For three years API prices only fell: GPT-4 cost $30 per million input tokens in 2023, and OpenAI's flagship today, gpt-6.2-astra, costs $8. That has stopped at the top end. Google's gemini-4-carbon and gemini-4-argon double on December 1, 2026, gemini-3.8-flash doubles on January 1, 2027, and DeepSeek's top tier already costs about twice what deepseek-reasoner cost in January 2025.
We read the official pricing pages of six providers on November 20, 2026, listed 18 models (plus gemini-4-carbon's announced December 1 price and DeepSeek's off-peak prices), compared them with archived 2023–2025 prices and priced three real workloads. Every figure is quoted from a linked source or computed with a formula shown below.
Who this is for: developers, product and finance teams who budget for AI used through an API, and journalists who need a sourced number. Not covered: consumer subscriptions (see ChatGPT Plus vs Pro vs API), image, audio and video models, fine-tuning and negotiated contracts.
Data as of: November 20, 2026. Next update: monthly — prices are re-checked on the 1st of each month and logged at the end of this report.
Key findings
- Small models keep getting cheaper: OpenAI's and Anthropic's small models cost $0.20 blended, 60–90% below their 2023–2024 predecessors.
- At the top end prices now move both ways: the most expensive flagship costs 15 times the cheapest, and Google has scheduled increases for gemini-4-carbon, gemini-4-argon and gemini-3.8-flash.
- Public discounts beat negotiation: caching cut our coding session by 61–79%, and Batch halves most offline jobs.
OpenAI
- Flagship
- $16.00
- Mid
- $4.00
- Small
- $0.20
Anthropic
- Flagship
- $8.00
- Mid
- $4.00
- Small
- $0.20
- Flagship
- $4.00
- Mid
- $1.50
- Small
- $0.85
xAI
- Flagship
- $3.00
- Mid
- $1.56
- Small
- —
Mistral
- Flagship
- $1.05
- Mid
- —
- Small
- $0.26
DeepSeek
- Flagship
- $1.98
- Mid
- —
- Small
- $0.53
| Provider | Flagship | Mid | Small |
|---|---|---|---|
| OpenAI | $16.00 | $4.00 | $0.20 |
| Anthropic | $8.00 | $4.00 | $0.20 |
| $4.00 | $1.50 | $0.85 | |
| xAI | $3.00 | $1.56 | — |
| Mistral | $1.05 | — | $0.26 |
| DeepSeek | $1.98 | — | $0.53 |
Source: OpenAI, Anthropic, Google, xAI, Mistral and DeepSeek pricing pages, read 2026-11-20; CalculatorAI analysis; calculatorai.app
Share / download
Where a provider sells two models in a tier, the table shows the cheaper one (Claude Fable 5.2 costs $16.00; gemini-4-hydrogen $3.00).
Headline findingsThree numbers from the 2027 report
Source: OpenAI, Anthropic, Google, xAI, Mistral and DeepSeek pricing pages (2026-11-20) and archived pricing pages (2023–2025); CalculatorAI analysis; calculatorai.app
Share / download
Budgets that assume each model generation will be cheaper than the last need a second look: on December 1 Google's flagship will cost more than the Gemini 1.5 Pro it replaced, $8.00 against $5.25 blended.

MethodologyHow the numbers were built
(3 × input price + 1 × output price) ÷ 4requests × (input tokens × input price + output tokens × output price) ÷ 1,000,000(blended today ÷ blended then) − 1(blended today ÷ blended then) ^ (1 ÷ years elapsed) − 1- Prices are Standard (real-time) list prices in US dollars from each provider's own pricing page, for prompts below the long-context threshold (272,000 tokens at OpenAI, 200,000 at Google and xAI, 100,000 for Claude Haiku 5.5).
- The 3:1 ratio assumes three input tokens per output token, typical of chat, retrieval and agents; at 1:1, output-heavy models look more expensive.
- Tiers (flagship, mid, small) are the providers' own labels, not a quality ranking. History pairs each model with the earlier model in the same product line on an archived official page.
