New guide: AI agent cost tracking in 2026, compared.

Read it
RAWRabbitt

RAW · THE AGENT OPTIMISATION PLATFORM

Optimise your AI agents for accuracy and cost.

Build, observe, optimise, attribute, and secure every agent in one platform. RAW routes, caches, compresses, and governs each request, then shows you what it saved.

Read the cost guide

For platform, engineering, security, and finance teams.

RabbittSupport triage agentIllustrative
Run #a91f · trace6 spans · 1.1 s · $0.004
  1. TriggerTicket #4821—
  2. Policy checkPII maskedGuardrail—
  3. Retrieve contextHelp-centre articlesCache hit$0.0002
  4. PlanSmall modelRouted$0.0009
  5. Draft reply2,410 → 1,380 tokensCompressed$0.0027
  6. Update caseSalesforce$0.0002
Illustrative run · sample values

YOUR MODELS. YOUR STACK.

Works with the models and apps you already run.

  • Slack
  • Notion
  • GitHub
  • Salesforce
  • OpenAI
  • Anthropic
  • Google Gemini
  • Amazon Bedrock
  • Ollama

Agents are in production. Most teams cannot see, tune, or explain what they cost.

  1. 01

    Black-box runs

    A run fans out into models, tools, and subagents. When it is wrong or slow, nobody can say which step caused it.

  2. 02

    Spend without an owner

    The invoice shows totals per API key. It cannot say which agent, team, or customer spent the money.

  3. 03

    Accuracy left to chance

    Every agent uses the biggest model and the longest prompt, because no one can prove a cheaper path keeps quality.

Ask about any run. RAW Assistant answers from your traces, spend, and policies.

Which agent drove last week's spend?

Illustrative prompt

Work with your stack. Connect agents to the apps and models your teams already use.

  • Slack
  • Notion
  • GitHubGitHub
  • Salesforce
  • OpenAIOpenAI
  • AnthropicAnthropic
  • Google Gemini
  • Amazon Bedrock
  • OllamaOllama

Illustrative app and model connections

FIVE MODULES. ONE TRACE.

Everything a production agent needs, around one core.

Each module replaces a tool you would otherwise stitch in. Together they share one trace, so a change in Optimise shows up in Observe, Attribute, and Secure.

  1. Build & Run

    Agents and workflows on a production harness.

    Like Glean, n8n

  2. Observe & Trace

    Every request, span, token, and latency in one trace.

    Like Langfuse, Grafana, Datadog
  3. CORE

    Optimise

    Routing, caching, compression, guardrails, and spend controls that make agents more accurate and cheaper.

    Like LiteLLM, GPTCache, LLMLingua, Headroom

  4. Attribute

    Spend by agent, team, customer, and run.

    Like CloudZero, Vantage

  5. Secure

    AI red teaming and controls your CISO can sign off.

    Like CISO tooling, AI red teaming

OPTIMISE · THE CORE

More accurate agents.
Lower cost per run.

Optimise sits in the request path. Every call is checked against policy, trimmed, served from cache when it can be, and routed to the model that meets your quality bar for the lowest price. Each change is measured against your evals.

One layer for the jobs of LiteLLM, GPTCache, LLMLingua, Headroom, and your own harness code.

The cheapest model that meets your quality bar.

Each request is scored by task and difficulty, then sent to the model that clears your eval threshold for the lowest price. Hard requests still reach the frontier model.

Frontier model for every call$0.019 / run
Routed per request$0.007 / run
Classification
small model
Draft reply
mid model
Escalation reasoning
frontier model
Illustrative values

OPTIMISE · ROUTING & GATEWAY

One gateway in front of every model.

Apps and agents call one endpoint. RAW chooses the model, applies policy, and records the request in the trace.

  • Bring your own providers and endpoints
  • Route per request on cost, latency, and quality
  • Fallbacks when a provider fails

The job of LiteLLM, OpenRouter, inside one trace.

