Straitly compass logostraitly

RESEARCH · 75 STATISTICS

LLM inference cost statistics

LLM inference keeps getting cheaper: Epoch AI measured a median 50x yearly fall in the price of reaching a fixed level of performance.[1] Usage is growing faster, so most bills still go up.

Last updated . Every number links to its source.

Read the statisticsCompare model prices

KEY TAKEAWAYS

  • 50x

    Median yearly fall in the price of a fixed level of model performance (Epoch AI, 2025). [1]

  • 3.2Q

    Tokens Google processed per month in May 2026, 7x more than a year earlier. [11]

  • $177B

    Enterprise AI deployment spend by September 2026, up 5.4x in a year (Menlo Ventures). [7]

  • 80%

    Share of OpenRouter tokens now driven by agents rather than people. [7]

What do frontier tokens cost today?

Inside one provider's lineup the top model costs many times the cheapest, and list prices can rise as well as fall.

  • $4 / $20

    Claude Opus 5.5 list price per million input / output tokens.[15] Anthropic docs, Checked 9 November 2026. Pricing table

  • $10 / $50

    Claude Fable 5.1 list price per million input / output tokens, Anthropic's top tier.[15] Anthropic docs, Checked 9 November 2026. Pricing table

  • 100x

    Spread in OpenAI's flagship input prices: GPT-6 Luna at $0.10 and GPT-6 Astra at $10.00 per million tokens.[17] OpenAI, Checked 9 November 2026

  • 2x

    Gemini 4 Argon input goes from $2.00 to $4.00 per million tokens on 1 December 2026. List prices can rise, not only fall.[19] Google, Checked 9 November 2026

List price per million input tokens, November 2026

Standard requests from each provider's pricing page. DeepSeek shows its off-peak rate.

  1. GPT-6 Luna$0.10
  2. Claude Haiku 5.5Prompts up to 100K tokens$0.10
  3. DeepSeek V4.1 FlashOff-peak$0.15
  4. Gemini 4 Argon$4.00 from 1 Dec 2026$2.00
  5. GPT-6.1 Sol$2.00
  6. Claude Sonnet 5.5$2.00
  7. Claude Opus 5.5$4.00
  8. GPT-6 Astra$10.00
  9. Claude Fable 5.1$10.00

Bar lengths use a log scale so small and large values fit on one chart.

[17] OpenAI, Checked 9 November 2026; [15] Anthropic docs, Checked 9 November 2026; [19] Google, Checked 9 November 2026; [20] DeepSeek API docs, Checked 9 November 2026

All 7 more stats in this section
What do frontier tokens cost today?
$15 to $5Anthropic's Opus tier input price per million tokens: Claude Opus 4.1 lists at $15, Claude Opus 5 at $5.[14] Anthropic, Checked 9 November 2026
5xClaude Haiku 5.5 input goes from $0.10 to $0.50 per million tokens once a prompt is over 100,000 tokens.[15] Anthropic docs, Checked 9 November 2026
50%Batch API discount on both input and output tokens at Anthropic, for work that can wait for asynchronous processing.[14] Anthropic, Checked 9 November 2026
HalfDeepSeek off-peak rates are half of peak: V4.1 Flash input is $0.15 per million tokens off-peak and $0.30 at peak.[20] DeepSeek API docs, Checked 9 November 2026
$0.73Median cost per million tokens across models in OpenRouter's usage study.[13] OpenRouter and a16z, Dec 2025
2.5xOpenAI's 12-hour cache write rate on GPT-5.6 and later, against 1.25x for the 30-minute default.[18] OpenAI docs, Checked 9 November 2026
$1B+Yearly saving Google says a company running about 1 trillion tokens a day would get by moving 80% of frontier workloads to Gemini 3.5 Flash. A vendor claim.[11] Google, 19 May 2026

How much do companies spend on AI?

Cheaper tokens have not shrunk budgets. Spend is rising fastest on the most capable closed models.

  • $177B

    Enterprise AI deployment spend by September 2026, up from $33B a year earlier: 5.4x in one year.[7] Menlo Ventures, Nov 2026, survey fielded Sep 2026

  • 86%

    Share of enterprise LLM API spend going to Anthropic and OpenAI, up from 67% a year earlier.[7] Menlo Ventures, Nov 2026, survey fielded Sep 2026

Share of tokens vs share of spend

Year to September 2026, Menlo Ventures estimates.