- Model names follow each pricing page: model IDs for OpenAI, Google, xAI and DeepSeek, product names for Anthropic and Mistral. "(peak)" marks DeepSeek's peak-hour price; off-peak prices are half. Prices are shown to the cent; percentages are computed from unrounded prices (Gemini 1.5 Flash was $0.525 blended, shown as $0.53). Changes of 10% or more are rounded to whole numbers, smaller ones to one decimal.
- Reproducibility: one checked script computes every number; re-run it with your own token counts in the AI Token Cost Calculator.
Price tableWhat 18 models cost on November 20, 2026
21 rows for 18 models: gemini-4-carbon also appears at its announced December 1 price, and DeepSeek's two models at their off-peak prices (outside 01:00–04:00 and 06:00–10:00 UTC on weekdays). List prices are not quality-adjusted; to compare quality, run 30–50 of your own tasks on each shortlisted model, score the answers blind and divide cost by passing answers.
| Model | Input | Blended 3:1 |
|---|---|---|
| OpenAI | ||
| gpt-6.2-astra | $8.00 | $16.00 |
| gpt-6.2-sol | $2.00 | $4.00 |
| gpt-6-luna | $0.10 | $0.20 |
| Anthropic | ||
| Claude Fable 5.2 | $8.00 | $16.00 |
| Claude Opus 5.5 | $4.00 | $8.00 |
| Claude Sonnet 5.5 | $2.00 | $4.00 |
| Claude Haiku 5.5 | $0.10 | $0.20 |
| gemini-4-carbon | $2.00 | $4.00 |
| gemini-4-carbon from Dec 1 | $4.00 | $8.00 |
| gemini-4-argon (doubles Dec 1) | $2.00 | $4.00 |
| gemini-3.8-flash (doubles Jan 1) | $0.75 | $1.50 |
| gemini-4-hydrogen | $1.50 | $3.00 |
| gemini-3.5-flash-lite | $0.30 | $0.85 |
| xAI | ||
| grok-4.7 | $2.00 | $3.00 |
| grok-4.3 | $1.25 | $1.56 |
| Mistral | ||
| Mistral Large 4 | $0.70 | $1.05 |
| Mistral Small 4 | $0.15 | $0.26 |
| DeepSeek | ||
| deepseek-v4-pro (peak) | $1.32 | $1.98 |
| deepseek-v4-pro (off-peak) | $0.66 | $0.99 |
| deepseek-flash (peak) | $0.30 | $0.53 |
| deepseek-flash (off-peak) | $0.15 | $0.26 |
All 7 columns
OpenAI
gpt-6.2-astra
- Tier
- Flagship
- Input
- $8.00
- Cached input
- $0.40
- Output
- $40.00
- Blended 3:1
- $16.00
- Batch discount
- 50%
gpt-6.2-sol
- Tier
- Mid
- Input
- $2.00
- Cached input
- $0.10
- Output
- $10.00
- Blended 3:1
- $4.00
- Batch discount
- 50%
gpt-6-luna
- Tier
- Small
- Input
- $0.10
- Cached input
- $0.01
- Output
- $0.50
- Blended 3:1
- $0.20
- Batch discount
- 50%
Anthropic
Claude Fable 5.2
- Tier
- Flagship
- Input
- $8.00
- Cached input
- $0.20
- Output
- $40.00
- Blended 3:1
- $16.00
- Batch discount
- 50%
Claude Opus 5.5
- Tier
- Flagship
- Input
- $4.00
- Cached input
- $0.20
- Output
- $20.00
- Blended 3:1
- $8.00
- Batch discount
- 50%
Claude Sonnet 5.5
- Tier
- Mid
- Input
- $2.00
- Cached input
- $0.10
- Output
- $10.00
- Blended 3:1
- $4.00
- Batch discount
- 50%
Claude Haiku 5.5
- Tier
- Small
- Input
- $0.10
- Cached input
- $0.01
- Output
- $0.50
- Blended 3:1
- $0.20
- Batch discount
