Short answer. Published list prices fell by roughly 80% a year at the cheap end between 2023 and 2025, then stopped falling at the top. The cheapest OpenAI text model went from $0.60 to $0.50 per million output tokens between November 2024 and November 2026; the top general-purpose line went from $10 to $40. Both are true, and which one matters depends on which slot you buy.
Dataset baseline: 20 November 2026. Every row was read from a provider pricing page, a deprecation notice or a dated launch announcement on that date. Rows we could not confirm are marked “reported” and excluded from the charts. We do not reconstruct a monthly series we cannot source.
Normalised prices, 2023 to 2026
Normalisation is the product. A price is comparable only if the tier, context band and token direction match. We use the standard synchronous tier, short context, text only, and we keep input and output separate because the ratio between them moved from 2:1 to 8:1 and back to 5:1.
| Date | Provider | Model | Input | Output | Basis |
|---|---|---|---|---|---|
| 2023-03-14 | OpenAI | GPT-4 (8K) | $30.00 | $60.00 | Launch |
| 2023-06-13 | OpenAI | GPT-3.5 Turbo (4K) | $1.50 | $2.00 | Price cut |
| 2023-11-06 | OpenAI | GPT-4 Turbo | $10.00 | $30.00 | Launch |
| 2024-01-25 | OpenAI | GPT-3.5 Turbo | $0.50 | $1.50 | Price cut |
| 2024-03-04 | Anthropic | Claude 3 Haiku | $0.25 | $1.25 | Launch |
| 2024-03-04 | Anthropic | Claude 3 Opus | $15.00 | $75.00 | Launch |
| 2024-05-13 | OpenAI | GPT-4o | $5.00 | $15.00 | Launch |
| 2024-07-18 | OpenAI | GPT-4o mini | $0.15 | $0.60 | Launch |
| 2024-08-06 | OpenAI | GPT-4o (2024-08-06) | $2.50 | $10.00 | Price cut |
| 2024-12-17 | Gemini 2.0 Flash | $0.10 | $0.40 | Launch | |
| 2024-12-26 | DeepSeek | DeepSeek V3 | $0.27 | $1.10 | Launch |
| 2025-04-14 | OpenAI | GPT-4.1 | $2.00 | $8.00 | Launch |
| 2025-08-07 | OpenAI | GPT-5 | $1.25 | $10.00 | Launch |
| 2026-03-17 | OpenAI | GPT-5.4 | $2.50 | $15.00 | Current page |
| 2026-04-23 | OpenAI | GPT-5.5 | $5.00 | $30.00 | Reported |
| 2026-09-03 | OpenAI | GPT-6 Astra | $10.00 | $50.00 | Reported |
| 2026-11-20 | OpenAI | GPT-6.2 Astra | $8.00 | $40.00 | Current page |
| 2026-11-20 | OpenAI | GPT-6 Luna | $0.10 | $0.50 | Current page |
USD per 1M tokens. “Reported” means a dated secondary source. GPT-4o and GPT-4.1 shut down on 14 January 2027 and GPT-5 on 11 December 2026, so their rows stay frozen at the last published price. The “Basis” column is the change log.
The 24-month chart
Both charts use the same dates and tier. The first tracks the top general-purpose line; the second tracks the cheapest standard text model. The gap is the story: capability at the top got more expensive in 2026 while the floor stayed flat.
Top general-purpose model (output $/1M)
Cheapest standard text model (output $/1M)
The 2026 flagship rows are 4× the November 2024 output price; the budget line moved from $0.60 to $0.50. If your workload fits the budget slot, “AI gets cheaper every year” still holds. If it needs the top slot, it does not.
Batch and cached rates
Batch and cache pricing changes the effective rate more than most list-price cuts. On the OpenAI pages checked today, batch is 50% of standard, cached input is 10% of input on the GPT-6 family, and a cache write costs 1.25× input for 30 minutes or 2.5× for 12 hours. Anthropic publishes a lower cache read — 2.5% of input on Fable 5.2 — and a write at 1.25×.
| Model | Standard in / out | Cached read | Cache write | Batch in / out | Effective input at 90% hit |
|---|---|---|---|---|---|
| OpenAI GPT-6.2 Astra | $8.00 / $40.00 | $0.40 | $10.00 (30m) | $4.00 / $20.00 | $1.16 |
| OpenAI GPT-6 Luna | $0.10 / $0.50 | $0.01 | $0.125 (30m) | $0.05 / $0.25 | $0.02 |
| Anthropic Fable 5.2 | $8.00 / $40.00 | $0.20 | $10.00 | 50% off | $0.98 |
| Anthropic Haiku 5.5 ≤100K | $0.10 / $0.50 | $0.01 | $0.125 | 50% off | $0.02 |
Effective input = 0.9 × cached read + 0.1 × standard input. It is arithmetic on published rates, not a measured workload. Cache writes are charged separately and excluded, which flatters the cache case.
