NFT AtomsAPPLIED AI & PRODUCT ENGINEERING
Explore the voice demoDiscuss your use case
Menu
Why own your model

Frontier models arerented by the token.Yours shouldn’t be.

They are repriced without notice and retired within months. Below is the evidence, every number dated and linked, and the case for a small model trained on your job and run on capacity you control.

34models retired by the big three
83 daysshortest lifetime, Gemini 3.1 Flash-Lite prev.
50public sources, all linked below
Anatomy of a token

Every request pays forthe whole encyclopedia.

A frontier model carries the world inside it: every language, every subject, every style. That is what makes it expensive to run, and you pay for all of it on every call.

Your product uses a sliver. Routing a ticket, extracting a field, deciding whether to approve a refund, transcribing a customer. None of that needs trigonometry or sonnets. The rest is weight you rent anyway.

A small model trained on your task keeps the part you use and drops the part you don’t. It answers faster because there is less of it to run, and because it sits next to your data on hardware you rent by the hour, not the token. Send it a million requests or a billion. The invoice does not move.

One frontier request$1.00
  • World knowledge
  • Code in 40 languages
  • Poetry and prose
  • Maths and science
  • History and law
  • Reasoning about anything
  • Safety for every user
  • Your taskkept
Keep the sliver. Drop the rest.
The evidence

The bill is real.So are these numbers.

Nothing here is invented. Each figure is public, dated and linked in the sources at the bottom of the page.

$0B

Enterprise generative-AI spend in 2025, up from $11.5B in 2024 and $1.7B in 2023. Foundation-model API spend alone: $12.5B. Survey of 495 US enterprise decision-makers.

Menlo Ventures, Dec 2025 24
0%

Of organisations had AI cost overruns in the past twelve months. Survey of 500 finance leaders at companies with 1,000+ employees.

DoiT / Sapio, Feb 2026 25
0×

Input price of OpenAI's newest flagship, GPT-6 Astra ($10 per M), against GPT-5 thirteen months earlier ($1.25 per M). Same API, new model, new number.

OpenAI pricing, Sept 2026 1
0%

Rise in Gemini 2.5 Flash output price when it went from preview to general availability: $0.60 to $2.50 per M tokens, June 17, 2025.

Willison, June 2025 11
0 days

Lifetime of GPT-4.5 Preview in the API: launched Feb 27, 2025 at $75 / $150 per M, removed July 14, 2025.

OpenAI deprecations 4
0 quadrillion

Tokens Google processes per month, seven times the year before. Every one of them is on somebody's invoice.

Google I/O, May 2026 26
0%

Of organisations have full real-time visibility into their AI operating costs. 49% scaled back or paused agent deployments when costs outweighed value.

KPMG AI Pulse, Q2 2026 27
0%

More tokens for the same text under Anthropic's Claude 4.7+ tokenizer, at the same list price. A price rise that never touched the price list.

Anthropic pricing 2
0 h

Elevated error rates across OpenAI APIs on June 10, 2025, from 06:36 to 22:00 on the status page. Their outage is your outage.

OpenAI status 20
0%

Scheduled increase for Gemini 3.6 to 3.8 Flash on Jan 1, 2027: $0.75 / $3.75 becomes $1.50 / $7.50 per M. It is already in the changelog.

Google pricing, Sept 2026 3
$0

Per million output tokens for o1-pro, still on OpenAI's price list today. Ten times o1. The ceiling of "it depends on the model".

OpenAI pricing 1
$0M

Average LLM spend a Global 2000 enterprise expects this year, up about 65% on last year and from roughly $4.5M two years ago. Growth in the bill, not in the number of problems solved.

a16z CIO survey, 2026 28
The graveyard

You built on it.They retired it.

Closed models are products with a shelf life. When one is retired, every prompt, eval and integration tuned against it becomes a migration project, on their schedule. Dates are from the vendors’ own deprecation pages.