- 50%
gemini-4-carbon
- Tier
- Flagship
- Input
- $2.00
- Cached input
- $0.10
- Output
- $10.00
- Blended 3:1
- $4.00
- Batch discount
- 50%
gemini-4-carbon from Dec 1
- Tier
- Flagship
- Input
- $4.00
- Cached input
- $0.20
- Output
- $20.00
- Blended 3:1
- $8.00
- Batch discount
- 50%
gemini-4-argon (doubles Dec 1)
- Tier
- Flagship
- Input
- $2.00
- Cached input
- $0.10
- Output
- $10.00
- Blended 3:1
- $4.00
- Batch discount
- 50%
gemini-3.8-flash (doubles Jan 1)
- Tier
- Mid
- Input
- $0.75
- Cached input
- $0.075
- Output
- $3.75
- Blended 3:1
- $1.50
- Batch discount
- 50%
gemini-4-hydrogen
- Tier
- Mid
- Input
- $1.50
- Cached input
- $0.075
- Output
- $7.50
- Blended 3:1
- $3.00
- Batch discount
- 50%
gemini-3.5-flash-lite
- Tier
- Small
- Input
- $0.30
- Cached input
- $0.03
- Output
- $2.50
- Blended 3:1
- $0.85
- Batch discount
- 50%
xAI
grok-4.7
- Tier
- Flagship
- Input
- $2.00
- Cached input
- $0.50
- Output
- $6.00
- Blended 3:1
- $3.00
- Batch discount
- 20%
grok-4.3
- Tier
- Mid
- Input
- $1.25
- Cached input
- $0.20
- Output
- $2.50
- Blended 3:1
- $1.56
- Batch discount
- 20%
Mistral
Mistral Large 4
- Tier
- Flagship
- Input
- $0.70
- Cached input
- $0.07
- Output
- $2.10
- Blended 3:1
- $1.05
- Batch discount
- 50%
Mistral Small 4
- Tier
- Small
- Input
- $0.15
- Cached input
- Not listed
- Output
- $0.60
- Blended 3:1
- $0.26
- Batch discount
- 50%
DeepSeek
deepseek-v4-pro (peak)
- Tier
- Flagship
- Input
- $1.32
- Cached input
- $0.044
- Output
- $3.96
- Blended 3:1
- $1.98
- Batch discount
- None listed
deepseek-v4-pro (off-peak)
- Tier
- Flagship
- Input
- $0.66
- Cached input
- $0.022
- Output
- $1.98
- Blended 3:1
- $0.99
- Batch discount
- None listed
deepseek-flash (peak)
- Tier
- Small
- Input
- $0.30
- Cached input
- $0.006
- Output
- $1.20
- Blended 3:1
- $0.53
- Batch discount
- None listed
deepseek-flash (off-peak)
- Tier
- Small
- Input
- $0.15
- Cached input
- $0.003
- Output
- $0.60
- Blended 3:1
- $0.26
- Batch discount
- None listed
| Model | Tier | Input | Cached input | Output | Blended 3:1 | Batch discount |
|---|---|---|---|---|---|---|
| OpenAI | ||||||
| gpt-6.2-astra | Flagship | $8.00 | $0.40 | $40.00 | $16.00 | 50% |
| gpt-6.2-sol | Mid | $2.00 | $0.10 | $10.00 | $4.00 | 50% |
| gpt-6-luna | Small | $0.10 | $0.01 | $0.50 | $0.20 | 50% |
| Anthropic | ||||||
| Claude Fable 5.2 | Flagship | $8.00 | $0.20 | $40.00 | $16.00 | 50% |
| Claude Opus 5.5 | Flagship | $4.00 | $0.20 | $20.00 | $8.00 | 50% |
| Claude Sonnet 5.5 | Mid | $2.00 | $0.10 | $10.00 | $4.00 | 50% |
| Claude Haiku 5.5 | Small | $0.10 | $0.01 | $0.50 | $0.20 | 50% |
| gemini-4-carbon | Flagship | $2.00 | $0.10 | $10.00 | $4.00 | 50% |
| gemini-4-carbon from Dec 1 | Flagship | $4.00 | $0.20 | $20.00 | $8.00 | 50% |
| gemini-4-argon (doubles Dec 1) | Flagship | $2.00 | $0.10 | $10.00 | $4.00 | 50% |
