Home / Indexes / AI Business & Economics / LLM Token Price Index 2026
Thank you for reading this post, don't forget to subscribe!
AI Business & Economics · Index
LLM TOKEN PRICE INDEX 2026
The price of a given level of AI capability is collapsing faster than almost any technology in history — roughly an order of magnitude per year. Every figure is linked to its primary source and dated.
What changed · 1 October 2026
- Anthropic pricing corrected — the table was stale while marked High confidence. Previous: Opus 4.8 at $5.00/$25.00, Sonnet 4.6 at $3.00/$15.00 (verified Jul 2026). New: Opus 5.5 at $4.00/$20.00, Sonnet 5.5 at $2.00/$10.00, and Fable 5.1 at $10.00/$50.00 added. Verified against Anthropic’s own pricing page on 1 Oct 2026. Why: Opus 5.5 shipped 22 Sep and Sonnet 5.5 on 28 Sep, both priced below the models they replace. Haiku 4.5 unchanged at $1.00/$5.00.
- 2026 market pricing added. The page previously ended its series at October 2024. It now carries the Silicon Data LLM Token Expenditure Index at $0.97 per million tokens (1 Sep 2026) and −81.3% since March 2023.
- Unchanged and still correct: the 280× decline, the $20.00 → $0.07 series and the ~10×/yr rate. These are dated historical readings from Stanford HAI and a16z/Epoch, not current-state figures, so they do not go stale.
Key takeaways
- ~280× in ~2 years. A GPT-3.5-level system fell from $20.00 per million tokens (Nov 2022) to $0.07 (Oct 2024). (Stanford HAI)
- “LLMflation.” Cost to reach a fixed capability bar drops roughly 10× per year; Epoch AI finds a median ~50×/yr, ranging up to ~900×/yr by task. (a16z / Epoch)
- Frontier prices fall slower. The cliff is in capability-per-dollar, not the sticker price of the newest model.
- Falling token costs are the engine behind agentic and test-time-compute workloads — use cases uneconomic in 2023 are routine by 2026.
280×
~10×/yr
$0.07
The capability-cost curve
The clearest single series comes from Stanford HAI’s AI Index 2025, which tracks the inference price of reaching GPT-3.5 performance (MMLU 64.8) over time. The decline is steep and roughly log-linear — the cost to run a GPT-3.5-class model fell more than 280× in about two years, from $20.00 per million tokens in November 2022 to $0.07 in October 2024 (Stanford HAI AI Index 2025).
| Period | GPT-3.5-level inference cost ($/1M tokens) |
|---|---|
| Nov 2022 | $20.00 |
| Aug 2023 | $1.80 |
| Jun 2024 | $0.18 |
| Oct 2024 | $0.07 |
“The cliff is in capability-per-dollar, not the sticker price of the newest model.”
The market price in 2026: below a dollar
The Stanford series above ends in October 2024. The clearest 2026 reading comes from Silicon Data’s LLM Token Expenditure Index, which fell to 97 cents per million tokens on 1 September 2026 — a record low since the index launched in late 2025, and less than half the peak it set earlier that summer. Over the longer run, frontier LLM token prices sit roughly 81.3% below their March 2023 level.
Read that number carefully. The index is usage-weighted — it blends provider list pricing with real consumption volumes. So 97 cents describes what the market actually pays on average, dominated by cheap and mid-tier models, not what frontier capability costs. Anthropic’s Fable 5.1 is $10 input and $50 output per million tokens on the same date. Both figures are correct and they measure different things, and conflating them is the most common error in coverage of AI pricing.
| 2026 indicator | Value | Source · confidence |
|---|---|---|
| LLM Token Expenditure Index (usage-weighted) | $0.97 / 1M | Silicon Data, 1 Sep 2026 · Medium |
| Index vs its own summer 2026 peak | less than half | Silicon Data · Medium |
| Frontier token prices vs March 2023 | −81.3% | Reported Sep 2026 · Medium |
| Claude Opus 5.5 vs Opus 5, cost to run | −40% | Anthropic, 22 Sep 2026 · High |
| Claude Sonnet 5.5 vs Sonnet 5 | up to −30%, 30% faster | Anthropic, 28 Sep 2026 · High |
What is pushing it down. Three forces are named consistently: capable open-weight Chinese models resetting the floor, Moonshot’s Kimi K3 in particular; OpenAI cutting prices on two GPT-5.6 models in late July 2026; and dynamic pricing across the field. The consequence is margin pressure exactly where it is least welcome — on OpenAI and Anthropic, both widely discussed as IPO candidates. A falling index also conditions buyers to expect lower rates, eroding pricing power even where capability has improved.
