Genfeed.ai
Make your own

Are AI tokens getting cheaper or more expensive? Both.

Token prices are collapsing per unit of capability, flat or rising at the frontier, and the bill per agent job keeps climbing. Four charts and a conclusion on where prices are heading.

For three years the rule was simple: tokens get cheaper. This month Elon Musk told X that serving Grok Bot was "unsustainably high relative to revenue", Google put a date on doubling Gemini 4 prices, and the newest top-tier models launched above the price of the ones before them. So which is it?

The short answer

Down for what you did last year. Up, or at least not down, for the best model today. And up for what you spend, because agents use far more tokens. Three different things are happening at once, and mixing them up is how people end up with the wrong story.

  1. The price of a fixed level of capability keeps collapsing. Last year's intelligence is nearly free.
  2. The price of the best model you can buy has stopped falling, and in several places is rising.
  3. The bill per task, per user and per subscription is going up, because agents burn far more tokens than chat ever did.

The starting point: Grok Bot

On 1 November Elon Musk replied to a user asking about Grok Bot's limits: "The cost of serving Grok Bot was unsustainably high relative to revenue." A week later he posted that "Grok Bot fixes for improved token efficiency and smarter behavior are underway", then that a round table with the Grok Bot engineering team would ship fixes overnight. Builders on X said the nerfed limits broke their workflows and people were leaving the ecosystem. TechRadar reported users going from 50% to 100% of their allowance without using it, and calling the change a "rug pull".

The detail that matters: xAI's own API got cheaper. Grok 4.3 launched at $1.25 input and $2.50 output per million tokens, replacing Grok 3 at $3 and $15, an 83% cut on output. Cheap tokens and an unsustainable product at the same time. That is the whole story in one case.

Is Musk betting on tokens getting more expensive? Nothing in his posts says so. His actions point the other way: his latest update says xAI rewrote its entire RL and inference stack from scratch in C/C++ for Grok 4.8, the Grok Bot fix is framed as "token efficiency", and his API prices sit below most rivals. That is a bet on serving tokens more cheaply. The problem is on the demand side: flat-fee users running agents consume more than they pay for.

1. Capability per dollar: a collapse

Epoch AI tracked how fast the price of reaching a fixed benchmark score falls. Across six benchmarks it found declines of 9x to 900x per year, median 50x. Reaching GPT-4's level on PhD-level science questions got about 40x cheaper per year. Epoch's own caveats: the fastest drops started after January 2024, it is unclear they will persist, and it excluded reasoning models.

Output price cuts for the same slot in the lineup. Haiku and Sonnet and Opus are Anthropic list prices.

BenchLM's Token Price Index, which tracks the median blended price (3:1 input to output) of flagship models, sits at 15 against a March 2023 base of 100. Frontier tokens are 85% cheaper than at GPT-4's launch, a median of $5.63 per million.

2. The frontier: the floor stopped falling

The same index has a less comfortable number in it. Year over year, the frontier index is up 114%, while the mid-tier is down 41% and the budget tier down 28%. BenchLM warns it is a median over a changing basket, so a new, pricier model entering can move it without any price rising. That is largely what happened: the increase is a new top tier, not repricing of old models.

BenchLM Token Price Index sub-indices, year over year to November 2026.

What the providers' own pages show:

  • A new tier above the old flagship. Anthropic's Opus sits at $4 to $5 input and $20 to $25 output. Fable launched above it at $10 / $50, twice Opus. Fable 5.2, released 9 November, is $8 / $40: a 20% cut, but still double Opus 5.5.
  • Pro tiers priced like luxury goods. OpenAI's GPT-5.5 Pro is $30 / $180 and o1-pro is $150 / $600, per BenchLM's catalog. GPT-4 launched at $30 / $60.
  • Flagship slot repriced up (aggregator data). OpenAI's GPT-5 was $1.25 / $10. GPT-5.5 is listed at $5 / $30.
  • Introductory prices expiring. Google's pricing page lists Gemini 4 Carbon and Gemini 4 Argon at $2 / $10 through 30 November, then $4 / $20 from 1 December. Gemini 3.8 Flash goes from $0.75 / $3.75 to $1.50 / $7.50 on 1 January 2027.
  • Time-of-day pricing. DeepSeek now charges double at peak hours (01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays), for example $0.30 / $1.20 peak against $0.15 / $0.60 off-peak on its Flash model.
  • Stealth increases. Anthropic's docs say Claude 4.7 and later use a new tokenizer that produces roughly 30% more tokens for the same text. Same sticker price, about 30% higher effective price on the same prompt.
Current output list prices, one model per tier. The dashed line marks what GPT-4 cost at launch.