Share of tokensShare of spend

Frontier models
42% share of tokens
63% share of spend
Open-source models
19% share of tokens
4% share of spend

[7] Menlo Ventures, Nov 2026, survey fielded Sep 2026

All 9 more stats in this section
How much do companies spend on AI?
18% to 42%Frontier models' share of tokens in one year. Their share of spend went from 35% to 63%.[7] Menlo Ventures, Nov 2026, survey fielded Sep 2026
$12.5BEnterprise spend on foundation model APIs in 2025, out of $37B total generative AI spend.[8] Menlo Ventures, 9 Dec 2025
$3.5B to $8.4BModel API spend from November 2024 to mid-2025: more than double in six months.[9] Menlo Ventures, 31 Jul 2025
74%Startups saying most of their compute is inference, up from 48% a year earlier. Among large enterprises it was 49%, up from 29%.[9] Menlo Ventures, 31 Jul 2025
11%Teams that switched model provider in the past year. 66% upgraded to a newer model from the same provider.[9] Menlo Ventures, 31 Jul 2025
$150B+Google's capital expenditure in 2025, as major cloud providers accelerated spending.[6] Stanford HAI, Apr 2026
29.6 GWAI data center power capacity in 2025, comparable to New York state at peak demand.[5] Stanford HAI, Apr 2026
5,427Data centers in the United States, more than ten times any other country.[5] Stanford HAI, Apr 2026
60%+Nvidia's share of total AI compute, with Google and Amazon supplying much of the rest.[5] Stanford HAI, Apr 2026

How fast is token usage growing?

Faster than prices fall, driven by agents and coding.

  • 3.2Q

    Tokens Google processed per month across its surfaces in May 2026, 7x the roughly 480 trillion of a year earlier.[11] Google, 19 May 2026

  • 80%

    Share of OpenRouter tokens driven by agents, roughly 7x human-driven usage.[7] Menlo Ventures, Nov 2026, survey fielded Sep 2026

Tokens Google processes per month

Across Google's products and APIs, as reported at Google I/O each May.

  1. May 20249.7T
  2. May 2025480T
  3. May 20263.2Q

Bar lengths use a log scale so small and large values fit on one chart.

[11] Google, 19 May 2026; [10] Google, 20 May 2025

All 13 more stats in this section
How fast is token usage growing?
19BTokens per minute Google's model APIs were processing in May 2026.[11] Google, 19 May 2026
375+Google Cloud customers that each processed more than 1 trillion tokens in the 12 months to May 2026.[11] Google, 19 May 2026
4xYear-over-year growth in Microsoft Foundry customers running at a 1 trillion token annualized rate.[12] Microsoft Investor Relations, 29 Jul 2026
160T+Tokens a week processed by OpenRouter, according to Menlo's methodology note.[7] Menlo Ventures, Nov 2026, survey fielded Sep 2026
37xGrowth in OpenRouter agent traffic since it overtook human traffic in February 2026, against 5x for human-driven usage.[7] Menlo Ventures, Nov 2026, survey fielded Sep 2026
2.9QOpen-source tokens per month, up 22.5x in a year. Closed-source tokens grew 12.1x.[7] Menlo Ventures, Nov 2026, survey fielded Sep 2026
480TTokens Google processed per month in May 2025, up from 9.7 trillion a year earlier: 50 times more.[10] Google, 20 May 2025
4xGrowth in average prompt tokens per request on OpenRouter, from about 1.5K to over 6K. Completions nearly tripled, from about 150 to 400 tokens.[13] OpenRouter and a16z, Dec 2025
11% to 50%+Programming's share of OpenRouter token volume, early 2025 to late 2025.[13] OpenRouter and a16z, Dec 2025
0.5-0.7%Rise in usage for a 10% price cut. The authors describe demand as relatively price-inelastic.[13] OpenRouter and a16z, Dec 2025
1/3Share of OpenRouter usage served by open-weight models by late 2025.[13] OpenRouter and a16z, Dec 2025
100TTokens of real-world LLM traffic analyzed in OpenRouter's State of AI study, across tasks, geographies and time.[13] OpenRouter and a16z, Dec 2025
~52%Roleplay's share of open-source model tokens on OpenRouter. Roleplay and programming together are the majority.[13] OpenRouter and a16z, Dec 2025

How do caching and agents change the bill?