  1. Claude Opus 4.1365 days
  2. Claude Sonnet 4 / Opus 4389 days
  3. Gemini 2.0 Flash481 days
  4. Gemini 2.0 Flash-Lite461 days
  5. Gemini 3.1 Flash-Lite prev.preview83 days
  6. Claude 3 Haiku768 days
  7. gpt-4-03141108 days
  8. GPT-4 Turbo previewpreview871 days
  9. Gemini 3 Pro Previewpreview111 days
  10. Claude 3.7 Sonnet360 days
  11. Claude 3.5 Haiku485 days
  12. Gemini 2.5 Flash prev. (Sep)preview145 days
  13. GPT-4o (ChatGPT)641 days
  14. Claude 3 Opus672 days
  15. Gemini 2.5 Pro prev. (Jun)preview180 days
  16. Gemini 2.5 Pro prev. (May)preview210 days
  17. Gemini 2.5 Pro prev. (Mar)preview252 days
  18. Gemini 2.5 Flash prev. (May)preview182 days
  19. Claude 3.5 Sonnet (Oct)371 days
  20. Claude 3.5 Sonnet (Jun)495 days
  21. o1-mini410 days
  22. Gemini 1.5 Pro489 days
  23. Gemini 1.5 Flash489 days
  24. o1-previewpreview319 days
  25. Claude 3 Sonnet504 days
  26. Claude 2.1608 days
  27. Claude 2.0741 days
  28. GPT-4.5 Previewpreview137 days
  29. gpt-4-32k815 days
  30. GPT-4 (ChatGPT)778 days
  31. gpt-4-vision-previewpreview396 days
  32. Claude Instant 1.2455 days
  33. Claude 1.x603 days
  34. text-davinci-003402 days

34 retirements shown. Preview and dated-snapshot models are marked; they were still billable endpoints that teams shipped on.

Volatility

The price is not a number.It’s a mood.

Cuts make headlines. Rises arrive in a changelog. Either way, a cost you don’t control is a cost you can’t forecast. Drag or scroll sideways.

  1. 30×

    GPT-4.5 Preview launches at $75 / $150 per M

    Thirty times GPT-4o on input, fifteen on output. Pulled from the API 137 days later.

    Willison 7
  2. $600

    o1-pro: $150 in / $600 out per M

    Ten times o1. Still on the price list in Sept 2026.

    Neowin 8
  3. +100%

    Gemini 2.5 Pro adds a long-context surcharge

    Prompts over 200K tokens: input doubles to $2.50, output rises to $15 per M.

    TechCrunch 9
  4. −80%

    OpenAI cuts o3 from $10 / $40 to $2 / $8

    A cut you did not plan for is still a number you did not control.

    VentureBeat 10
  5. +317%

    Gemini 2.5 Flash goes GA: output $0.60 → $2.50 per M

    Input doubled too, $0.15 → $0.30. The same weights, a new invoice.

    Willison 11
  6. Caps

    Anthropic adds weekly limits on top of 5-hour limits

    Pro and Max subscribers get weekly hour ceilings for Sonnet and Opus.

    TechCrunch 17
  7. −67%

    Claude Opus 4.5 priced at $5 / $25, down from $15 / $75

    Good news, if your margins were not already built on the old number.

    Anthropic 12
  8. ×

    GPT-4o retired from ChatGPT, for the second time

    Retired at GPT-5 launch, restored after user backlash, retired again with GPT-4.1 and o4-mini.

    OpenAI 18
  9. +30%

    Claude 4.7+ tokenizer yields ~30% more tokens per text

    Same list price per token. About 30% more tokens per request. Anthropic says so on its own pricing page.

    Anthropic pricing 2
  10. 5×

    Gemini 3.5 Flash replaces the Flash tier at $1.50 / $9

    The tier that cost $0.30 / $2.50 a year earlier. Five times on input, 3.6 on output.

    XDA 13
  11. 4×

    GPT-5.6 Sol lands at $5 / $30 per M

    GPT-5, eleven months earlier, was $1.25 / $10. Same "flagship" slot, four times the input price.

    PYMNTS 15
  12. −80%

    GPT-5.6 Luna cut to $0.20 / $1.20

    Terra cut 20% the same day. "Priority processing" renamed "Fast mode", still about 2× standard.

    OpenAI pricing 1
  13. −33%

    Sol promo: $4 / $20 "at least through November 21, 2026"

    A discount with an expiry date is a price rise with a calendar.

    OpenAI forum 14
  14. ±0

    Anthropic cancels a planned Sonnet 5 rise to $3 / $15

    The "introductory" $2 / $10 becomes standard. The rise was announced first; the cancellation came later.

    Anthropic pricing 2
  15. $50

    GPT-6 Astra: $10 / $50 per M, Fast mode $20 / $100

    Eight times GPT-5's input price in thirteen months. Long context: $20 / $75.

    OpenAI pricing 1
  16. −17%

    Claude Code weekly limits fall 17% vs the summer promo

    Temporary +50% becomes permanent +25%. Anthropic's own arithmetic.

    Anthropic 19
  17. +100%

    Scheduled: Gemini 3.6–3.8 Flash doubles to $1.50 / $7.50

    Already published. If you are budgeting 2027 on today's Flash price, you are budgeting wrong.