| gemini-3.8-flash (doubles Jan 1) | Mid | $0.75 | $0.075 | $3.75 | $1.50 | 50% |
| gemini-4-hydrogen | Mid | $1.50 | $0.075 | $7.50 | $3.00 | 50% |
| gemini-3.5-flash-lite | Small | $0.30 | $0.03 | $2.50 | $0.85 | 50% |
| xAI | ||||||
| grok-4.7 | Flagship | $2.00 | $0.50 | $6.00 | $3.00 | 20% |
| grok-4.3 | Mid | $1.25 | $0.20 | $2.50 | $1.56 | 20% |
| Mistral | ||||||
| Mistral Large 4 | Flagship | $0.70 | $0.07 | $2.10 | $1.05 | 50% |
| Mistral Small 4 | Small | $0.15 | Not listed | $0.60 | $0.26 | 50% |
| DeepSeek | ||||||
| deepseek-v4-pro (peak) | Flagship | $1.32 | $0.044 | $3.96 | $1.98 | None listed |
| deepseek-v4-pro (off-peak) | Flagship | $0.66 | $0.022 | $1.98 | $0.99 | None listed |
| deepseek-flash (peak) | Small | $0.30 | $0.006 | $1.20 | $0.53 | None listed |
| deepseek-flash (off-peak) | Small | $0.15 | $0.003 | $0.60 | $0.26 | None listed |
Source: OpenAI, Anthropic, Google, xAI, Mistral and DeepSeek pricing pages, read 2026-11-20; CalculatorAI analysis (blended price); calculatorai.app
Dataset: CSV · CC BY 4.0
Share / download
What changes real bills beyond the list price:
- Caching: a cache hit costs 2.5% of the input price on Claude Fable 5.2 and 5% on gpt-6.2-astra, Claude Opus 5.5 and Google's Gemini 4 models; xAI charges 25% on grok-4.7 and Mistral Small 4 lists no cached rate. See prompt caching explained.
- Scheduled changes: gemini-4-carbon and gemini-4-argon go from $2.00 / $10.00 to $4.00 / $20.00 on December 1 ($4.00 → $8.00 blended); gemini-3.8-flash from $0.75 / $3.75 to $1.50 / $7.50 on January 1 ($1.50 → $3.00). Batch and cached prices double on the same dates.
- Speed: OpenAI's Fast mode, xAI's Priority and Anthropic's Fast mode cost twice Standard (Claude Opus 5.5 $8 / $40, Claude Fable 5.2 $16 / $80; no Batch). OpenAI's Ultrafast on gpt-6.2-astra costs three times Standard until February 28, 2027 and six times from March 1. No price in this report includes these modes.
Price historyThree years of launch prices vs today
Each row runs from the price on an archived official page to today's model in the same line; further left is cheaper.
Per year, the flagship falls work out to about −21% (OpenAI), −39% (Anthropic), −59% (Mistral) and −36% (xAI); DeepSeek's off-peak price is only 3.1% above deepseek-reasoner's.
- Anthropic SonnetClaude 3 Sonnet, Mar 2024$6.00 → $4.00−33%
- Mistral SmallMistral Small, Feb 2024$3.00 → $0.26−91%
- OpenAI smallgpt-3.5-turbo, Apr 2023$2.00 → $0.20−90%
- Google Flash, todayGemini 1.5 Flash, Jun 2024$0.53 → $1.50+186%
- Jan 1: Google FlashScheduled · vs Gemini 1.5 Flash$0.53 → $3.00+471%
- Anthropic HaikuClaude 3 Haiku, Mar 2024$0.50 → $0.20−60%
- DeepSeek smalldeepseek-chat, Jan 2025 · peak$0.48 → $0.53+9.9%
Values: $ per 1M tokens, blended
Google's Flash line is the outlier: nearly 3× its 2024 price today and 5.7× from January 1.
