3.1 Pro Previewpersistence baseline · Anchor: Gemini 3.1 Pro Preview | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $2.00 | Not calibrated |
| Mistral Mistral Small linehistorical ensemble · Anchor: Mistral Small 4 | 3 / 1Sep 17, 2024 – Nov 9, 2026 | persistence: 0.405damped-trend: 1.161robust-trend: 1.161 | $0.15 | $0.13 – $0.15 |
| Mistral Codestralpersistence baseline · Anchor: Codestral | 2 / 0Sep 17, 2024 – Nov 9, 2026 | persistence: not measured | $0.30 | Not calibrated |
| Mistral Mistral Large linepersistence baseline · Anchor: Mistral Large 4 | 2 / 0Sep 17, 2024 – Nov 9, 2026 | persistence: not measured | $0.70 | Not calibrated |
| Mistral Ministral 3 3Bpersistence baseline · Anchor: Ministral 3 3B | 2 / 0Oct 16, 2024 – Nov 9, 2026 | persistence: not measured | $0.10 | Not calibrated |
| Mistral Ministral 3 8Bpersistence baseline · Anchor: Ministral 3 8B | 2 / 0Oct 16, 2024 – Nov 9, 2026 | persistence: not measured | $0.15 | Not calibrated |
| Mistral Ministral 3 14Bpersistence baseline · Anchor: Ministral 3 14B | 1 / 0Nov 9, 2026 – Nov 9, 2026 | persistence: not measured | $0.20 | Not calibrated |
| DeepSeek DeepSeek general / Pro linehistorical ensemble · Anchor: DeepSeek V4 Pro 0813 | 6 / 4Feb 8, 2025 – Nov 9, 2026 | persistence: 0.699damped-trend: 0.831robust-trend: 0.819 | $1.92 | $1.32 – $4.43 |
| DeepSeek DeepSeek Flash linepersistence baseline · Anchor: DeepSeek V4.1 Flash 1017 | 4 / 2Apr 24, 2026 – Nov 9, 2026 | persistence: 0.191 | $0.30 | Not calibrated |
| xAI Grok general-model linehistorical ensemble · Anchor: Grok 4.7 | 4 / 2Dec 12, 2024 – Nov 9, 2026 | persistence: 0.000damped-trend: 0.000robust-trend: 0.000 | $2.00 | $2.00 – $2.00 |
| Cohere Command R efficiency linehistorical ensemble · Anchor: Command R7B 12-2024 | 3 / 1Aug 30, 2024 – Nov 9, 2026 | persistence: 0.000damped-trend: 0.717robust-trend: 0.717 | $0.0368 | $0.018 – $0.0375 |
| Cohere Command R+ linehistorical ensemble · Anchor: Command R+ 08-2024 | 3 / 1Aug 30, 2024 – Nov 9, 2026 | persistence: 0.000damped-trend: 0.000robust-trend: 0.000 | $2.50 | $2.50 – $2.50 |
| Amazon Nova Micro · AWS us-east-1historical ensemble · Anchor: Nova Micro | 5 / 3Jan 9, 2025 – Nov 9, 2026 | persistence: 0.000damped-trend: 0.000robust-trend: 0.000 | $0.035 | $0.035 – $0.035 |
| Amazon Nova Lite · AWS us-east-1historical ensemble · Anchor: Nova Lite | 5 / 3Jan 9, 2025 – Nov 9, 2026 | persistence: 0.000damped-trend: 0.000robust-trend: 0.000 | $0.06 | $0.06 – $0.06 |
| Amazon Nova Pro · AWS us-east-1historical ensemble · Anchor: Nova Pro | 5 / 3Jan 9, 2025 – Nov 9, 2026 | persistence: 0.000damped-trend: 0.000robust-trend: 0.000 | $0.80 | $0.80 – $0.80 |
| Alibaba Qwen Plus · Singapore / Internationalpersistence baseline · Anchor: Qwen3.7-Plus | 2 / 0Jul 14, 2025 – Nov 9, 2026 | persistence: not measured | $0.40 | Not calibrated |
| Observation | Model | Input | Output | Source |
|---|---|---|---|---|
| Aug 22, 2022Quote | OpenAI GPT-3 Ada | $0.80 | $0.80 | Original source ↗ |
| Aug 22, 2022Quote | OpenAI GPT-3 Babbage | $1.20 | $1.20 | Original source ↗ |
| Aug 22, 2022Quote | OpenAI GPT-3 Curie | $6.00 | $6.00 | Original source ↗ |
| Aug 22, 2022Quote | OpenAI GPT-3 DavinciPre-cut rate quoted in the 22 Aug announcement. One rate for prompt and completion; not a 2020 launch price. | $60.00 | $60.00 | Original source ↗ |
| Sep 1, 2022Price change | OpenAI GPT-3 Ada | $0.40 | $0.40 | Original source ↗ |
