Open-weight models processed 71.5% 09 of the tokens on Vercel's AI Gateway in September 2026. They took 23% 10 of the dollars. Vercel published both figures in its October 2026 Production Index. Most of the work already runs on cheap models, and most of the money still flows to a few expensive ones.
That gap is a price signal you can act on. Divide 71.5 by 23 25 and open weights carry about 3.1 times their spend share in tokens. In Vercel's numbers, closed models run the other way: 28.5% 11 of tokens for 77% of spend.
These figures cover one gateway, and it skews toward developers and web apps. Enterprise Technology Research's September survey puts open weights at 34% 12 of enterprise token usage, well below Vercel's figure. So the honest claim is narrower. On the traffic that is easiest to move, open models already own the volume.
The supply side keeps adding pressure. Reflection AI just announced Beam, a 501-billion-parameter 04 mixture-of-experts model, and released its model weights under the Apache 2.0 license. Mixture-of-experts means only a slice of the model fires on each token, which keeps serving costs down. Apache 2.0 on the released weights means you can run them and sell products built on them.
The volume layer and the premium layer
Vercel's data splits the market in two. In the volume layer, tokens pile up at low prices. In the premium layer, a smaller stream of tokens collects most of the money.
What open and frontier tokens cost, and what switching could save
Some napkin math shows how wide the spread is. Open weights earn 3.1 27 points of token share per point of spend share. Closed models earn 0.37 26, which is 28.5 divided by 77. Divide one by the other and the average open token on Vercel's gateway costs about 8.4 times less than a closed one.
The September model names sort cleanly into those two layers. On the volume side, DeepSeek handled 58% 23 of gateway tokens and 6.6% 24 of spend, per Vercel. On the premium side, Vercel shows Anthropic's Claude Opus 5.5 taking 11.03% 19 of spend and OpenAI's GPT-6 Astra taking 4.86% 20.
The interesting case sits in between. Moonshot AI's Kimi handled 3.6% of tokens and 13.2% of spend, per Vercel, which means buyers are paying premium-layer prices for an open model. On Terminal-Bench 2.1, Kimi K3 scores 88.3 01 as reported by Moonshot AI, ahead of GLM-5.2 at 81.0 as reported by Z.AI and Beam at 80.1 as reported by Reflection AI. Nobody can say for sure whether Vercel's Kimi spend is K3 specifically. Still, the pattern fits: money moves toward measured quality, whatever the license.
Where the 77% of closed spend goes
Start with the before and after, because the dates matter. In December 2025, Vercel's data put open weights at fewer than one in ten gateway tokens. Vercel's August 2026 index showed 56% of tokens and 14% 13 of spend. By September the figures had climbed to 71.5% 07 of tokens and 23% of spend.
Over that stretch, Vercel says September spend ran about 13 times December's while the average price per token fell 47% 14. Your own bill probably looks similar. Ramp's AI Index found the average price US businesses paid per million tokens fell 41% 08 from its March peak to $0.68, Wired Enterprise reports. Ramp also found frontier models' share of business tokens dropped from 53% in August to 45% as companies set cheaper defaults.
Now put that on a monthly invoice. Computer Weekly, citing McKinsey and Silicon Data, puts open-weight API prices at roughly $1 21 to $6 per million output tokens and frontier APIs at $10 22 to $50. A triage bot writing 100 million 28 output tokens a month costs $100 to $600 on open weights and $1,000 to $5,000 on frontier models.
The headline price leaves out one cost. That Computer Weekly report says agentic systems can burn 5 to 30 times more tokens per task. A cheap model that loops three extra times on a hard ticket can end up costing more than an expensive model that finishes in one pass. So stop counting cost per token and count cost per successful task.
Cost per successful task is your total spend divided by the tasks that passed. That spend has to include retries and human review. Think of a print shop. The open model runs the presses all day on standard jobs. The closed model is the senior proofreader, paid by the hour, who only sees the pages that failed a check.
Three numbers will get more attention this month than they deserve. The first is the 71.5% 07 headline read as a market-wide figure, when it describes one gateway. The second is launch buzz: Vercel says TypeSafe AI's Jev reached 35% of paid teams in two weeks yet took 0.27% of spend. The third is a small leaderboard gap. Beam trails GLM-5.2 by 0.9 30 points on Terminal-Bench v2.1, and your own tests may never notice it.
Two things deserve more attention: license and price. Beam's Apache 2.0 license lets you self-host and fine-tune it with no usage fee owed to Reflection AI. Other open models carry their own terms, so read each license before you ship on it.
On price, MIT Sloan reports a paper by Frank Nagle of the MIT Initiative on the Digital Economy and Daniel Yue of Georgia Tech. They found open models reach about 90% of closed-model performance at release and cost 87% less to run. By their estimate, shifting demand from closed to open models could cut average spending by more than 70% 15, about $25 billion a year globally.
