1. Builds
  2. Cheap AI agent team (Sonnet/Opus lead + Haiku subagents)

Cheap AI agent team (Sonnet/Opus lead + Haiku subagents)

Run a coding or ops agent team on the Claude API. A lead model plans and reviews, and cheap Haiku subagents do the routine work, for about 68% less than an all-Sonnet team.

Priced from vendor pages
Prices checked
16 Oct 2026
Exchange rate
1 USD = £0.7489
Sources
4
How we check

The stack

AIClaude Sonnet 5.5Plans, splits work, reviewsusage
Layer
Lead (default)
Plan
$2.00 in / $2.50 cache write / $0.10 cache read / $10.00 out per MTok
What it does here
Plans, splits work, reviews
Note
Verified
AIClaude Opus 5.5Hard tasks onlyusage
Layer
Lead (premium)
Plan
$4.00 / $5.00 / $0.20 / $20.00 per MTok
What it does here
Hard tasks only
Note
About doubles the bill
AIClaude Haiku 5.5 (prompts ≤100k)Reads, search, tests, summariesusage
Layer
Subagents
Plan
$0.10 / $0.125 / $0.01 / $0.50 per MTok
What it does here
Reads, search, tests, summaries
Note
Prompts over 100k cost 5×: $0.50 / $0.625 / $0.05 / $2.50
HOSTVercel ProHobby is non-commercial£14.98
Layer
Hosting (only if you sell it)
Plan
$20/seat
What it does here
Hobby is non-commercial
Note
Not in the totals below
OPSGitHubDependabot + secret scanningFree
Layer
Monitoring and security
Plan
Free
What it does here
Dependabot + secret scanning
Note
Secret scanning + push protection free on public repos; private Secret Protection price not found

Steps

  1. Create a Claude Console account, set a spend limit and create an API key. (You do this yourself; we don't sign up on your behalf.)

    Done when a spend limit is set and a test call returns.

  2. Put the lead agent on Sonnet 5.5 to plan, split work and review.

    Done when the lead turns a task into subagent jobs.

  3. Put the subagents on Haiku 5.5 for file reads, search, tests, summaries and boilerplate.

    Done when a subagent completes a routine job and the lead accepts it.

  4. Turn on prompt caching for shared system prompts and repo context.

    Done when usage shows cache-read tokens on repeat calls.

  5. Keep every Haiku prompt under 100k tokens by passing summaries, not whole files.

    Done when logged prompt sizes stay under 100k.

  6. Escalate to an Opus lead only for hard tasks.

    Done when you have a written rule for when to escalate.

  7. Log usage per call and compare it with the assumptions after a week.

    Done when you have a measured cost per task to replace the assumption.

Watch out for

  • Haiku prompts over 100k tokens cost 5×, and the higher rate applies to every token in that request.
  • Output tokens dominate the lead's cost. Cap max_tokens and ask for diffs, not full files.
  • An Opus lead roughly doubles the bill, so make Sonnet the default.
  • Benchmarks suggest Haiku is weaker at complex agentic coding: 39.2% vs Sonnet's 70.6% on Terminal-Bench 4.0. That figure is secondary: Anthropic's own figure, which we haven't re-checked. A cheap team that needs reruns isn't cheap.

What it costs each month

API spend with the default Sonnet lead + 4 Haiku subagents at $0.1367 per task (assumed token profile):

Stage Platform (no Lovable) Stripe fees Total + Lovable Pro
(a) Hobby (100 tasks/mo, $13.67) £10.24 £0.00 £10.24 £32.24
(b) Daily driver (1,000 tasks/mo, $136.70) £102.37 £0.00 £102.37 £124.37
(c) Heavy / production (10,000 tasks/mo, $1,367.00) £1,023.75 £0.00 £1,023.75 £1,045.75

Comparison at the same assumed tokens: | Stage | Opus lead + 4 Haiku | All-Sonnet (4 Sonnet subagents) | |-------|--------------------|-------------------------------| | (a) | $25.77, £19.30 | $42.30, £31.68 | | (b) | $257.70, £192.99 | $423.00, £316.78 | | (c) | $2,577.00, £1,929.92 | $4,230.00, £3,167.85 |

  • Assumed tokens per task (one task is one ticket or a PR-sized change):
  • Lead: 20k uncached input, 10k cache write, 60k cache read, 5k output. Sonnet costs $0.1210; Opus costs $0.2420.
  • Each of the 4 Haiku subagents: 15k / 5k / 30k / 3k, costing $0.003925 each ($0.0157 for all four).
  • Not assumed: Batch API, US-only routing, web search or Managed Agents.
  • The Sonnet lead accounts for 89% of the remaining cost.
  • Max credit offset (optional): Max 5x's $100 credit covers all of (a). At (b) you'd pay $36.70, or £27.48, of API spend on top of the Max subscription, which itself isn't included. Eligibility and terms are unknown.
  • Stripe doesn't apply because this prices the agent team, not a product built around it.
  • Excluded: tax/VAT (Anthropic says its prices exclude applicable tax). Claude Pro/Max/Team subscriptions. Web search, Managed Agents runtime, and code execution beyond the free hours. Haiku prompts over 100k. US-only inference (+10%), fast mode, retries and failed runs. The ~30% tokeniser uplift if your estimates come from older-model token counts.

Still open

  1. The tokens-per-task profile is unknown until measured, and every total scales linearly with it.
  2. Max API credit eligibility, rollover and terms are unknown. The Help Center article wasn't opened.
  3. The Max 20x monthly price is unknown. The fetched page says "from $100".
  4. Nothing needs a browser check, because both Anthropic pages returned prices as text.

Sources

Other verified figures: Batch API 50% off; US-only inference 1.1×; web search $10 per 1,000; Managed Agents $0.08 per session-hour; code execution 1,550 free hours/month, then $0.05/h; newer tokeniser produces about 30% more tokens. Max 5x/20x include $100/$200 a month in API credits (Team up to $500). The Max 20x price is unknown.

Change log

Date Change Source
2026-10-16 New desk-costed build Stackguide desk research
2026-10-23 USD prices re-converted at the Bank of England reference rate, 1 USD = 0.7489 GBP (16 Oct 2026), the one rate used site-wide (was 0.7535, exchangerates.org.uk 15 Oct 2026 close); £ figures and totals re-derived Bank of England series XUDLUSS, checked 2026-10-23