YOUR APPS & AGENTS

RAW AI Gateway

Model access + request evidence + usage

BRING YOUR OWN MODEL
OpenAI
Anthropic
Google Gemini
Amazon Bedrock
Ollama
Your endpoint
Illustrative model choices · confirm your provider and endpoint requirements during evaluation.

OPTIMISE · EVALUATION

Prove the cheaper path is just as good.

Every optimisation runs against your eval set before it takes traffic, so cost goes down and accuracy does not.

  • Compare models, prompts, and workflow versions
  • Quality, latency, cost, and errors side by side
  • Ship the change only when the eval holds

The job of custom eval scripts, inside one trace.

Compare two approaches
Illustrative comparison of two example agent configurations. Quality scores are example human review ratings, not product results.
Sample metricRun ARun B
Review rating4 / 54 / 5
Latency3.2 s2.8 s
Cost$0.24$0.19
Tool errors00
Compare → review → adjust → evaluate again
Illustrative values · review rating uses an example human evaluation.

BUILD & RUN

Build agents and workflows on a production harness.

Turn a business task into an agent, then into a repeatable workflow. Every run is traced and optimised from day one.

  • Instructions, models, tools, and context in one agent
  • Workflows with triggers, review steps, and results
  • Runs on the same harness Optimise improves

The job of Glean, n8n, inside one trace.

Support-request triageIllustrative workflow
Agent · Triage assistant

Reason through the next step

Combine instructions, a selected model, and available tools to work through the request.

Select a step. App connections are illustrative examples.

OBSERVE & TRACE

Every model call has a story. See all of it.

When an agent is slow, wrong, or expensive, start with the trace. Follow the run from trigger to result, step by step.

  • Requests, spans, tool calls, and status
  • Tokens, latency, and cost on every step
  • Each call tied to its agent and run

The job of Langfuse, Grafana, Datadog, inside one trace.

AI request logs
Sample data

Select a request to inspect its metadata.

req_04dfSuccess
Latency
2.8s
Tokens
1,240
Request cost
$0.18
Message content masked

Content access and retention follow your company requirements.

Illustrative requests and costs. Select a row to explore.

ATTRIBUTE

AI spend, explained. Down to the agent run.

Put every dollar on whoever spent it, minutes after the run, not days after the invoice.

  • Cost by agent, team, customer, and step
  • Cached and reasoning tokens priced correctly
  • Export to the FinOps tools finance already uses

The job of CloudZero, Vantage, inside one trace.

SPEND · LAST 7 DAYS$2,184−24% vs last week
AgentRunsCostShareTrend
Support triage6,120$1,04248%−31%
Document review2,480$61228%−12%
Sales research1,310$34816%+4%
Internal assistant3,050$1828%−22%
Export toCloudZeroSnowflakeBigQuery
Illustrative spend · sample customers and teams

SECURE

Red team your agents before attackers do.

Probe every agent with attacks built for its tools and data, turn findings into guardrails in Optimise, and keep the evidence your security team needs to sign off.

  • Automated red teaming: injection, jailbreak, data leaks
  • Runtime guardrails for PII, secrets, and tool misuse
  • Access, retention, and audit evidence for the CISO

The job of CISO tooling, AI red teaming, inside one trace.

Red-team run · Support triageIllustrative
148probes run
139blocked
9to review
  • 91Indirect prompt injection via ticket bodyInjectionBlocked
  • 88Exfiltrate customer email through tool callData leakBlocked
  • 74Role-play jailbreak to reveal system promptJailbreakFlagged
  • 63Unapproved model requested by agentPolicyBlocked
  • 22Refund issued above approval limitTool misusePassed
Illustrative findings · score is example risk severity

HOW IT WORKS

From first trace to cheaper, more accurate agents.

No rewrite. Point traffic at RAW, let the trace show where the cost and errors are, switch on the levers that fix them, and prove the result.

  1. # one line in your agent
client = OpenAI(
  base_url=RAW_GATEWAY
)

    Framework unchanged

    01

    Connect

    Point model traffic at the RAW gateway or add the SDK.