| Sep 1, 2022Price change | OpenAI GPT-3 Babbage | $0.50 | $0.50 | Original source ↗ |
| Sep 1, 2022Price change | OpenAI GPT-3 Curie | $2.00 | $2.00 | Original source ↗ |
| Sep 1, 2022Price change | OpenAI GPT-3 Davinci | $20.00 | $20.00 | Original source ↗ |
| Mar 1, 2023Release | OpenAI GPT-3.5 Turbo | $2.00 | $2.00 | Original source ↗ |
| Jul 2023Quote | Anthropic Claude 2July 2023 price sheet. Plotted at month start; not an exact-day price observation. | $11.02 | $32.68 | Original source ↗ |
| Jul 2023Quote | Anthropic Claude Instant | $1.63 | $5.51 | Original source ↗ |
| Nov 6, 2023Release | OpenAI GPT-3.5 Turbo 1106 | $1.00 | $2.00 | Original source ↗ |
| Nov 6, 2023Quote | OpenAI GPT-4Derived from the dated vendor comparison: Turbo was 3× cheaper on input at $10/M and 2× cheaper on output at $30/M. This is a November observation, not a launch-date quote. | $30.00 | $60.00 | Original source ↗ |
| Nov 6, 2023Release | OpenAI GPT-4 Turbo | $10.00 | $30.00 | Original source ↗ |
| Mar 4, 2024Release | Anthropic Claude 3 Opus | $15.00 | $75.00 | Original source ↗ |
| Mar 4, 2024Release | Anthropic Claude 3 Sonnet | $3.00 | $15.00 | Original source ↗ |
| Mar 13, 2024Release | Anthropic Claude 3 Haiku | $0.25 | $1.25 | Original source ↗ |
| Jun 21, 2024Release | Anthropic Claude 3.5 SonnetDate displayed on the provider announcement. | $3.00 | $15.00 | Original source ↗ |
| Jul 18, 2024Release | OpenAI GPT-4o mini | $0.15 | $0.60 | Original source ↗ |
| Aug 12, 2024Price change | Google Gemini 1.5 FlashPrompts under 128K tokens. | $0.075 | $0.30 | Original source ↗ |
| Aug 30, 2024Release | Cohere Command R+ 08-2024The R+ line is separate from Command R and R7B. | $2.50 | $10.00 | Original source ↗ |
| Aug 30, 2024Release | Cohere Command R 08-2024Command R efficiency line. Versions and parameter sizes differ; this is not a constant-capability index. | $0.15 | $0.60 | Original source ↗ |
| Sep 17, 2024Price change | Mistral Codestral | $0.20 | $0.60 | Original source ↗ |
| Sep 17, 2024Price change | Mistral Mistral Large 2 | $2.00 | $6.00 | Original source ↗ |
| Sep 17, 2024Price change | Mistral Mistral NeMo | $0.15 | $0.15 | Original source ↗ |
| Sep 17, 2024Price change | Mistral Mistral Small 24.09 | $0.20 | $0.60 | Original source ↗ |
| Oct 14, 2024Quote | Google Gemini 1.5 Flash-8BAnnounced 3 Oct; paid-tier billing began 14 Oct. Under 128K context. This historical model is retired. | $0.0375 | $0.15 | Original source ↗ |
| Oct 16, 2024Release | Mistral Ministral 3BOriginal 2024 Ministral 3B. Different generation and capability from Ministral 3 3B (2512). | $0.04 | $0.04 | Original source ↗ |
| Oct 16, 2024Release | Mistral Ministral 8BOriginal 2024 Ministral 8B. Kept within its size family; not a fixed-quality index. | $0.10 | $0.10 | Original source ↗ |
| Dec 12, 2024Price change | xAI Grok-2 1212 | $2.00 | $10.00 | Original source ↗ |
| Jan 9, 2025Quote | Amazon Nova LiteAWS archived in-region standard on-demand rate card, published 2025-01-09. Term effective 2024-12-01. No batch, cross-region, cache, priority or custom-model rates. Date is a published rate-card observation, not a release or measured charge. | $0.06 | $0.24 | Original source ↗ |
| Jan 9, 2025Quote | Amazon Nova MicroAWS archived in-region standard on-demand rate card, published 2025-01-09. Term effective 2024-12-01. No batch, cross-region, cache, priority or custom-model rates. Date is a published rate-card observation, not a release or measured charge. | $0.035 | $0.14 | Original source ↗ |
