Workloads

What real AI workloads cost on Anvat.

10 common workload shapes — agentic coding, RAG, support triage, batch classification, deep research — with the cost math fully visible. No invented customers, no cherry-picked numbers. Plug your own token estimates into /cost-calculator to reproduce.

Method: we apply the publisher's list price per million tokens (input + output, exact tier math) to a realistic monthly volume, then show the same workload at Anvat's 30%-off metered rate, then the effective cost after the 2× prepaid credit match. The middle column is what you owe before any credit pack; the right column is the effective burn rate against your wallet.

Coding agents

Daily Claude Code sessions, 1 engineer

Claude Sonnet 4.6

One engineer running claude-code interactively most of an 8-hour day. Mostly Sonnet 4.6 with occasional Opus 4.8 escalation for hard refactors. Heavy prompt caching keeps reads cheap.

Input / mo

24M tok

@ $3/M

Output / mo

6M tok

@ $15/M

Total

30M tok

Model

Claude Sonnet 4.6

Direct (list price)

$162

upstream provider / mo

Anvat metered (30% off)

$113

pay-as-you-go / mo

Anvat + 2× credit match

$56.7

effective burn rate / mo

Annual savings vs direct — $1,264 (65% lower)

Note — Assumes 60-70% cache hit rate on common system prompts. Real-world spend will be 30-50% lower than the blended number once cache reads are weighted in.

Claude Code across a 10-engineer team

Claude Sonnet 4.6

10 engineers, daily Claude Code sessions, weekday-only. Same usage shape as the single-engineer case scaled 10×.

Input / mo

240M tok

@ $3/M

Output / mo

60M tok

@ $15/M

Total

300M tok

Model

Claude Sonnet 4.6

Direct (list price)

$1,620

upstream provider / mo

Anvat metered (30% off)

$1,134

pay-as-you-go / mo

Anvat + 2× credit match

$567

effective burn rate / mo

Annual savings vs direct — $12,636 (65% lower)

Note — Caching effects amplify at team scale because the same project context is shared across many sessions.

Autonomous PR code-review agent

Claude Opus 4.8

Agent picks up every PR opened in a 50-engineer org, reviews diffs end-to-end against repo conventions. Opus 4.8 for the actual review pass, Haiku 4.5 for triage.

Input / mo

80M tok

@ $15/M

Output / mo

18M tok

@ $75/M

Total

98M tok

Model

Claude Opus 4.8

Direct (list price)

$2,550

upstream provider / mo

Anvat metered (30% off)

$1,785

pay-as-you-go / mo

Anvat + 2× credit match

$893

effective burn rate / mo

Annual savings vs direct — $19,890 (65% lower)

RAG / knowledge

RAG over 100K internal docs

Claude Sonnet 4.6

Internal knowledge agent. Retrieval narrows to 8K tokens of context per question. ~200 questions / day across the org. Sonnet 4.6 is the right value tier.

Input / mo

50M tok

@ $3/M

Output / mo

8M tok

@ $15/M

Total

58M tok

Model

Claude Sonnet 4.6

Direct (list price)

$270

upstream provider / mo

Anvat metered (30% off)

$189

pay-as-you-go / mo

Anvat + 2× credit match

$94.5

effective burn rate / mo

Annual savings vs direct — $2,106 (65% lower)

Customer support

Customer-support email triage

Claude Haiku 4.5

Inbound support emails classified and pre-drafted before a human reviews. High volume, short turns. Haiku 4.5 with prompt caching keeps cost trivial.

Input / mo

60M tok

@ $0.8/M

Output / mo

18M tok

@ $4/M

Total

78M tok

Model

Claude Haiku 4.5

Direct (list price)

$120

upstream provider / mo

Anvat metered (30% off)

$84.0

pay-as-you-go / mo

Anvat + 2× credit match

$42.0

effective burn rate / mo

Annual savings vs direct — $936 (65% lower)

GPT-5 powered customer chat

GPT-5

Branded customer-facing chat at ~5K conversations / day, average 10 turns each. GPT-5 because the brand voice is already tuned for the OpenAI Responses shape.

Input / mo

180M tok

@ $1.25/M

Output / mo

50M tok

@ $10/M

Total

230M tok

Model

GPT-5

Direct (list price)

$725

upstream provider / mo

Anvat metered (30% off)

$507

pay-as-you-go / mo

Anvat + 2× credit match

$254

effective burn rate / mo

Annual savings vs direct — $5,655 (65% lower)

Research / audit

Deep-research agent, weekly cadence

Claude Opus 4.8

Long-running research agent that produces a 5-10K-token brief from web sources weekly. Opus 4.8 for synthesis quality; Brave Search MCP for retrieval.

Input / mo

15M tok

@ $15/M

Output / mo

4M tok

@ $75/M

Total

19M tok

Model

Claude Opus 4.8

Direct (list price)

$525

upstream provider / mo

Anvat metered (30% off)

$368

pay-as-you-go / mo

Anvat + 2× credit match

$184

effective burn rate / mo

Annual savings vs direct — $4,095 (65% lower)

Codebase security audit pass

Claude Opus 4.8

Opus 4.8 reviewing security-critical code at a 50K-LOC scale, similar workflow to the Zcash discovery. Heavy reasoning, low token volume relative to compute.

Input / mo

8M tok

@ $15/M

Output / mo

3M tok

@ $75/M

Total

11M tok

Model

Claude Opus 4.8

Direct (list price)

$345

upstream provider / mo

Anvat metered (30% off)

$241

pay-as-you-go / mo

Anvat + 2× credit match

$121

effective burn rate / mo

Annual savings vs direct — $2,691 (65% lower)

Batch processing

Nightly batch classification, 200K rows

DeepSeek V4 Pro

Overnight batch job classifying 200K records (e.g., support tickets, transaction descriptions). DeepSeek V4 Pro is the cheap-enough tier for this scale.

Input / mo

60M tok

@ $1.74/M

Output / mo

12M tok

@ $3.48/M

Total

72M tok

Model

DeepSeek V4 Pro

Direct (list price)

$146

upstream provider / mo

Anvat metered (30% off)

$102

pay-as-you-go / mo

Anvat + 2× credit match

$51.2

effective burn rate / mo

Annual savings vs direct — $1,140 (65% lower)

Multimodal batch with Gemini 3.5 Pro

Gemini 3.5 Pro

10K images + text labels processed daily for catalog enrichment. Gemini 3.5 Pro's vision tier on Anvat is straight passthrough — no shape changes.

Input / mo

100M tok

@ $1.25/M

Output / mo

20M tok

@ $10/M

Total

120M tok

Model

Gemini 3.5 Pro

Direct (list price)

$325

upstream provider / mo

Anvat metered (30% off)

$227

pay-as-you-go / mo

Anvat + 2× credit match

$114

effective burn rate / mo

Annual savings vs direct — $2,535 (65% lower)

Your workload not here?

Plug your own token estimates into the cost calculator to see the same math for your numbers. If you want a custom workload sketched (we add one every couple of weeks), email hello@anvat.app.

Related: paste your monthly bill for an instant before/after, or see the value leaderboard for the best $/intelligence pick this week.