Show these figures as a table
| Then ($ per 1M tokens, blended) | Now ($ per 1M tokens, blended) | |
|---|---|---|
| Anthropic Sonnet | $6.00 | $4.00 |
| Mistral Small | $3.00 | $0.26 |
| OpenAI small | $2.00 | $0.20 |
| Google Flash, today | $0.53 | $1.50 |
| Jan 1: Google Flash | $0.53 | $3.00 |
| Anthropic Haiku | $0.50 | $0.20 |
| DeepSeek small | $0.48 | $0.53 |
Source: Provider pricing pages: archived (then) and read 2026-11-20 (now); CalculatorAI analysis; calculatorai.app
Share / download
Mid-tier prices fell least: Claude Sonnet went from $3 / $15 to $2 / $10 (−33%) while the Opus line fell 73% and Haiku 60%, and mid-tier models carry most production traffic. The Flash line now costs nearly three times its 2024 price, and nearly six times from January 1, for a much more capable model. DeepSeek's peak-hour prices are higher on both tiers than its January 2025 list; off-peak deepseek-flash is −45%.
WorkloadsWhat three real jobs cost on each model
Three workloads, priced the same way on every model
Support chatbot day: 1,000 messages, each sending 2,000 input tokens (system prompt, policy, history and the question) and receiving 250 output tokens. Document summaries: 1,000 documents of 6,000 tokens each, with a 500-token summary. Coding-agent session: 40 turns, each re-sending 60,000 tokens of files, tools and history and writing 2,000 tokens. No caching unless stated.
- gpt-6.2-astraOpenAI flagship$26.00
- Claude Fable 5.2Anthropic flagship$26.00
- Claude Opus 5.5Anthropic flagship$13.00
- gpt-6.2-solOpenAI mid$6.50
- Claude Sonnet 5.5Anthropic mid$6.50
- gemini-4-carbonGoogle flagship$6.50
- gemini-4-argonGoogle flagship$6.50
- grok-4.7xAI flagship$5.50
- gemini-4-hydrogenGoogle mid$4.88
- deepseek-v4-pro (peak)DeepSeek flagship$3.63
Show all 18Show fewer
- grok-4.3xAI mid$3.13
- gemini-3.8-flashGoogle mid$2.44
- Mistral Large 4Mistral flagship$1.93
- gemini-3.5-flash-liteGoogle small$1.23
- deepseek-flash (peak)DeepSeek small$0.90
- Mistral Small 4Mistral small$0.45
- gpt-6-lunaOpenAI small$0.33cheapest
- Claude Haiku 5.5Anthropic small$0.33cheapest
Values: $ per day
A small model handles the whole day for 33 cents.
Show these figures as a table
| Value ($ per day) | |
|---|---|
| gpt-6.2-astra — OpenAI flagship | $26.00 |
| Claude Fable 5.2 — Anthropic flagship | $26.00 |
| Claude Opus 5.5 — Anthropic flagship | $13.00 |
| gpt-6.2-sol — OpenAI mid | $6.50 |
| Claude Sonnet 5.5 — Anthropic mid | $6.50 |
| gemini-4-carbon — Google flagship | $6.50 |
| gemini-4-argon — Google flagship | $6.50 |
| grok-4.7 — xAI flagship | $5.50 |
| gemini-4-hydrogen — Google mid | $4.88 |
| deepseek-v4-pro (peak) — DeepSeek flagship | $3.63 |
| grok-4.3 — xAI mid | $3.13 |
| gemini-3.8-flash — Google mid | $2.44 |
| Mistral Large 4 — Mistral flagship | $1.93 |
| gemini-3.5-flash-lite — Google small | $1.23 |
| deepseek-flash (peak) — DeepSeek small | $0.90 |
| Mistral Small 4 — Mistral small | $0.45 |
| gpt-6-luna — OpenAI small | $0.33 |
| Claude Haiku 5.5 — Anthropic small | $0.33 |
Source: Provider pricing pages, read 2026-11-20; CalculatorAI analysis; calculatorai.app
Share / download
With the 1,200-token system prompt served from cache, the day costs 35% less on gpt-6.2-astra ($16.88) and Claude Sonnet 5.5 ($4.22). For a support bot the tier matters far more than the provider, as our cost-per-task comparison also found.