| Jan 9, 2025Quote | Amazon Nova ProAWS archived in-region standard on-demand rate card, published 2025-01-09. Term effective 2024-12-01. No batch, cross-region, cache, priority or custom-model rates. Date is a published rate-card observation, not a release or measured charge. | $0.80 | $3.20 | Original source ↗ |
| Jan 20, 2025Release | DeepSeek DeepSeek R1Reasoning-model launch. Kept separate from the general-model and Flash families. | $0.55 | $2.19 | Original source ↗ |
| Feb 8, 2025Quote | DeepSeek DeepSeek V3Announced standard post-promotion rates; cache-miss input. V3 launched 26 Dec 2024. | $0.27 | $1.10 | Original source ↗ |
| Apr 4, 2025Quote | Amazon Nova LiteApproximate prices quoted by AWS. | $0.06 | $0.24 | Original source ↗ |
| Apr 4, 2025Quote | Amazon Nova MicroApproximate prices as quoted in the dated AWS article, not a launch-price assertion. | $0.035 | $0.14 | Original source ↗ |
| Apr 14, 2025Release | OpenAI GPT-4.1 | $2.00 | $8.00 | Original source ↗ |
| Apr 14, 2025Release | OpenAI GPT-4.1 mini | $0.40 | $1.60 | Original source ↗ |
| Apr 14, 2025Release | OpenAI GPT-4.1 nano | $0.10 | $0.40 | Original source ↗ |
| May 2, 2025Quote | Amazon Nova LiteAWS archived in-region standard on-demand rate card, published 2025-05-02. Term effective 2025-04-01. No batch, cross-region, cache, priority or custom-model rates. Date is a published rate-card observation, not a release or measured charge. | $0.06 | $0.24 | Original source ↗ |
| May 2, 2025Quote | Amazon Nova MicroAWS archived in-region standard on-demand rate card, published 2025-05-02. Term effective 2025-04-01. No batch, cross-region, cache, priority or custom-model rates. Date is a published rate-card observation, not a release or measured charge. | $0.035 | $0.14 | Original source ↗ |
| May 2, 2025Quote | Amazon Nova ProAWS archived in-region standard on-demand rate card, published 2025-05-02. Term effective 2025-04-01. No batch, cross-region, cache, priority or custom-model rates. Date is a published rate-card observation, not a release or measured charge. | $0.80 | $3.20 | Original source ↗ |
| May 22, 2025Release | Anthropic Claude Opus 4 | $15.00 | $75.00 | Original source ↗ |
| May 22, 2025Release | Anthropic Claude Sonnet 4 | $3.00 | $15.00 | Original source ↗ |
| Jul 10, 2025Quote | Mistral Mistral Small 3.1Mistral explicitly quotes Small 3.1 in its 10 July Devstral announcement. This is the quote date, not an inferred March launch price. | $0.10 | $0.30 | Original source ↗ |
| Jul 14, 2025Quote | Alibaba Qwen-Plus (non-thinking)Standard non-thinking rates quoted in the dated Alibaba Cloud USD notice. Thinking-mode and mainland CNY offers are excluded. This does not establish an April launch price. | $0.40 | $1.20 | Original source ↗ |
| Aug 7, 2025Release | OpenAI GPT-5 | $1.25 | $10.00 | Original source ↗ |
| Aug 7, 2025Release | OpenAI GPT-5 mini | $0.25 | $2.00 | Original source ↗ |
| Aug 7, 2025Release | OpenAI GPT-5 nano | $0.05 | $0.40 | Original source ↗ |
| Sep 5, 2025Price change | DeepSeek DeepSeek V3.1Price increase effective 16:00 UTC; standard cache-miss input, off-peak discounts ended. | $0.56 | $1.68 | Original source ↗ |
| Sep 19, 2025Release | xAI Grok 4 FastUnder 128K input tier; reasoning and non-reasoning variants shared this token rate. | $0.20 | $0.50 | Original source ↗ |