The counter-current worth holding onto: unit prices are falling while total bills rise. Agentic and test-time-compute workloads consume orders of magnitude more tokens per task than a 2023 chat turn did, so cheaper tokens have not produced cheaper AI for most buyers. Token management is becoming the competitive variable, not token price.
Current frontier API pricing
Below is current published API pricing per 1M tokens, verified against Anthropic’s own pricing page on 1 October 2026. Other vendors’ list prices vary and should be confirmed against each provider’s pricing page before relying on them. Note that batch processing cuts Anthropic prices by 50%, US-only inference costs 1.1× standard, and Opus 5.5 fast mode costs 2×.
| Model | Input ($/1M) | Output ($/1M) | Confidence |
|---|---|---|---|
| Anthropic Claude Fable 5.1 | $10.00 | $50.00 | High |
| Anthropic Claude Opus 5.5 | $4.00 | $20.00 | High |
| Anthropic Claude Sonnet 5.5 | $2.00 | $10.00 | High |
| Anthropic Claude Haiku 4.5 | $1.00 | $5.00 | High |
| Anthropic cached reads (all tiers) | $0.10–$0.25 | — | High |
| Other frontier vendors (OpenAI, Google, DeepSeek, xAI) | List prices vary; confirm on each provider’s page | Vendor-listed |
Why “price per token” understates the drop
Capability is rising at the same time. A dollar buys both cheaper tokens and a more capable model, so cost-per-useful-task falls faster than cost-per-token. Open-weight models reset the floor — competitive open models (DeepSeek, Llama-class) pull hosted prices down across the board. And frontier is the exception: the newest top model is priced for willingness-to-pay; the deflation shows up once a capability tier becomes commoditized.
The numbers in full
| Indicator | Value | Source |
|---|---|---|
| GPT-3.5-level inference cost (Nov 2022) | $20.00 / 1M | Stanford HAI |
| GPT-3.5-level inference cost (Oct 2024) | $0.07 / 1M | Stanford HAI |
| Total decline, GPT-3.5-class | >280× | Stanford HAI |
| Cost decline at fixed capability | ~10×/yr | a16z |
| Task-specific cost decline (longer run) | ~50×/yr median · up to ~900×/yr | Epoch AI |
| LLM Token Expenditure Index (usage-weighted) | $0.97 / 1M · 1 Sep 2026 | Silicon Data |
| Frontier token prices vs March 2023 | −81.3% | Reported Sep 2026 |
| Claude Fable 5.1 (input / output) | $10.00 / $50.00 | Anthropic |
| Claude Opus 5.5 (input / output) | $4.00 / $20.00 | Anthropic |
| Claude Sonnet 5.5 (input / output) | $2.00 / $10.00 | Anthropic |
Frequently asked
How fast are AI inference costs falling?
For a fixed capability level, roughly 10× per year per a16z; Stanford HAI documents a ~280× fall for GPT-3.5-class inference between November 2022 and October 2024. (a16z / Stanford HAI)
Does this mean frontier models are getting cheaper?
Not as quickly. The steep decline applies to reaching a given capability bar. The newest frontier models hold their price longer; deflation hits a tier once cheaper or open-weight models match it.
What does a GPT-3.5-level system cost today?
About $0.07 per million tokens as of October 2024, down from $20.00 in November 2022 — a more than 280× decline. (Stanford HAI)
Cite this page
The AI Index (2026). LLM Token Price Index 2026. Retrieved July 2026, from report-ai.org/indexes/ai-economics/llm-token-price-index/
Related: AI Business & Economics · Compare year over year · Tokens · Large Language Model
On this page
- Key takeaways
- The capability-cost curve
- The market price in 2026
- Frontier API pricing
- Why per-token understates
- The numbers in full
- Frequently asked
- Cite this page
Primary sources
- Stanford HAI — AI Index 2025. GPT-3.5-level capability-cost curve
- a16z / Epoch AI — LLMflation. ~10×/yr · up to ~900×/yr task-specific (median ~50×)
- Anthropic — provider pricing, verified 1 Oct 2026. Fable / Opus / Sonnet / Haiku
- Silicon Data — LLM Token Expenditure Index (usage-weighted). $0.97/1M, 1 Sep 2026
280×
Drop in GPT-3.5-class inference cost in under two years — $20.00 to $0.07 per 1M tokens.
Share