- gpt-6.2-astraOpenAI flagship$22.40
- Claude Fable 5.2Anthropic flagship$22.40
- Claude Opus 5.5Anthropic flagship$11.20
- gemini-4-carbonGoogle flagship$5.60
- gemini-4-argonGoogle flagship$5.60
- grok-4.7xAI flagship$5.28
- deepseek-v4-pro (peak)DeepSeek flagship$3.48
- Mistral Large 4Mistral flagship$1.85cheapest
- gemini-4-carbon, Dec 1Scheduled price, not in effect yet$11.20
Values: $ per session
12× cheaper than gpt-6.2-astra on this input-heavy session, before any quality difference is counted.
Show these figures as a table
| Value ($ per session) | |
|---|---|
| gpt-6.2-astra — OpenAI flagship | $22.40 |
| Claude Fable 5.2 — Anthropic flagship | $22.40 |
| Claude Opus 5.5 — Anthropic flagship | $11.20 |
| gemini-4-carbon — Google flagship | $5.60 |
| gemini-4-argon — Google flagship | $5.60 |
| grok-4.7 — xAI flagship | $5.28 |
| deepseek-v4-pro (peak) — DeepSeek flagship | $3.48 |
| Mistral Large 4 — Mistral flagship | $1.85 |
| gemini-4-carbon, Dec 1 — Scheduled price, not in effect yet | $11.20 |
Source: Provider pricing pages, read 2026-11-20; CalculatorAI analysis; calculatorai.app
Share / download
Reading 90% of each turn's input from cache (the other 10% written at the listed cache-write price) cuts the session by 61% to 79%: gpt-6.2-astra $22.40 → $6.46, Claude Opus 5.5 $11.20 → $3.23, grok-4.7 $5.28 → $2.04. Try your own session in the AI Agent Cost Calculator.
Summarising 1,000 documents costs $68.00 on gpt-6.2-astra and Claude Fable 5.2 down to $0.85 on gpt-6-luna and Claude Haiku 5.5 in real time; OpenAI, Anthropic, Google and Mistral halve that with Batch ($34.00 and $0.43). Only two providers differ:
grok-4.7
- Real time
- $15.00
- Cheapest option
- $12.00
- How
- Batch, 20% off
grok-4.3
- Real time
- $8.75
- Cheapest option
- $7.00
- How
- Batch, 20% off
deepseek-v4-pro
- Real time
- $9.90
- Cheapest option
- $4.95
- How
- No Batch; run off-peak
deepseek-flash
- Real time
- $2.40
- Cheapest option
- $1.20
- How
- No Batch; run off-peak
| Model | Real time | Cheapest option | How |
|---|---|---|---|
| grok-4.7 | $15.00 | $12.00 | Batch, 20% off |
| grok-4.3 | $8.75 | $7.00 | Batch, 20% off |
| deepseek-v4-pro | $9.90 | $4.95 | No Batch; run off-peak |
| deepseek-flash | $2.40 | $1.20 | No Batch; run off-peak |
Source: xAI and DeepSeek pricing pages, read 2026-11-20; CalculatorAI analysis; calculatorai.app
Share / download
Cheapest per tierThe lowest list price at each provider-designated tier
The cheapest flagship is Mistral Large 4 ($1.05), then deepseek-v4-pro at peak ($1.98); the cheapest mid model is gemini-3.8-flash ($1.50), then grok-4.3 ($1.56); the cheapest small models are gpt-6-luna and Claude Haiku 5.5 ($0.20), then Mistral Small 4 ($0.26). The table at the top shows each provider's cheapest model per tier, and the full list is in the dataset (CSV).
The flagship gap is the widest: 15 times per blended token ($16.00 vs $1.05) and 12 times on the input-heavy coding session ($22.40 vs $1.85) — the same two price lists at a different input-to-output mix. It does not mean Mistral Large 4 can replace gpt-6.2-astra, only that a flagship-only budget has the most room to shrink if a cheaper model passes your tests. At the small tier the spread is about 4×. The mid-tier ranking is set to change: when gemini-3.8-flash doubles to $3.00 on January 1, grok-4.3 at $1.56 becomes the cheapest.