    Run #a91f · 4 spans · $0.019

    02

    Trace

    Every request, span, token, and cost lands in one trace per run.

    • Model routing−$0.009
    • Semantic cache−$0.004
    • Prompt compression−$0.002

    Eval score held at 0.94

    03 · CORE

    Optimise

    Route, cache, compress, and enforce policy, checked against evals.

    • Support
    • Sales ops
    • Finance

    −79% spend by team, audit attached

    04

    Prove it

    Attribute the savings by agent and team, with security evidence.

ACROSS THE WHOLE AI STACK

One policy and one trace, wherever agents run.

Homegrown, bought, or on a developer's laptop, every agent goes through the same Optimise layer, with no custom integration per tool.

  • Homegrown agents

    Agents your teams build in code, on any framework.

  • Agentic platforms

    Agents configured outside your codebase, still under one policy.

  • Business apps

    AI inside the SaaS tools your company already runs.

  • Coding agents

    Spend and policy for the agents on every developer's machine.

INSIDE RAW

The product, as
your teams see it.

Real screens from the RAW workspace. Pick a team to see the view they work in every day.

Every model, key, and provider in one view.

See which models and virtual keys carry your traffic, and how spend splits across providers.

  • Top models and virtual keys
  • Spend, success, and failures by provider
  • Table and chart views
RAW usage view with top virtual keys, top model names, and spend by provider
Product screenshot · personal details redacted

INTEGRATIONS

Wherever your agents run, RAW connects.

Keep your models, frameworks, and tools. RAW sits in the request path and the trace, not in your way.

  • Model providers 9
  • Agent frameworks 10
  • Coding agents 5
  • Browser agents 5
  • Apps & data 8
  • Telemetry & FinOps 7
Illustrative list · confirm the integrations you need with our team.

FROM THE BLOG

Guides for running agents in production.

All articles
RAW usage dashboard with total spend, request metrics, and a daily spend chart
Attribute

How to track AI agent costs: gateways, tracing, and FinOps tools

RAW Team · October 8, 2026

Your provider invoice shows spend per API key, not per agent, customer, or step. A practical comparison of the three tool layers that fix that, and how to pick between them.

Read more
RAW guardrails table listing content filters such as PII masking and prompt injection shield
Optimise

How to cut LLM agent costs without hurting accuracy

RAW Team · October 1, 2026

Model routing, semantic caching, prompt compression, and better prompts: what each one saves, where it goes wrong, and how to test it before it reaches users.

Read more
RAW usage view with top virtual keys, top model names, and spend by provider
Platform

The five jobs of an AI agent platform

RAW Team · October 5, 2026

Build, observe, optimise, attribute, secure. Why teams end up with five separate tools for one agent, and what changes when they share a single trace.

Read more

Enterprise checklist

  • In the request path
  • Eval-gated changes
  • SSO and roles
  • Your models, your keys
  • Split-plane deployment
  • Horizontal by design
  • Every decision on record
Confirm deployment and performance requirements with our team.

Built for enterprise AI at scale.

Performance

In the request path

Routing, caching, and policy run inline without slowing the agent down.

Accuracy

Eval-gated changes

No optimisation takes traffic until it holds against your eval set.

Access control

SSO and roles

Decide who can change routing, policies, and budgets, team by team.

Model choice

Your models, your keys

Bring your own providers, endpoints, and credentials. Switch without a rewrite.

Data sovereignty

Split-plane deployment

Keep the data plane in your cloud or on-prem, so prompts stay where your policies require.

Scale

Horizontal by design

Add gateway nodes as traffic grows, across cloud, hybrid, and on-prem.

Audit trail

Every decision on record

Each request, policy decision, and change is logged with the run it belongs to.

Enterprise ready07 capabilities

QUESTIONS

Questions for your platform evaluation.

SEE IT ON YOUR AGENTS

Find out what your agents could save, without giving up accuracy.

Bring one agent. We will trace it, optimise it, and show you the cost and quality difference.

Read the blog