What AI used to cost — and what it costs now.
One consistent place to see how AI prices have moved, across text and creative work. The numbers live here; the why lives on our two sister sites.
Lean toward price, brains, speed or eco — watch the models climb and fall, and see your #1 pop up.
Text & token pricing
Per 1M tokens (input / output). Representative standard list rates from OpenAI, Anthropic, Google Gemini and DeepSeek, checked 22 Oct 2026 against TokenScale’s nightly checked record, which marks any carried observation. Active promotions are excluded here. The complete market is the 22-provider, 66-tier table on TokenScale.
Snapshot: standard list rates
| Model | Input | Output |
|---|---|---|
| Claude Opus 5.5 officialAnthropic | $4.00 | $20.00 |
| GPT-5.6 Terra officialOpenAI | $2.00 | $12.00 |
| Pro 3.1 officialGoogle Gemini | $2.00 | $12.00 |
| Claude Sonnet 5.5 officialAnthropic | $2.00 | $10.00 |
| GPT-6 Sol officialOpenAI | $2.00 | $10.00 |
| Flash 3.6 officialGoogle Gemini | $1.50 | $7.50 |
| V4 Pro · peak rate officialDeepSeek | $1.32 | $3.96 |
| Claude Haiku 4.5 officialAnthropic | $1.00 | $5.00 |
| Flash-Lite 3.5 officialGoogle Gemini | $0.30 | $2.50 |
| V4.1 Flash · peak rate officialDeepSeek | $0.30 | $1.20 |
| GPT-6 Luna officialOpenAI | $0.10 | $0.50 |
The trend — input $ / 1M tokens over time
Log scale. Flagship input fell from $30 for GPT‑4 in March 2023, then rose at the newest frontier: Claude Fable 5 is $10.00 per 1M input tokens. The budget record is not a straight line either. Gemini Flash-Lite moved from $0.10 to $0.30 in July 2026. An unsupported OpenAI Luna $0.20 reading was withdrawn on 11 August; a fresh official same-model repricing on 12 August then set Luna to $0.20. Both events remain in Price Moves.
Why did a token ever cost that — and why is "cheap" so confusing? TokenScale explains it in real reading.
TokenScale →Image, video & audio pricing
Representative cost per usable unit — one image, one second of video, one minute of voice. These creative rates still carry PixelScale’s June 2026 snapshot, they were not part of the 8 August text refresh, so treat them as older than the table above.
Snapshot: June 2026 creative rates, not current
| Model | Unit | Price |
|---|---|---|
| Imagen 4 Fast officialGoogle | per image | $0.020 |
| FLUX 1.1 Pro officialBlack Forest Labs | per image | $0.040 |
| GPT Image 1.5 est.OpenAI | per image | $0.040 |
| Midjourney Std est.Midjourney | per image | $0.035 |
| Veo 3.1 Fast officialGoogle | per second | $0.100 |
| Sora 2 officialOpenAI | per second | $0.100 |
| Gen-4 Turbo officialRunway | per second | $0.050 |
| Dream Machine est.Luma | per second | $0.080 |
| Multilingual v2 est.ElevenLabs | per min voice | $0.160 |
The trend — video $ / second (Veo, illustrative)
Illustrative, Google Veo 3.1 tiers (public pricing + creator reports, June 2026). Some API rates fell even as subscription plans felt tighter — the unit changed, not just the number.
One final image is rarely one generation. PixelScale shows the real project cost after retries, upscales and edits.
PixelScale →cheaper to reach GPT‑4‑class text in three years
GPT‑4 launched at $30 / $60 per 1M tokens in March 2023. A capable budget model now runs $0.10 per 1M input (GPT-6 Luna), a 300× drop. Go past the frontier labs and TokenScale's floor across all 22 providers is $0.02 per 1M (Llama 3.1 8B on DeepInfra), which is 1,500×.
The twist worth remembering: the very newest flagships have started ticking up again. Cheaper isn't a straight line — it depends entirely on which tier you pick.
This comparison has changed as current models changed. Nothing about 2023 was rewritten. The present 300× figure compares GPT-4's $30 launch input rate with GPT-6 Luna's current $0.10 input rate (GPT-5.6 Luna's $0.20 gave 150× until 24 Sep 2026); lower specialist and hosted-open-model rates exist elsewhere in TokenScale's complete board.
How to read this
Text list rates: OpenAI API pricing · Anthropic / Claude pricing · Google Gemini API pricing · DeepSeek API pricing, via TokenScale’s nightly check. Active promotions are excluded. DeepSeek rows use the labelled peak rate; off-peak is 50% lower outside 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, with weekends and all other hours off-peak. Creative rates: Black Forest Labs · Runway · ElevenLabs · public provider docs and creator reports, checked June 2026 and not yet re-verified.