We think open weights should be the default for the bulk path. Pay the premium only where your own tests show a gap that costs you money.
What actually moves the AI bill
Buyers pay premium prices for an open model.
Moonshot AI's Kimi handled 3.6% of gateway tokens and took 13.2% of spend, per Vercel. Kimi K3 scores 88.3 01 on Terminal-Bench 2.1, ahead of GLM-5.2 and Beam, which suggests money follows measured quality whatever the license.
Cheap tokens can still make an expensive task.
Computer Weekly reports agentic systems can burn 5 to 30 times more tokens per task. A cheap model that loops three extra times on a hard ticket can cost more than a closed model that finishes in one pass, so count cost per successful task.
Switching an open model in or out keeps getting cheaper.
Vercel says nearly 70% of gateway spend on Chinese models went through US-based providers, up from less than 10% in December. Almost one in three teams changed their most-used lab in September, an index record.
Spend share follows tokens into 2027
The dated points show the trend. Vercel puts open-weight spend share near 5% in June and 14% 13 in August. By September it had reached 23% 10. That works out to about 18 31 points in three months, or 6 points a month. Token share has less room to run, since 71.5% 07 leaves only 28.5 26 points on the table.
If spend share kept adding 6 points a month, it would cross 50% 32 in about four and a half months, around February 2027. We doubt the line stays that straight. Enterprises move slower: Enterprise Technology Research measured a climb from 23% 16 to 34% of enterprise tokens in a year, and its respondents expect 41% 17 within twelve months.
Switching costs are falling too, and that compounds the shift. Vercel says almost one in three teams changed their most-used lab in September, an index record. Vercel also reports that nearly 70% 18 of gateway spend on Chinese models went through US-based providers, up from less than 10% in December. Both trends make it cheaper to try an open model on Monday and drop it on Friday.
The risk runs one way. Build a routing layer now and your downside is a few days of test work, easy to unwind if closed prices fall. Lock every call into one closed vendor and you pay full premium on the share of work the gateway market has already moved off closed models. Ramp found only 6.4% of US businesses paying for AI pay for open-source models, against 43.8% paying Anthropic and 39.8% paying OpenAI.
So most buyers have not switched yet. That gives the teams that do a counterpositioning opening: their unit costs fall while competitors keep paying list price.
Price one workload per successful task
Start with the workload that has the most calls and the cheapest mistakes, such as ticket tagging or field extraction. Volume is where the 8.4 27 times price spread pays off. A low cost of error keeps the test safe.
Step one: pull 200 real requests from your logs, along with the answers a person accepted. That set becomes your eval, a fixed test you rerun every time something changes. Keep it frozen so results stay comparable month to month.
Step two: write the pass rule before you run anything. Tags need an exact match. Extraction needs every required field present. Any output that breaks the rule counts as a failed task.
Step three: run all 200 through your current closed model and two open models, such as GLM-5.2 and Beam through a hosted provider. Log tokens and retries for every call. Expect some runs to time out or return junk, and count those as failures.
Step four: compute cost per successful task for each model. Add token spend to human review minutes priced at your hourly rate, then divide by passes. Suppose the open model passes 180 33 of 200 and the closed model passes 192. You now know exactly what those 12 extra passes cost you.
Step five: route. Send every request to the cheapest model that cleared your bar, and escalate failed checks to the closed model. The open model will break on some class of input, like long documents or odd formats. Write each break down, because those breaks are your routing rules.
Step six: rerun the 200 on the first of every month. Vercel publishes a fresh index monthly, and its average price per token fell 47% 14 between December and September. Your routing table should move as fast as the prices do.
Price one high-volume workload per successful task.
- Freeze a 200-request 33 eval. Pull 200 real requests from your logs for a high-volume, low-risk job like ticket tagging, along with the answers a person accepted. Keep the set fixed so monthly results stay comparable.
- Write the pass rule, then run three models. Define pass before you start, such as an exact tag match or every required field present. Run your current closed model plus GLM-5.2 and Beam through a hosted provider, and log tokens, retries and timeouts as failures.
- Compute cost per pass and route. Add token spend to human review minutes at your hourly rate, then divide by passes. Send traffic to the cheapest model that clears your bar, escalate failed checks to the closed model, and rerun on the first of every month.
Make open weights the default and make closed models earn their premium.
On Vercel's gateway, open weights carry about 3.1 25 times their spend share in tokens, and the average open token costs about 8.4 27 times less than a closed one. Ramp finds only 6.4% of US businesses paying for open-source models, so most buyers are still paying list price. A frozen eval, a written pass rule and a routing table cost a few days of work and are easy to unwind. Measure cost per successful task, route by the results, and rerun every month as prices fall.