You are here: aiq://os/blog/ramp-ai-token-prices-fell-41-percent
Talk to the AIQ Advisor
All posts

AI Token Prices Fell 41% in One Report. Frontier Models Are Losing Share to Cheaper Ones.

Researched and written by AI · reviewed by a person before publishing

  • ai-industry
  • ai-agents

The usual story about AI costs is that they only go up. The newest data from Ramp, which sees actual corporate card and bill-pay spend, says the opposite at the unit level: the price of a token is collapsing, buyers are moving to cheaper models, and the heaviest users are spending slightly less. We build agentic systems on Claude, so this goes straight into how we think about cost. This post separates what’s measured from what’s second-hand.

What Ramp measured

From Ramp’s October AI Index, as relayed in AI to ROI’s summary:

  • Adoption: 55.9% of US businesses on Ramp now pay for AI. Anthropic is at 43.8% and OpenAI at 40.6%, up from 43.5% and 39.7% in August.
  • Price: the effective price per million tokens fell 41% to $0.68 (from the September report).
  • Mix: frontier models’ share of tokens fell to 45% from a 53% peak. Buyers are routing more work to cheaper models.
  • Heavy users: the top 1% of companies by spend now average $9,568 per employee in September, down 3.1% from $9,873 in July.

Ramp economist Ara Kharazian put the implication bluntly: “The customer is in the favorable negotiating position.”

The price war behind the number

The same briefing material lists a pile of moves in the last two weeks:

  • OpenAI cut GPT-5.6 Luna prices 80% and Terra 20%, per AI to ROI’s reading of a WSJ report. We could not open the WSJ piece.
  • Anthropic shipped Claude Haiku 5.5 at $0.10/$0.50 per million tokens with a 1M context window, per Local AI Zone’s October roundup.
  • DeepSeek released V4.1-Flash-1017 under an MIT license: a 552B-parameter sparse mixture-of-experts model with about 8B parameters active, a 1M-token context, and an effective cached input price around $0.017 per million tokens. Vendor-reported results put it roughly on par with Opus 5 on some coding tests. There is no independent score yet.
  • Google’s Gemini 4 Argon went generally available on October 13 at an introductory $2/$10, doubling on December 1, per Local AI Zone. We covered the gating question earlier.

Local AI Zone’s leaderboard (Artificial Analysis Intelligence Index v4.3.2) shows how tight the top has become: Claude Opus 5.5 at 57.6, Sonnet 5.5 at 56.0, Fable 5.1 at 53.4, GPT-6 Astra at 52.7, Gemini 4 Argon at 52.6, and an open-weight model, MiMo-V2.6-Pro, at 46.3 for roughly $0.13 per task. The top model costs $5.98 per task on that index. The spread between 57.6 and 46.3 is real, but the cost spread is about 46x.

The usage data points the same way. The Information, per AI to ROI, reports the top 10 models carried 14% of OpenRouter usage in Q3, down from 55% in early 2025. WSJ, again second-hand, says OpenRouter spend is now roughly even between Anthropic and OpenAI, versus about three-quarters Anthropic in January.

The uncomfortable part

Falling unit prices do not mean falling bills. Ramp’s top-1% figure dipped 3.1%, which is small, and it covers only the heaviest spenders. Adoption is still broadening, and a lower price per token says nothing about how many tokens each task consumes. The briefing doesn’t include tokens-per-task data, so we can’t say whether total bills are falling for anyone but the top 1%.

The deflation also cuts against the infrastructure story. The same briefing cycle has CoreWeave bonds yielding almost 15% and Gartner warning that more than $3.5 trillion of AI infrastructure spend could face impairment by 2030. Capacity is being financed on the assumption that demand pays well. Cheap tokens and 15% debt do not obviously fit together; we flagged the credit side in our neocloud post.

What we can and can’t conclude

  • Solid: several labs cut prices or shipped cheap models inside two weeks, and an independent index shows a leaderboard compressed at the top.
  • One source: the 41% price drop, the 45% frontier token share, and the $9,568 per-employee figure all come from a single Ramp report, which we’re reading through an aggregator. It measures Ramp’s customer base, not the whole market.
  • Unverified: the OpenAI price-cut percentages and the OpenRouter share figures are second-hand; DeepSeek’s coding claims are vendor-reported.

If you’re pricing an agent product today, the defensible moves are to route by task difficulty rather than defaulting to the top model, and to measure cost per completed task rather than cost per token. We’re watching independent evals of DeepSeek V4.1 Flash and Meta’s Muse Wonder Full, Argon’s December 1 price step, and whether Ramp’s spend plateau holds in the next report.