Agents resend their context, and cache reads cost a fraction. See caching savings and agent cost.

  • 95%+

    Share of an agent's tokens that cache reads can make up. At Opus 5.5 prices that brings an agent session to about $0.70 per million tokens.[7] Menlo Ventures, Nov 2026, survey fielded Sep 2026

  • 41-80%

    API cost reduction from prompt caching across OpenAI, Anthropic and Google, measured on 500+ agent sessions with 10,000-token system prompts.[22] Lumer et al., 9 Jan 2026

What one million Claude Opus 5.5 tokens cost

List price by token type, plus Menlo's estimate for a blended agent session.

  1. Cache read$0.20
  2. Agent session, blendedMenlo estimate~$0.70
  3. Fresh input$4.00
  4. 5-minute cache write$5.00
  5. Output$20.00

[15] Anthropic docs, Checked 9 November 2026; [7] Menlo Ventures, Nov 2026, survey fielded Sep 2026

All 16 more stats in this section
How do caching and agents change the bill?
0.1xAnthropic cache reads cost 0.1x the base input price. Writes cost 1.25x for a 5-minute cache and 2x for a 1-hour cache.[15] Anthropic docs, Checked 9 November 2026
0.025xCache hits on Claude Fable 5.1 and Mythos 5.1. Opus 5.5 and Sonnet 5.5 hits cost 0.05x base input.[15] Anthropic docs, Checked 9 November 2026
2.15xOpenAI: one cache write and nine full reads cost 2.15x the prefix's input price, against 10x without caching.[18] OpenAI docs, Checked 9 November 2026
1,024Minimum cacheable prompt length in tokens on OpenAI GPT-5.6 and later.[18] OpenAI docs, Checked 9 November 2026
2%DeepSeek V4.1 Flash cache hits cost $0.003 per million tokens off-peak, against $0.15 for a cache miss.[20] DeepSeek API docs, Checked 9 November 2026
$0.10Gemini 4 Argon implicit cache price per million tokens, against $2.00 for uncached input (through 30 November 2026).[19] Google, Checked 9 November 2026
90%Anthropic's launch claim: caching cuts cost by up to 90% and latency by up to 85% for long prompts.[16] Anthropic, Page dated 14 Aug 2025
91%Share of input tokens served from cache in the sample session readout in Anthropic's Claude Code cost docs (14 requests, 2 misses).[21] Claude Code docs, Checked 9 November 2026. Example output
$34BEnterprise spend on coding, the largest AI app category in 2026. Claude Code and Codex account for $24B of it (70%).[7] Menlo Ventures, Nov 2026, survey fielded Sep 2026
50%Developers using AI coding tools daily in 2025, 65% in top-quartile organizations.[8] Menlo Ventures, 9 Dec 2025
$13Average Claude Code cost per developer per active day in enterprise deployments, $150-250 per developer per month. 90% of users stay below $30 a day.[21] Claude Code docs, Checked 9 November 2026
20K+Input tokens that code understanding, debugging and generation requests routinely exceed on OpenRouter.[13] OpenRouter and a16z, Dec 2025
21-36%Cut in total computational cost from removing wasted context in coding-agent trajectories (input tokens fell 39.9% to 59.7%) with the same agent performance.[23] Xiao et al., 28 Sep 2025
+6.8%Billed cost increase from the most aggressive tool-output compression, which removed 38.4% of those tokens. Cache creation and reads dominated input-side cost.[24] Weinberger and Hozez, 13 Jul 2026
18-31%Average token cost reduction from an external controller across Claude Code, Codex and GitHub Copilot CLI, with comparable task success.[25] Li et al., 15 Oct 2026
32 in 571Prompt-caching mis-integrations found across 571 open-source agent projects. Most such bugs fail silently.[26] Zhang et al., 26 Oct 2026

What do LLM gateways and routers charge?

A percentage fee grows with your bill.