    Google pricing 3
Proof

Small models winon narrow tasks.

“But aren’t the big models smarter?” Yes, at everything. Everything is not your job. On the one task you train for, the small model is measured on exactly the thing it was trained on.

+10 points over GPT-4

LoRA Land: 310 fine-tunes of ten open models of 8B parameters or fewer, across 31 tasks. On average the 4-bit fine-tuned models beat their base model by 34 points and GPT-4 by 10. Twenty-five of them served from a single A100.

Zhao et al., arXiv, Apr 2024 29
5–30× cheaper, and better

Small models (Qwen3-8B, Gemini Flash-Lite and others) fine-tuned on curated GPT-4.1 outputs beat the teacher on named-entity extraction, multi-hop retrieval and τ-bench, at 5 to 30 times lower inference cost.

TensorZero, Jul 2025 30
95.0% vs 75–80%

A fine-tuned Qwen3.5-27B classifying security documents into 7 categories and 51 subcategories, against commercial frontier models under the same prompting. Narrow task, your data, small model wins.

TorchSight, arXiv, May 2026 31
7B beats GPT-4o at tool calling

NVIDIA's Tool-N1 7B and 14B models, trained with reinforcement learning for function calling, outperform GPT-4o on the Berkeley Function Calling Leaderboard and API-Bank. Agents do not need an encyclopedia either.

NVIDIA, arXiv, Apr 2025 34
5.33% WER, 2B params, Apache 2.0

Top of the Open ASR Leaderboard in mid-2026 is a two-billion-parameter open model. Whisper large-v3 sits near 7.4%. NVIDIA's 0.6B Parakeet runs at over 3,000× real time. Speech is already an open-weights game.

Open ASR Leaderboard via MarkTechPost, Jul 2026 32
4 months behind, on average

Epoch AI measures the open-weight frontier as trailing the closed frontier by about four months in 2026. For a task you fine-tune on, that gap disappears: the small model is trained on exactly the thing you measure it on.

Epoch AI, May 2026 33

For balance: in a16z's 2025 CIO survey, some enterprises reported that prompt engineering matched fine-tuning at lower cost. That is often true for low-volume work. Our audit exists to find out which of your workloads it is not true for, before you spend anything.

3.0M

Public models on Hugging Face, Aug 2026. About 2,700 new ones a day.

35
83%

Of all-time Hugging Face downloads are models under one billion parameters. Small is what people actually run.

36
Apache 2.0

Gemma 4, Qwen3.5, Mistral Small 4 and gpt-oss all ship under permissive licences you can build a business on.

38
Speed

Real time is not a feature.It’s the product.

A voice agent that pauses for a second is a bad voice agent. A fraud decision that takes a second is not in the payment path. A rented model answers from a shared data centre after a queue. A small model answers from a GPU next to your data.

LLM decision · time to first token

GPT-6 Astra (low reasoning), P502,220 ms
Llama 3.1 8B, one H100, 1,000-token prompt19 ms

Deloitte and Google measured 0.1 s faster mobile pages lifting retail conversion 8.4%. A decision that takes seconds cannot sit in a checkout, a payment path or a live chat.

Speech-to-text · first partial after you stop talking

ElevenLabs Scribe v2 Realtime, AA benchmark130 ms
NVIDIA Nemotron 3 ASR (open weights), same benchmark40 ms

People answer each other within about 200 ms. A voice agent that waits longer is already late, and a rented API adds a 20–200 ms round trip on top of the model, with your audio leaving your network.

Text-to-speech · time to first audio

ElevenLabs Flash v2.5, P50 incl. network288 ms
Qwen3-TTS 0.6B (Apache 2.0), single stream97 ms

ITU-T G.114: under 150 ms one-way feels transparent, over 400 ms is unacceptable for conversation. First audio is only one leg of that budget.

Rented figures are independent P50 measurements (Artificial Analysis, Coval) that include the vendor's network path. Owned figures are published single-GPU measurements of open-weight models, not our own claims; our targets are on the Pulse card above. Even the fastest frontier model Artificial Analysis measured, Claude Sonnet 5 without reasoning, takes about 1,010 ms to its first token. Rented speech APIs are quick at the model; Deepgram's own docs put the network round trip at 20 to 200 ms on top. Bars use a logarithmic scale.

Estimator

Run yourown numbers.

Pick the frontier model you use today. Its list price comes from the strip at the top of the page. Everything else is yours to edit. The build cost and the flat monthly figure are placeholders to think with, not a quote.