What it meansFor builders and for the people who pay the bill
Plan for price rises
Google has announced increases, and DeepSeek's top tier already costs more than the model it replaced. Put the model's current price and its scheduled change in the budget, and re-check pricing pages before every renewal or launch.
Price the task, not the token
Tokenizer changes, thinking tokens and retries all change cost per task. Run 50 real requests per model, count tokens from the API response, and divide by the outputs that pass your test.
Route by tier
In our chatbot workload OpenAI's and Anthropic's small models cost 1–3% of their own flagships. Send routine traffic to a small or mid model and only escalate the hard cases.
Cache and batch before you negotiate
Caching cut the coding session by 61–79% and Batch halves most offline jobs. Those discounts are on the public price list and need no contract.
Keep a second provider ready
With prices moving both ways, being able to switch models is a cost control. A thin abstraction layer and a saved test set make a switch a one-day job.
Choosing between seats and API usage? The AI Subscription vs API Calculator finds the break-even.
LimitsBiases and what this report does not measure
- Disclosure. CalculatorAI makes AI-powered calculators and sells PDF planners; we receive no payment from any provider in this report, and none of the links in it are affiliate or paid links.
- Not quality-adjusted. A newer model that solves a task in one attempt can cost less per finished task than a cheaper one that needs three.
- Real bills differ. Commitments, credits and marketplace pricing can undercut list prices. Some deployments cost more: OpenAI adds 10% for regional endpoints on models released since March 5, 2026; Anthropic lists a 10% premium for regional endpoints on Amazon Bedrock and Google Cloud, and
inference_geo: "us"multiplies every token price by 1.1 on Claude 4.6 and later (Claude Opus 5.5 becomes $4.40 / $22.00). - Prices change. The archived "then" prices are snapshots; cuts between snapshot dates are not captured. At a 1:1 ratio xAI and Mistral look cheaper still.
- Not verified: Amazon Nova (prices loaded by a script we could not read) and Meta's Llama (no first-party price) are excluded. Mistral's FAQ quotes Large at "$0.5 /M tokens in and $1.5 /M tokens out" against $0.70 / $2.10 in its documentation table; we use the table.
Where these numbers come from
All current prices were read on November 20, 2026:
- OpenAI: developers.openai.com/api/docs/pricing (Standard, Batch, Flex, Fast and Ultrafast; ≤272K input tokens).
- Anthropic: platform.claude.com/docs/en/about-claude/pricing (models, cache multipliers, Batch 50%, tokenizer note, Fast mode, 1.1×
inference_geo). - Google: aistudio.google.com/docs/pricing, where ai.google.dev/gemini-api/docs/pricing now redirects (Standard and Batch prices and the December 1 and January 1 changes).
- xAI: docs.x.ai/developers/pricing (20% batch discount, 2× Priority).
- Mistral: docs.mistral.ai/inference/pricing; mistral.ai/pricing (batch 50%, cache up to 90%).
- DeepSeek: api-docs.deepseek.com/quick_start/pricing (peak and off-peak prices).
History, from Wayback Machine snapshots of official pages:
- OpenAI, April 1, 2023 (GPT-4 8K $0.03 / $0.06 per 1K tokens; gpt-3.5-turbo $0.002 per 1K): web.archive.org/web/20230401063613/openai.com/pricing
- Anthropic, March 4, 2024 (Claude 3 Opus $15 / $75, Sonnet $3 / $15, Haiku $0.25 / $1.25): web.archive.org/web/20240311010050/anthropic.com/news/claude-3-family
- Google, June 3, 2024 (Gemini 1.5 Pro $3.50 / $10.50, Gemini 1.5 Flash $0.35 / $1.05, prompts up to 128K): web.archive.org/web/20240603135204/ai.google.dev/pricing
- Mistral, February 29, 2024 (Mistral Large $8 / $24, Mistral Small $2 / $6): web.archive.org/web/20240229171744/docs.mistral.ai/platform/pricing
- xAI, May 1, 2025 (grok-3-beta $3 / $15): web.archive.org/web/20250501021006/docs.x.ai/docs/models
- DeepSeek, January 31, 2025 (deepseek-chat $0.27 / $1.10, deepseek-reasoner $0.55 / $2.19): web.archive.org/web/20250131222749/api-docs.deepseek.com/quick_start/pricing
Workload costs ignore tool fees, search grounding, cache storage and retries (which raise bills) and off-peak scheduling or negotiated discounts (which lower them).
