Nous Ranks the Models, Then Sells the Router. Who Signs the Savings?

Nous Research now writes the scoreboard and sells the router that reads it. That is a referee who also owns a team. The scoreboard is useful. A savings number the router reports about itself proves nothing until your contract defines how savings get measured, and who signs them.

What Nous published

Nous raised $90 million [Vendor-reported, Nous fundraise note, Oct 2026]. TechCrunch reports a Series B at a $1.5 billion valuation, led by Robot Ventures, bringing total funding to $158 million [Press report, TechCrunch, Oct 7]. The money funds Hermes for Businesses, which promises "One harness across every AI model, with full visibility on cost and ROI." The page shows no price [Vendor page, Nous Portal, Oct 9].

The Hermes Index ranks 14 models by mean score and mean dollars per task across four suites [Vendor-reported, Nous leaderboard, Oct 9]. Five models matter for this read:

All table figures: [Vendor-reported, Nous leaderboard, Oct 9]. Nous marks the Opus and Sonnet average costs provisional, meaning a partial run or a cost estimate pending a rerun.

Read every number with these limits. Results come from the Nous harness only, on these four suites, at high reasoning effort where the model offers it, with one attempt per task. Nous sells the router this index feeds. The tasks are agent, terminal, science and skills benchmarks; none is a GTM workload such as account research or outbound drafting. SkillsBench figures above are the clean scores, which drop tasks where a model found the public answers.

For the list-price view of Sol and Astra, see our GPT-6.1 Sol cost-per-task Order Form. The index tests GPT 6 Sol, an earlier model than GPT-6.1 Sol.

One hard task sets the bill

CTOL Digital ran a simple mix: 99 routine tasks on GPT 6 Luna at its Hermes Bench cost of $0.015 each, plus one science task on Sonnet 5.5 at $8.02 [Press report, CTOL Digital, Oct 7; costs Vendor-reported, Nous leaderboard, Oct 9]. The routine work costs $1.485, the bill is $9.505, and the science task is 84% of it [RAR calculation from Nous leaderboard, Oct 9]. The arithmetic holds.

Two checks. Price Luna at its index average of $0.141 and the science share falls to 37% [RAR calculation from Nous leaderboard, Oct 9]. Send the science task to Opus at $12.65 and it rises to 89% [RAR calculation from Nous leaderboard, Oct 9]. The router's choice on the hard step decides most of the spend. Audit that step.

The Proof Gap

What Nous claims: the router will "cut your spend by letting you choose the best models at the best prices," with full visibility on cost and ROI [Vendor-reported, Nous fundraise note and business page].

What a buyer can verify today: public benchmark scores from the vendor's own harness. The price, savings method, baseline and definition of success in your workflow are unpublished.

Order Form lines

Add these before signing:

• Cost per accepted outcome, per model, per step, monthly. You define "accepted": meeting held, ticket closed without reopen.
• Router decision log export. Every call: model chosen, reason, tokens, dollars, outcome. Raw rows, delivered to your warehouse on your schedule.
• Pre-router baseline. Measure cost per accepted outcome on your own workflows for a fixed window before routing starts. Freeze it in the contract. Savings are computed against it, and only it.
• Right to pin and block models per workflow or step, at no fee.
• Your eval set. A sample of your accepted outcomes runs on each candidate model before any routing change, and again after any model swap.
• Separate price lines. Router fee on one line, model pass-through on another, any markup shown.
• Named savings owner. One person in RevOps signs the savings figure each quarter; Finance co-signs. The vendor reports. Your owner certifies.

Eight-Seat Read

• CRO: Ask for cost per accepted outcome by workflow. Ignore blended savings.
• CMO: Content and research agents hit hard steps often. Expect those steps to dominate spend.
• RevOps: You own the baseline, the log export and the savings sign-off.
• Sales: Watch for quiet downgrades on research steps that feed your calls.
• CS: Define "accepted" with reopen and escalation rates, then hold the router to it.
• Marketing Ops: Pin models on brand-sensitive steps until your eval set clears them.
• Finance/CFO: Co-sign savings only against the frozen baseline, with the fee on its own line.
• IT/Security: Confirm where logs live, who can read prompts, and how the self-hosted option ships.

Why we might be wrong

Hermes Agent is MIT licensed. A team that self-hosts can inspect routing directly, which narrows the Proof Gap. Most GTM work is routine, so hard steps may be rare and cheap in your mix. The provisional Opus and Sonnet costs may move after the rerun. Nous also publishes its method and flags contaminated runs.

Sources

1. Nous Research, Hermes Index leaderboard (read Oct 9, 2026): portal.nousresearch.com/bench
2. Nous Research, "A Note on Our Fundraise," Dillon Rolnick, Oct 2026: nousresearch.com
3. Nous Portal, Hermes for Businesses (read Oct 9, 2026; no price shown): portal.nousresearch.com/business
4. Marina Temkin, TechCrunch, Oct 7, 2026: techcrunch.com
5. CTOL Digital, Oct 7, 2026: ctol.digital

Get the weekly read on where AI is moving revenue. Subscribe to the Report.