  • $0.80

    Minimum OpenRouter card fee per credit purchase. Crypto payments are charged 5%.[28] OpenRouter docs, Checked 9 November 2026

  • $25,000

    Monthly list-price inference OpenRouter allows through your own provider keys (BYOK) with no fee. Above it, the fee is 5%.[27] OpenRouter, Checked 9 November 2026

  • 365 days

    After purchase, OpenRouter may expire unused credits. Platform fees are non-refundable.[29] OpenRouter, Last updated 31 Aug 2026

Fee on top of the provider's list price

Per dollar of credit or usage, from each gateway's own pricing page.

  1. OpenRouter, Business plan8%
  2. OpenRouter, Standard plan5.5%
  3. OpenRouter, crypto payments5%
  4. OpenRouter, own keys over $25,000 a month5%
  5. Vercel AI Gateway0%
  6. StraitlyBills 40% below the published rate0%

[27] OpenRouter, Checked 9 November 2026; [28] OpenRouter docs, Checked 9 November 2026; [30] Vercel docs, Checked 9 November 2026; [32] Straitly, Checked 9 November 2026

“OpenRouter's fee is charged when you buy credits, 5.5% on Standard and 8% on Business, never on individual requests.”
[27] OpenRouter, Checked 9 November 2026
All 5 more stats in this section
What do LLM gateways and routers charge?
5.5%OpenRouter's fee on credit purchases on the Standard plan. The Business plan fee is 8%. Inference is billed at the provider's list price.[27] OpenRouter, Checked 9 November 2026
600+Models OpenRouter lists on its Standard and Business plans, from 80+ providers.[27] OpenRouter, Checked 9 November 2026
0%Vercel AI Gateway markup on tokens: provider list price, including with your own key on the paid tier.[30] Vercel docs, Checked 9 November 2026
$0Cloudflare AI Gateway core features (analytics, caching and rate limiting) are free on all plans.[31] Cloudflare docs, Last updated 2 Nov 2026
40%Straitly bills each request 40% below the provider's published rate, with no fee on top, no subscription and no minimum.[32] Straitly, Checked 9 November 2026

What does this mean for your bill?

Per-token prices fall, but usage grows faster and buyers trade up, so most teams see bills rise.

  1. 1The price falls. A fixed level of performance gets a median 50x cheaper each year.[1]
  2. 2Usage grows faster. Google's monthly tokens grew 7x in a year, to over 3.2 quadrillion.[11]
  3. 3The bill grows. Enterprise AI deployment spend went from $33B to $177B in a year.[7]
STRAITLY, 9 NOVEMBER 2026

  • 40%

    below the provider's published rate on Claude, GPT, Gemini and Grok, with no fee on top.[32] Straitly, Checked 9 November 2026

  • 40%

    median price below official list across 102 matched models in our 164-model snapshot.[33] Straitly, Snapshot 9 November 2026

  • 129 of 141

    shared models cheaper than the cheapest OpenRouter seller, fee included.[33] Straitly, Snapshot 9 November 2026

Per-model prices: LLM API price index.

What do people ask about LLM inference costs?

Short answers, each drawn from the sources below.

How fast are LLM API prices falling?

Fast, but unevenly. Epoch AI found the price of reaching a fixed level of performance fell between 9x and 900x a year depending on the task, with a median of 50x (March 2025). a16z estimates about 10x a year for equivalent performance (November 2024). Menlo Ventures reports frontier labs cut small-model prices by 80% or more in the year to September 2026.

If prices are falling, why is my LLM bill going up?

Usage is growing faster than prices fall, and buyers trade up to newer, pricier models. Google went from about 480 trillion to over 3.2 quadrillion tokens a month between May 2025 and May 2026. Agents now drive 80% of OpenRouter tokens, and frontier models went from 18% to 42% of tokens in a year.

How much does prompt caching save?

Cache reads cost 0.1x normal input at Anthropic and OpenAI (0.05x or less on the newest models), and about 2% of the miss price on DeepSeek. A 2026 study measured 41-80% lower API costs on agent sessions. Menlo reports cache reads can be 95% or more of an agent's tokens.

How much does a coding agent cost per developer?

Anthropic's Claude Code docs put the average at about $13 per developer per active day and $150-250 per developer per month in enterprise deployments, with 90% of users below $30 per active day. Enterprise coding spend reached $34B in 2026, per Menlo Ventures.

What do LLM gateways and routers charge?

OpenRouter charges 5.5% on credit purchases on its Standard plan and 8% on Business, with a $0.80 minimum card fee. Vercel AI Gateway charges no markup on tokens. Straitly charges no fee and bills each request below the provider's published rate.