Rented, per month–
Owned, per month–
Build pays back in–
Month 1Month 36–
Rented, cumulativeOwned incl. build

Rented is list-price arithmetic. Real bills also carry retries, rate-limit backoff and migration work when a model is retired. Those are not in the curve, and neither is a price change in either direction.

Sources

Every number, and where it came from.

Prices and dates as published by the vendors and the cited authors. Verified Sept 9, 2026.

  1. 1OpenAI API pricing (official)
  2. 2Anthropic Claude pricing (official)
  3. 3Google Gemini API pricing (official)
  4. 4OpenAI model deprecations (official)
  5. 5Anthropic model deprecations (official)
  6. 6Google Gemini API deprecations (official)
  7. 7Simon Willison, "Introducing GPT-4.5"
  8. 8Neowin, o1-pro priced at $600 per million output tokens
  9. 9TechCrunch, "Gemini 2.5 Pro is Google's most expensive AI model yet"
  10. 10VentureBeat, OpenAI announces 80% price drop for o3
  11. 11Simon Willison, Gemini 2.5 GA pricing
  12. 12Anthropic, "Introducing Claude Opus 4.5"
  13. 13XDA, Gemini 3.5 Flash costs 3× the model it replaced
  14. 14OpenAI community, 20% price reduction for GPT-5.6 Sol
  15. 15PYMNTS, OpenAI cuts prices on select models
  16. 16Yahoo Finance, GPT-6 Astra pricing
  17. 17TechCrunch, Anthropic unveils new rate limits for Claude Code power users
  18. 18OpenAI, "Retiring GPT-4o and older models"
  19. 19Anthropic (@ClaudeDevs), weekly limit change effective Sept 14, 2026
  20. 20OpenAI status, elevated error rates across APIs and ChatGPT
  21. 21The Register, ChatGPT, Claude and Grok outages on the same day
  22. 22Cybersecurity News, Claude outages logged through August 2026
  23. 23OpenAI API pricing, audio (Whisper $0.006/min; TTS $15 per 1M chars)
  24. 24Menlo Ventures, "2025: The State of Generative AI in the Enterprise"
  25. 25DoiT, AI spending survey of 500 finance leaders (Sapio Research)
  26. 26Google, Sundar Pichai at I/O 2026: over 3.2 quadrillion tokens per month
  27. 27KPMG Global AI Pulse, Q2 2026
  28. 28a16z, third annual enterprise CIO survey
  29. 29Zhao et al., "LoRA Land: 310 Fine-tuned LLMs that Rival GPT-4"
  30. 30TensorZero, distillation with programmatic data curation
  31. 31TorchSight, fine-tuned Qwen3.5-27B for security document classification
  32. 32MarkTechPost, best open speech recognition models in 2026 (Open ASR Leaderboard)
  33. 33Epoch AI, open vs closed ECI gap
  34. 34NVIDIA, "Nemotron-Research-Tool-N1" (Tool-N1)
  35. 35Hugging Face, three million models and counting
  36. 36Hugging Face, State of Open Models, Summer 2026
  37. 37a16z, "How 100 enterprise CIOs are building and buying gen AI in 2025"
  38. 38Google, Gemma 4 model card (Apache 2.0)
  39. 39Northflank, open-source speech-to-text benchmarks 2026 (Whisper large-v3 WER)
  40. 40Mavvrik / Benchmarkit, 2025 State of AI Cost Management
  41. 41Artificial Analysis, GPT-6 Astra (low) vs GPT-5.6 Terra: median time to first token
  42. 42NVIDIA NIM benchmarking, Llama 3.1 8B Instruct on one H100 (FP8, TTFT by concurrency)
  43. 43Artificial Analysis, streaming speech-to-text benchmark (latency to first partial, AA-WER)
  44. 44Coval TTS latency benchmark 2026 (time to first audio, P50), via Gradium
  45. 45Qwen3-TTS technical report, first-packet latency
  46. 46Levinson & Torreira, timing in turn-taking (Frontiers in Psychology)
  47. 47ITU-T Recommendation G.114, one-way transmission time
  48. 48Deloitte / Google, "Milliseconds Make Millions"
  49. 49Artificial Analysis, Claude Opus 5 vs Claude Sonnet 5 (non-reasoning): time to first token
  50. 50Deepgram docs, measuring streaming latency (network 20–200 ms)
Get in touch

Tell us aboutthe work.

A workflow that takes too long, a model that keeps missing, a product that needs its next version. Write to us and we reply within a working day.

Discuss your use case hello@nftatoms.com