CSV and API
The CSV below is the normalised extract, small enough to paste into a notebook. The live model list comes from the OpenAI-compatible /v1/models endpoint, so a price table and a catalogue can be joined without scraping.
provider,model_id,observed,input_usd_mtok,output_usd_mtok,cached_input,batch_input,confidence
OpenAI,gpt-4o-2024-08-06,2024-08-06,2.50,10.00,1.25,1.25,primary
OpenAI,gpt-4.1-2025-04-14,2025-04-14,2.00,8.00,0.50,1.00,primary
OpenAI,gpt-5,2025-08-07,1.25,10.00,0.125,0.625,primary
OpenAI,gpt-6-luna,2026-11-20,0.10,0.50,0.01,0.05,primary
Anthropic,claude-sonnet-5-5,2026-11-20,2.00,10.00,0.10,1.00,primary
Anthropic,claude-fable-5-2,2026-11-20,8.00,40.00,0.20,4.00,primary
curl -s https://api.plugsky.com/v1/models \
-H "Authorization: Bearer $PLUGSKY_API_KEY" \
| jq -r '.data[] | [.id, .engine_context_window] | @csv'
Keep a dated snapshot: the endpoint returns today’s catalogue, not historical prices, and that is the only way to answer “what did we pay in March?” after a model is retired.
Methodology and limits
- Unit. USD per 1M tokens; per-1K quotes are multiplied by 1,000. Tokenisers differ, so a token is not the same amount of text across providers.
- Tier. Standard synchronous only. Batch, flex, priority, provisioned throughput, fine-tuning and long-context surcharges are separate rows or excluded.
- Confidence.
primaryis the provider page;reportedis a dated third-party source. Reported rows are excluded from the charts. - Coverage. Six providers and 18 model lines: OpenAI, Anthropic, Google, DeepSeek, Mistral and Plugsky. Resellers and aggregators are excluded because their price is not a list price.
Two limits: private discounts are invisible, and model generations are not strictly comparable, so this is a price index, not a quality-adjusted one.
Sources
Links checked 20 November 2026. We cite the primary page for each figure. Where a provider publishes a price behind a login, we say so and do not guess.
- OpenAI API pricing — standard, batch, cached and long-context rates for the GPT-6, GPT-5 and GPT-4 families
- OpenAI deprecations — shutdown dates plus historical list prices for retired models
- Anthropic pricing — Fable, Opus, Sonnet and Haiku rates, cache reads and writes
- Anthropic model deprecations — retirement dates and replacement models
FAQ
How much did LLM API prices fall between 2023 and 2026?
At the cheap end, the floor fell by roughly two orders of magnitude: GPT-4 launched at $30 per 1M input tokens in March 2023, and GPT-6 Luna lists at $0.10 in November 2026. At the top end the decline reversed: the flagship output price went from $10 in November 2024 to $40 in November 2026.
Is this dataset free to use?
Yes. The table and CSV extract are published under CC BY 4.0 with attribution to Plugsky. Provider names and prices remain the property of the providers, and the underlying terms of service still apply.
Why do you show input and output separately?
Because the ratio between them changed from 2:1 on GPT-4 to 8:1 on GPT-5 and back to 5:1 on GPT-6 Astra. A blended price hides that shift and will misprice any workload that reads long documents and writes short answers.
Does the dataset include batch and cached prices?
Yes, where the provider publishes them. Batch is typically 50% of standard, and cached reads range from 2.5% to 10% of input. Cache writes are charged above input, so a cache only pays off when the same prefix is reused.
Why are some rows marked “reported”?
Because the provider no longer lists the launch price on a page we can read, and a dated third-party report is the best available source. Those rows are excluded from the charts and flagged in the confidence column rather than silently blended with primary data.
How often is the table updated?
Monthly, and within one business day of a price change we can verify. The baseline for this edition is 20 November 2026; the change log records every revision so a row can be traced back to its source.
Use the data, then test it on your own workload
Plugsky publishes its plan caps and model limits. Start free with no card, or compare the flat plans against per-token billing.
Start Free → See plans