Pixel art: water runs down a hillside of terraced fields one step at a time to the sea.

Pay 40% below list

One OpenAI-compatible API for Claude, GPT, Gemini and Grok. No fee, no subscription, no minimum.

Sources and method

How these numbers were checked

Every source was fetched and read on 9 November 2026, and only numbers the source states are shown. Provider prices are list prices in USD per million tokens for standard requests unless a row says otherwise. Research and survey results are reported as the authors state them; see each source for its method and sample. Menlo Ventures figures are estimates from a survey of nearly 400 US enterprise decision-makers plus a bottoms-up market model. Straitly figures come from the live Straitly catalog snapshot on the same day.

  1. [1]LLM inference prices have fallen rapidly but unequally across tasks Epoch AI (Cottier et al.), 12 Mar 2025
  2. [2]Welcome to LLMflation: LLM inference cost is going down fast Andreessen Horowitz (Guido Appenzeller), 12 Nov 2024
  3. [3]The 2025 AI Index Report: top takeaways Stanford HAI, 2025
  4. [4]The 2026 AI Index Report Stanford HAI, Apr 2026
  5. [5]2026 AI Index, chapter 1: Research and development Stanford HAI, Apr 2026
  6. [6]2026 AI Index, chapter 4: Economy Stanford HAI, Apr 2026
  7. [7]2026: The State of Enterprise AI Menlo Ventures, Nov 2026, survey fielded Sep 2026
  8. [8]2025: The State of Generative AI in the Enterprise Menlo Ventures, 9 Dec 2025
  9. [9]2025 Mid-Year LLM Market Update Menlo Ventures, 31 Jul 2025
  10. [10]Google I/O 2025 keynote Google, 20 May 2025
  11. [11]I/O 2026: Welcome to the agentic Gemini era Google (Sundar Pichai), 19 May 2026
  12. [12]FY26 Q4 earnings conference call Microsoft Investor Relations, 29 Jul 2026
  13. [13]State of AI: an empirical 100 trillion token study with OpenRouter OpenRouter and a16z (Aubakirova et al.), Dec 2025
  14. [14]Claude API pricing Anthropic, Checked 9 November 2026
  15. [15]Prompt caching Anthropic docs, Checked 9 November 2026
  16. [16]Prompt caching with Claude Anthropic, Page dated 14 Aug 2025
  17. [17]API pricing OpenAI, Checked 9 November 2026
  18. [18]Prompt caching guide OpenAI docs, Checked 9 November 2026
  19. [19]Gemini Developer API pricing Google, Checked 9 November 2026
  20. [20]Models and pricing DeepSeek API docs, Checked 9 November 2026
  21. [21]Manage costs effectively Claude Code docs, Checked 9 November 2026
  22. [22]Don't Break the Cache: An Evaluation of Prompt Caching for Long-Horizon Agentic Tasks Lumer et al., arXiv 2601.06007, 9 Jan 2026
  23. [23]Reducing Cost of LLM Agents with Trajectory Reduction Xiao et al., arXiv 2509.23586, 28 Sep 2025
  24. [24]Token Reduction Is Not Cost Reduction Weinberger and Hozez, arXiv 2607.12161, 13 Jul 2026
  25. [25]TokenEconomizer: Experience-Driven Black-Box Cost Control for Coding Agents Li et al., arXiv 2610.22732, 15 Oct 2026
  26. [26]Silent and Costly: Characterizing and Detecting Prompt-Caching Mis-Integration in LLM Agents Zhang et al., arXiv 2610.34817, 26 Oct 2026
  27. [27]OpenRouter pricing OpenRouter, Checked 9 November 2026
  28. [28]OpenRouter FAQ: pricing and fees OpenRouter docs, Checked 9 November 2026
  29. [29]OpenRouter Terms of Service OpenRouter, Last updated 31 Aug 2026
  30. [30]AI Gateway pricing Vercel docs, Checked 9 November 2026
  31. [31]AI Gateway pricing Cloudflare docs, Last updated 2 Nov 2026
  32. [32]Straitly pricing Straitly, Checked 9 November 2026
  33. [33]Straitly price snapshot vs official list and OpenRouter (164 models) Straitly, Snapshot 9 November 2026
Show all 33 sources