Multi-agent swarms hit a cost wall, and token-cutting middleware forms
Gateways and audits that cut agent token spend, paid for by enterprises whose agent bills outrun the value delivered.
Crowded · Vals AI finds teams cost 5.1x more
Written for: Investor or analyst
Next step
Thesis: agent spend shifts from swarm orchestration toward single agents plus token-cutting middleware, which pressures seat-based agent pricing. Confirming signal: an agent platform vendor reprices around cost per completed task, or cites swarm cost in guidance. Already building: Vals AI on measurement, mcpdelta and local-indexing projects, and Anthropic's internal scaling research.
Why now
New since the last card: third-party data and a lab's own data now agree that 'more agents' is not a quality lever. Builders are shipping fixes. mcpdelta claims up to 24.1x fewer tokens by running one program per task instead of a tool call per step, and local indexing projects cut tool calls. Snapshot-and-branch recovery beats retrying coding agents from scratch. Swarm-first products such as Bops face harder ROI questions.
What became possible
Independent testing from Vals AI shows agent teams cost up to 5.1x more than solo agents. Only one of four tests with GPT-6 Sol and Claude Opus 5.5 showed a measurable quality gain. Anthropic's own data shows quality plateaus beyond ten agents.
Two ways in
- MCP program gateway
Compiles multi-step tool calls into single programs, records every change and can undo it.
- Customer
- Enterprise platform teams running MCP-heavy agents
- Model
- Per-seat or share of token savings
- Agent architecture audit
Compares solo-agent and swarm setups on cost per completed task before a team scales to production.
- Customer
- Heads of AI at mid-size enterprises
- Model
- Fixed-fee engagement turning into ongoing monitoring
Competition
The ledger side's main opening is tokenizer-normalised accounting across vendors. Revenium, Paid, CloudZero and Vantage all report cost per outcome or per unit, but none of them says how it normalises token counts across different tokenizers.
Full competition notes (12 competitors)
The ledger side's main opening is tokenizer-normalised accounting across vendors. Revenium, Paid, CloudZero and Vantage all report cost per outcome or per unit, but none of them says how it normalises token counts across different tokenizers. The Haiku 5.5 launch, which makes the same text use roughly 30% more tokens, makes this concrete, and today the explanations come only from blogs such as Yotta Labs and TensorZero. A neutral ledger that neither routes nor bills could sell as an audit-grade record of cost per finished task across Anthropic, OpenAI and open-weight models, and it could include budgeting against the price tiers above 100K tokens. The risk is that this is a feature CloudZero or Revenium can add quickly. The router idea has little room left: Not Diamond already trains routers on customer-defined evals, and OpenRouter, Martian, Portkey, LiteLLM, TensorZero and Braintrust cover multi-vendor routing. Routing per token across vendors is unclaimed, but different tokenizers and KV caches make it barely workable, so it is a weak basis for a company. A newcomer probably does better building the measurement layer and feeding its data to existing routers than building another gateway.
- Revenium AI Outcomes
Ties all AI operations in a workflow to a single outcome identifier and reports cost per conversion and ROI per workflow. This is nearly the task-level ledger idea.
- Paid.ai
Raised $33M. Tracks spend across multiple LLM vendors and reports cost and margin per customer and per agent action. Aimed at companies selling agents, with outcome-based billing.
- CloudZero
FinOps platform that pulls LiteLLM usage across OpenAI, Anthropic and open models into unit-cost metrics per feature, transaction or provider. Publishes guidance on per-task cost rollups for agents.
- Vantage
Connects directly to OpenAI and Anthropic billing APIs and provides per-request costs and model-level attribution next to cloud spend.
- Finout
Enterprise FinOps vendor with AI cost tooling across Vertex AI, OpenAI and Anthropic. Positioned for company-wide cost allocation.
- Gravitee
API and LLM gateway vendor promoting cost per successful task, wasted tokens and runtime budgets. Correlates OpenTelemetry spans with finance tags.
- Not Diamond
Model router that trains custom routers on customer evaluation data, with explicit cost, quality and latency tradeoffs. Supports custom models. Directly overlaps the policy-driven router.
- Martian
Commercial model router and gateway with provider metrics and integrations for coding agents.
- OpenRouter
Unified API over 400+ models from 70+ providers with routing and fallbacks. Passes provider prices through and charges a platform fee on credit purchases.
- TensorZero
Open-source LLM gateway with routing across providers. Has published analysis showing that tokenizer differences make per-million-token price comparisons misleading.
- Braintrust
Eval and observability platform with a gateway. Argues that routing should be tied to measured quality from live traffic.
- Respan
Vendor offering quality-based routing driven by online evals and LLM-as-judge scores from production traffic.
Checked 2026-10-09
Risks
- Crowded, with many open-source token savers
- Labs and agent platforms will add these optimisations natively
- Falling token prices reduce the urgency
- Token-saving benchmarks are self-reported
Best on tech readiness and team fit; open space is the weak point (4).
tech 7 · pull 6 · space 4 · fit 7 · build 7
Exposed
Companies that could benefit or be threatened. News and context, not investment advice.
Momentum of its main topic
Sources
- AI agent teams waste massive tokens for barely measurable quality gains, research findsThe Decoder
- mcpdelta/mcpdeltaGitHub
- Local Indexing to reduce tool calling and token usageHacker News
- OrgoAI/bopsGitHub
- Returning to Development States: When Full Sandbox Snapshots Help Coding AgentsarXiv
- AI agents overstate their results and remain far from autonomous research, study findsThe Decoder
What happened next
- +7d · 16 Oct 2026 nothing yet. No product launches, funding announcements, or major press coverage of token-cutting middleware, MCP gateways, or agent audit platforms were found after October 9, 2026. While the category is actively discussed in vendor blogs and industry commentary, no specific new entrant has entered the market in the week following the flag date.
Status history
- 2026-10-09 Flagged as buildable