| Sep 29, 2025Price change | DeepSeek DeepSeek V3.2-ExpNew standard cache-miss rates effective 10:00 UTC. Experimental successor; not a fixed-capability comparison. | $0.28 | $0.42 | Original source ↗ |
| Nov 24, 2025Release | Anthropic Claude Opus 4.5 | $5.00 | $25.00 | Original source ↗ |
| Dec 18, 2025Quote | Cohere Command R+ 08-2024Vendor documentation at immutable 18 Dec 2025 commit; the documented rate is unchanged. | $2.50 | $10.00 | Original source ↗ |
| Dec 18, 2025Quote | Cohere Command R 08-2024Rate in the vendor documentation at immutable commit 83d07951c17c21725c349d80c75dc67ef065f4d6, committed 18 Dec 2025. Not backdated from model ID or frontmatter. | $0.15 | $0.60 | Original source ↗ |
| Dec 18, 2025Quote | Cohere Command R7B 12-2024Smaller R-family option in the same immutable 18 Dec 2025 documentation snapshot. Model size/capability differs from R08. | $0.0375 | $0.15 | Original source ↗ |
| Dec 23, 2025Quote | Amazon Nova LiteAWS archived in-region standard on-demand rate card, published 2025-12-23. Term effective 2025-12-01. No batch, cross-region, cache, priority or custom-model rates. Date is a published rate-card observation, not a release or measured charge. | $0.06 | $0.24 | Original source ↗ |
| Dec 23, 2025Quote | Amazon Nova MicroAWS archived in-region standard on-demand rate card, published 2025-12-23. Term effective 2025-12-01. No batch, cross-region, cache, priority or custom-model rates. Date is a published rate-card observation, not a release or measured charge. | $0.035 | $0.14 | Original source ↗ |
| Dec 23, 2025Quote | Amazon Nova ProAWS archived in-region standard on-demand rate card, published 2025-12-23. Term effective 2025-12-01. No batch, cross-region, cache, priority or custom-model rates. Date is a published rate-card observation, not a release or measured charge. | $0.80 | $3.20 | Original source ↗ |
| Apr 24, 2026Release | DeepSeek DeepSeek V4 Flash PreviewOriginal V4 Flash standard cache-miss launch rates. Flash history starts here; V3 flagship history is a separate line. | $0.14 | $0.28 | Original source ↗ |
| Apr 24, 2026Release | DeepSeek DeepSeek V4 Pro PreviewOriginal V4 preview standard cache-miss rates. Subsequent temporary discounts are not silently applied to earlier quotes. | $1.74 | $3.48 | Original source ↗ |
| May 4, 2026Quote | Amazon Nova LiteAWS archived in-region standard on-demand rate card, published 2026-05-04. Term effective 2026-04-01. No batch, cross-region, cache, priority or custom-model rates. Date is a published rate-card observation, not a release or measured charge. | $0.06 | $0.24 | Original source ↗ |
| May 4, 2026Quote | Amazon Nova MicroAWS archived in-region standard on-demand rate card, published 2026-05-04. Term effective 2026-04-01. No batch, cross-region, cache, priority or custom-model rates. Date is a published rate-card observation, not a release or measured charge. | $0.035 | $0.14 | Original source ↗ |
| May 4, 2026Quote | Amazon Nova ProAWS archived in-region standard on-demand rate card, published 2026-05-04. Term effective 2026-04-01. No batch, cross-region, cache, priority or custom-model rates. Date is a published rate-card observation, not a release or measured charge. | $0.80 | $3.20 | Original source ↗ |
| Jul 16, 2026Release | xAI Grok 4.5Standard API rates stated in the dated launch article. Input pricing is unchanged from the December 2024 Grok-2 quote. | $2.00 | $6.00 | Original source ↗ |