The counter-evidence is real too: Fable 5.2 and Opus 5.5 both launched cheaper than their predecessors, and Sonnet 5.5 is cheaper than Sonnet 4.6. The pattern is not "everything is rising". It is that the discount curve stopped being automatic at the top, and providers now price new top-tier models at what the market will bear.

3. Why per-token price and per-task cost diverge

OpenRouter's October recap is the cleanest picture of the split. Token volume was 790 trillion in October, 25x January. Its blended effective price per token fell 57 to 70% since January (closed models from $1.21 to $0.52 per million, open-weight from $0.28 to $0.084). Part of that fall was mix shift to cheaper models, not lower list prices.

OpenRouter, January to October 2026. The implied spend bar is our own multiplication of the other two.

The volume came from agents: 76% of tokens came from agentic keys, up from 8% a year earlier. Agents resend big contexts every turn, and OpenRouter says most providers served cached prefixes for 85 to 97% of agentic prompt tokens, which is the only reason the bills are survivable.

That is the Grok Bot mechanism. A flat monthly subscription sells an unbounded quantity (agent loops running overnight) priced against a per-token cost curve. When tokens get 50% cheaper but each user's agent runs 10x more tokens, the margin goes negative. The provider's only levers are to cut limits, raise tiers, or push users to cheaper models. Grok Bot pulled the first one, and TechRadar's read is the likely next move across the industry: tighter limits and higher prices as subsidies end.

Where we are going

This section is my read of the evidence above, not a measured fact.

  1. Intelligence keeps getting cheaper, on a roughly yearly lag. Whatever the best model does today, a budget-tier model does for a fraction of the price within about a year. Opus-class work now costs a quarter of what it did, and Haiku-class work a tenth. Plan on this continuing.
  2. The top of the market becomes a premium tier, not a falling curve. Fable at $8 to $10 in and $40 to $50 out, Pro models at $180 or more, and Google's doubling all point the same way. Expect new top models to launch at or above the old ones, with price moves as steps (end of an introductory price, a new tier) rather than a glide down.
  3. The unit of pricing moves away from the token. Flat plans get capped (Grok Bot, and reportedly OpenAI's shared limits), peak-hour rates appear, and caching discounts become the main lever. Providers are protecting margin from agents, not from chat.
  4. Total spend goes up while the unit price goes down. OpenRouter's numbers imply the average bill rose several-fold since January even as price per token fell more than half.

So the answer to "cheaper or more expensive?" is that the question has to name a model tier and a workload. For anyone building on these models: do not budget on the per-token price, budget on the cost of the finished output. Route by tier (cheap models for extraction and drafting, premium only for the last hard step), exploit caching, and avoid selling unbounded usage at a flat price. Customers care about the post, video or article they get, not the tokens behind it, and the teams that price the output will absorb this volatility better than the teams that price the tokens.

What we could not verify

  • OpenAI's pricing page blocked automated access, so OpenAI prices come from BenchLM's catalog, not OpenAI directly.
  • One aggregator claimed Claude Sonnet 5 rises from $2 / $10 to $3 / $15 on 1 September. Anthropic's pricing page lists Sonnet 5 and 5.5 at $2 / $10, so it is left out.
  • BenchLM's index is a median over a changing basket. The +114% is a composition effect and should not be quoted as "prices doubled".
  • xAI has not published Grok Bot's cost per user or margin. "Unsustainable" is Musk's word, not a number.
  • No source measures average tokens per task over time. OpenRouter shows total tokens and agent share, not per-task cost. The implied-spend figure is our own arithmetic.

Sources

Want to make your own?

Go from a brief to articles, images and videos like these with Genfeed.

Start creating
Newsletter
V

About Vincent

Building genfeed.ai, the open-source content agent your coding agent can run. Host of the @shipshitshow

More from Vincent