How to citeUse these numbers in your own work
Full citation: CalculatorAI. (2026, November 20). AI API Prices Are Rising Again: The 2027 Cost Report. CalculatorAI. calculatorai.app/blog/ai-api-cost-report-2027
Short in-line citation: (CalculatorAI, AI API Cost Report 2027, Nov 2026) — or in prose: "according to CalculatorAI's 2027 AI API Cost Report".
Data: Download the full dataset (CSV) — every price in the table above, licensed CC BY 4.0.
Quote any figure with a link to this page. Each chart has Link, PNG, CSV, Embed and Cite buttons. Prices were verified on November 20, 2026; after the 1st of a later month, check the update log for the latest re-check.
Questions about the data: research@calculatorai.app
Update logWhat changed in this report and when
- 2026-11-20 — First edition. Prices of 18 models from OpenAI, Anthropic, Google, xAI, Mistral and DeepSeek, compared with archived 2023–2025 prices and priced on three workloads.
- Next update: monthly, on the 1st; the first re-check, December 1, 2026, is the day Google's new Gemini 4 prices take effect.
FAQFrequently asked questions
Are AI API prices still going down?
At the small end, yes. OpenAI's gpt-6-luna and Anthropic's Claude Haiku 5.5 cost $0.10 per million input tokens, and their blended prices are 60–90% below their 2023–2024 predecessors. At the top end the trend has broken: Google's Gemini 4 flagships double on December 1, 2026, gemini-3.8-flash doubles on January 1, 2027, and DeepSeek's top tier costs about twice its January 2025 price at peak hours.
Which AI API prices are going up?
Google's gemini-4-carbon and gemini-4-argon double on December 1, 2026, from $2 / $10 to $4 / $20 per million input / output tokens, which puts gemini-4-carbon at $8.00 blended against $5.25 for the Gemini 1.5 Pro it replaced. gemini-3.8-flash doubles on January 1, 2027, to $1.50 / $7.50. DeepSeek's top tier, deepseek-v4-pro, already costs $1.98 blended at peak hours, about twice deepseek-reasoner's $0.96 in January 2025.
What is the cheapest flagship LLM API in November 2026?
By blended list price at a 3:1 input-to-output ratio, Mistral Large 4 at $1.05 per million tokens, followed by deepseek-v4-pro at $1.98 during peak hours. The most expensive flagships are gpt-6.2-astra and Claude Fable 5.2 at $16.00. The most expensive flagship costs 15 times the cheapest per blended token, or 12 times on our input-heavy coding session. These are list prices, not quality rankings.
How much has GPT-4-class pricing fallen since 2023?
GPT-4 cost $30 per million input tokens and $60 per million output tokens in April 2023, or $37.50 blended at 3:1. gpt-6.2-astra costs $8 and $40, or $16.00 blended. That is 57% lower, an average of about 21% a year.
What is a blended price per million tokens?
One number that combines the input and output prices at a stated ratio. This report uses three input tokens per output token: (3 × input price + output price) ÷ 4. Change the ratio and the ranking can change, because some models have much higher output prices than others.
How much does prompt caching save?
In our 40-turn coding-agent session, reading 90% of each turn's input from cache cut the cost by 61% to 79% depending on the model. Savings are smaller where the cached-input price is a larger share of the normal price, as on xAI's grok-4.7.
Why do Claude prices look lower per token than per page of text?
Anthropic says Claude 4.7 and later models use a tokenizer that produces about 30% more tokens for the same text. A price cut per token is therefore partly offset by more tokens per page.
Budget from your own numbers
This is one day's price lists on three stated workloads. Your bill depends on your token counts, cache hit rate and attempts per task: put them into the AI Token Cost Calculator. We re-check these prices on the 1st of every month.