| Aug 16, 2026Price change | DeepSeek DeepSeek V4 Flash 0731Peak-hour cache-miss rates effective 16:00 UTC. Off-peak input $0.22 / output $0.66 is excluded from this standard/peak comparison. | $0.44 | $1.32 | Original source ↗ |
| Aug 16, 2026Price change | DeepSeek DeepSeek V4 Pro 0813Peak-hour cache-miss rates effective 16:00 UTC. Off-peak rates were half of these; model and time-tier changes are explicit. | $1.32 | $3.96 | Original source ↗ |
| Sep 10, 2026Price change | DeepSeek DeepSeek V4.1 FlashPeak-hour cache-miss rates effective 04:00 UTC. Off-peak input $0.15 / output $0.60 is a different billing tier. | $0.30 | $1.20 | Original source ↗ |
| Sep 21, 2026Release | xAI Grok 4.7Standard short-context starting rates in the dated launch article; the fast variant is a separate, more expensive offer. | $2.00 | $6.00 | Original source ↗ |
| Nov 9, 2026Quote | Mistral Codestral | $0.30 | $0.90 | Original source ↗ |
| Nov 9, 2026Quote | Cohere Command R 08-2024Current vendor model specification; distinct from the deprecated March 2024 Command R. | $0.15 | $0.60 | Original source ↗ |
| Nov 9, 2026Quote | Cohere Command R7B 12-2024Current vendor model specification. R7B is a smaller, cheaper R-family option; no equal-capability claim. | $0.0375 | $0.15 | Original source ↗ |
| Nov 9, 2026Quote | Cohere Command R 03-2024Legacy March 2024 price retained from Cohere FAQ. Vendor deprecation notice removed this model from current forecast anchors; newer R08 and R7B offers are separate observations. | $0.50 | $1.50 | Original source ↗ |
| Nov 9, 2026Quote | Cohere Command R+ 04-2024Legacy April 2024 price retained from Cohere FAQ; deprecated model is excluded from current forecast anchors. | $3.00 | $15.00 | Original source ↗ |
| Nov 9, 2026Quote | Cohere Command R+ 08-2024 | $2.50 | $10.00 | Original source ↗ |
| Nov 9, 2026Quote | DeepSeek DeepSeek V4.1 Flash 1017Peak-hour, cache-miss rate. The provider also offers 50% lower off-peak rates. | $0.30 | $1.20 | Original source ↗ |
| Nov 9, 2026Quote | DeepSeek DeepSeek V4 Pro 0813Peak-hour, cache-miss rate; off-peak is half. | $1.32 | $3.96 | Original source ↗ |
| Nov 9, 2026Quote | Anthropic Claude Fable 5.1 | $10.00 | $50.00 | Original source ↗ |
| Nov 9, 2026Quote | Google Gemini 2.5 Flash-Lite | $0.10 | $0.40 | Original source ↗ |
| Nov 9, 2026Quote | Google Gemini 2.5 ProAt or below 200K input. Deprecated; provider documents automatic routing to a newer model from 1 Dec 2026. | $1.25 | $10.00 | Original source ↗ |
| Nov 9, 2026Quote | Google Gemini 3 Flash Preview | $0.50 | $3.00 | Original source ↗ |
| Nov 9, 2026Quote | Google Gemini 3.1 Pro PreviewPrompts at or below 200K tokens. | $2.00 | $12.00 | Original source ↗ |
| Nov 9, 2026Quote | Google Gemini 3.6 FlashIntroductory price through 31 Dec 2026; documented to double on 1 Jan 2027. | $0.75 | $3.75 | Original source ↗ |
| Nov 9, 2026Quote | Google Gemini 3.8 FlashIntroductory price through 31 Dec 2026. Published scheduled rates from 1 Jan 2027 are $1.50 input / $7.50 output. | $0.75 | $3.75 | Original source ↗ |
| Nov 9, 2026Quote | OpenAI GPT-4o | $2.50 | $10.00 | Original source ↗ |
| Nov 9, 2026Quote | OpenAI GPT-5.1 | $1.25 | $10.00 | Original source ↗ |
| Nov 9, 2026Quote | OpenAI GPT-5.2 | $1.75 | $14.00 | Original source ↗ |
| Nov 9, 2026Quote | OpenAI GPT-5.4 mini | $0.75 | $4.50 | Original source ↗ |
| Nov 9, 2026Quote | OpenAI GPT-5.4 nano | $0.20 | $1.25 | Original source ↗ |
| Nov 9, 2026Quote | OpenAI GPT-5.4 | $2.50 | $15.00 | Original source ↗ |
| Nov 9, 2026Quote | OpenAI GPT-5.5 | $5.00 | $30.00 | Original source ↗ |
| Nov 9, 2026Quote | OpenAI GPT-5.6 Luna | $0.20 | $1.20 | Original source ↗ |
| Nov 9, 2026Quote | OpenAI GPT-5.6 SolPromotional rate currently documented through 30 Nov 2026. | $4.00 | $20.00 | Original source ↗ |
| Nov 9, 2026Quote | OpenAI GPT-5.6 Terra | $2.00 | $12.00 | Original source ↗ |
| Nov 9, 2026Quote | OpenAI GPT-6 Astra | $10.00 | $50.00 | Original source ↗ |
| Nov 9, 2026Quote | OpenAI GPT-6 Luna | $0.10 | $0.50 | Original source ↗ |
| Nov 9, 2026Quote | OpenAI GPT-6 Sol | $2.00 | $10.00 | Original source ↗ |
| Nov 9, 2026Quote | OpenAI GPT-6.1 Sol | $2.00 | $10.00 | Original source ↗ |
| Nov 9, 2026Quote | OpenAI GPT-6.1 Terra | $1.00 | $5.00 | Original source ↗ |
| Nov 9, 2026Quote | xAI Grok 4.7 | $2.00 | $6.00 | Original source ↗ |
| Nov 9, 2026Quote | Anthropic Claude Haiku 4.5 | $1.00 | $5.00 | Original source ↗ |
| Nov 9, 2026Quote | Anthropic Claude Haiku 5.5Prompts up to 100K tokens; higher rates above that threshold. | $0.10 | $0.50 | Original source ↗ |
| Nov 9, 2026Quote | Amazon Nova LiteAWS us-east-1 in-region standard on-demand text rate. USD per 1K tokens converted to USD per 1M. Batch, cross-region, cache, priority and custom-model prices are excluded. | $0.06 | $0.24 | Original source ↗ |
| Nov 9, 2026Quote | Amazon Nova ProAWS us-east-1 in-region standard on-demand text rate. USD per 1K tokens converted to USD per 1M. Batch, cross-region, cache, priority and custom-model prices are excluded. | $0.80 | $3.20 | Original source ↗ |
| Nov 9, 2026Quote | Amazon Nova MicroAWS us-east-1 in-region standard on-demand text rate. USD per 1K tokens converted to USD per 1M. Batch, cross-region, cache, priority and custom-model prices are excluded. | $0.035 | $0.14 | Original source ↗ |
| Nov 9, 2026Quote | Mistral Ministral 3 14B | $0.20 | $0.20 | Original source ↗ |
| Nov 9, 2026Quote | Mistral Ministral 3 3B | $0.10 | $0.10 | Original source ↗ |
| Nov 9, 2026Quote | Mistral Ministral 3 8B | $0.15 | $0.15 | Original source ↗ |
| Nov 9, 2026Quote | Mistral Mistral Small 4 | $0.15 | $0.60 | Original source ↗ |
| Nov 9, 2026Quote | Mistral Mistral Large 4 | $0.70 | $2.10 | Original source ↗ |
| Nov 9, 2026Quote | OpenAI o3 | $2.00 | $8.00 | Original source ↗ |
| Nov 9, 2026Quote | Anthropic Claude Opus 4.6 | $5.00 | $25.00 | Original source ↗ |
| Nov 9, 2026Quote | Anthropic Claude Opus 4.7 | $5.00 | $25.00 | Original source ↗ |
| Nov 9, 2026Quote | Anthropic Claude Opus 4.8 | $5.00 | $25.00 | Original source ↗ |
| Nov 9, 2026Quote | Anthropic Claude Opus 5 | $5.00 | $25.00 | Original source ↗ |
| Nov 9, 2026Quote | Anthropic Claude Opus 5.5 | $4.00 | $20.00 | Original source ↗ |
| Nov 9, 2026Quote | Alibaba Qwen3.7-PlusSingapore international deployment; at or below 256K input. Other regional and promotional rates differ. | $0.40 | $1.60 | Original source ↗ |
| Nov 9, 2026Quote | Anthropic Claude Sonnet 4.6 | $3.00 | $15.00 | Original source ↗ |
| Nov 9, 2026Quote | Anthropic Claude Sonnet 5.5 | $2.00 | $10.00 | Original source ↗ |
Compute has
a price, too.
GPU hours can get more expensive while each completed job gets cheaper. Capacity, power, utilization and model efficiency all move the bill.
How Runpod adjusts prices with demand ↗What if capacity gets tight?
The rental record starts Nov 2026. Sparse series use persistence; older prices are not invented. GPU-hour and chip-hour curves compare billing rates, not equal capability. Listed prices do not guarantee stock.
Demand is only
half the equation.
Change supply, demand and efficiency separately. These controls describe a possible world; their coefficients are not calibrated predictions.
What does running
the model actually cost?
Hardware is only the start. Include idle time, power and how much of the benchmark your workload can sustain.
Llama 2 7B · Q4_0 · llama.cppCommunity-reported single-stream benchmark, commit 8e672ef. The $599 purchase reference is Apple’s October 2024 US launch price, not a current resale quote.
Benchmark / specification source ↗Historical purchase-price reference ↗- Hardware / rent
- $16.64
- Power
- $2.47
- Other costs
- $0
- Total / 720-hour month
- $19.11
- Effective output capacity
- 11.8 tok/s
- Productive time
- 360 hours
The 24.11 tok/s decode benchmark becomes 16.8 effective output tok/s when its 512-token prompt-processing time is included. The combined value is a calculation, not a separately measured benchmark. Costs exclude unspecified labor, taxes and network charges. Different models are not a quality-adjusted comparison.
Cost formulas, benchmark limits & forecast rules +
Output tokens = achieved aggregate output tokens/s × productive hours × 3600. Owned-machine monthly cost = purchase price / amortization months + active/idle kWh × electricity rate + other costs. Rental cost = hourly quote × 720 × billed chips/GPUs + other costs. Output cost = total monthly cost / output tokens × 1,000,000. A single-stream reference’s effective output rate is output length / (input length / prefill rate + output length / decode rate). Memory fit, context, KV cache, batching, quantization and runtime affect actual results. Benchmark values do not establish your workload’s speed or quality.
Rental forecasts use the same persistence / damped-trend / robust-trend ensemble as the API chart, trained separately per hardware, provider, region and billing unit. Today’s short record produces persistence baselines. News/cost regression requires 18 complete observations over a year. The structural scenario uses scarcity = ((1 + demand growth) / (1 + capacity growth))^(elasticity × years); hourly cost factor = scarcity × (1 − power share) + (1 + power growth)^years × power share; per-token cost divides that by efficiency growth. API pass-through is a separate user assumption. The scenario is not a learned market forecast.
The default electricity reference is 0.196 USD/kWh, 2026-09-01: US city-average residential electricity from BLS, not an industrial data-center tariff or the owner’s actual bill. Active and idle watts are editable assumptions. Server rental electricity is already included in the listed rate; it is not charged twice.
Electricity data source ↗Follow the inputs.
Keep the receipts.
Published prices, hardware costs and the news around them. Every forecast needs a record it can be measured against.
Snapshot: Nov 9, 07:18 PM UTCDownload the public data ↗15 model companies · OpenRouter catalog
historical manual observations remain labeled
failures and conditional requests count
local compute / hosting costs are separate
Compare what’s listed.
These are paid OpenRouter text-output routes, not guaranteed direct-vendor rates or all available models. Free routes are excluded; batch routes are hidden by default because they have different turnaround requirements. Prices are converted from USD/token to USD/1M tokens. Context, availability, tokenizers, model quality, tools, caching, residency and fees differ. Listing is not a successful inference test.
Price the workload.
Compare token bills for the same input/output counts. Check that each model can actually do the job.
Choose a model above and pin it to compare costs here.
A small model.
An existing laptop.
This narrator was generated on an M4 Max MacBook with Qwen3-TTS, an open-weight voice model. A reference recording gives it the voice; MLX runs the model on Apple silicon.
Generate once. Cache it. Serve it.
The model does not run for every visitor.
One measured run on M4 Max / 36 GB, with the model already loaded. This is pre-generated narration, not a live-conversation latency benchmark. Hardware, electricity, engineering and delivery still cost money.
Study the incumbent.
Build for the gap.
Niche arbitrage: find a job people already pay for, understand where the existing product gets in the way, and build a focused alternative. That is the factory’s selection rule.
Start where someone already pays.
Study the incumbent, the pricing, and the job customers are trying to do. Look for a specific workflow with an identifiable buyer.
Build small. Charge for the job. Reinvest in what gets used.