AI Coding Agent Cost Calculator

Compare Amp, Claude Code, Cursor, and Factory Droid modes. Each mode applies explicit assumptions about context size, autonomy loops, model choice, and retry behavior on top of the same token-cost formula.

Updated July 27, 2026 — model catalog refresh (click to expand what changed)

Amp: Replaced the deprecated smart/deep/rush mode rows with the current low/medium/high/ultra dial (ampcode.com/news/the-dial, July 9, 2026). New writer models: GLM-5.2 (low), GPT-5.6 Sol (medium/high), Claude Fable 5 (ultra), each paired with an oracle. Default selection changed from amp-smart-opus47 to amp-medium-gpt56sol. Legacy smart/deep/rush rows kept and labeled (classic, deprecated) for historical comparison.
Claude: Added Claude Opus 5, Sonnet 5 (new Claude Code default, $2/$10 intro through Aug 31 2026), and Fable 5. Default selection changed from claude-sonnet-46 to claude-sonnet-5. Opus 4.8/4.7 and Sonnet 4.6/4.5 rows relabeled (legacy) — still selectable in the API but superseded. Updated claude-default note: opus alias now resolves to Opus 5, sonnet to Sonnet 5.
Pricing sources: Anthropic pricing docs (Fable 5 $10/$50, Opus 5 $5/$25, Sonnet 5 $2/$10 intro, Haiku 4.5 $1/$5); OpenAI GPT-5.6 launch (Sol $5/$30, Terra $2.50/$15, Luna $1/$6). GLM-5.2 per-token price is a labeled planning proxy — Amp hosts it and has not published public per-token dollars.
Note on Fable 5: Claude Fable 5/Mythos 5 access was suspended June 12, 2026 under a U.S. export-control directive. The ultra tier and Fable rows remain in the catalog but may not be reachable for all orgs; Amp may fall back to Opus-class models.
Not changed: The underlying cost formula, UI, Cursor and Factory Droid catalogs, and all default team-size/usage assumptions are unchanged. Factory Droid still references GPT-5.5/Opus 4.7-class multipliers because Factory's published model list has not been re-verified for GPT-5.6.

Inputs ?

Amp mode

Governed agent workflow with context discipline and observability overhead.

Choose agent ? Chosen LLM / model ?

Engineers ? Adoption % ?

Sessions / engineer / day ? Workdays / month ?

Input tokens / session ? Output tokens / session ?

Input price / 1M tokens ($) ? Output price / 1M tokens ($) ?

Fixed-seat benchmark / engineer ($) ? Monthly budget ($) ?

Context multiplier ? Agent loop multiplier ?

Manual-backed assumption: Amp has deep, smart, and rush modes; can use subagents, Oracle, and Librarian; and charges pass-through provider/API costs for individual and non-enterprise workspaces. Mode rows below are planning proxies, not invoices.

Results ?

Cost / session ? $0 Input $0 + output $0

Cost / active engineer / month ? $0 $0 per active engineer per workday

Team monthly cost ? $0 0 active engineers

Team annualized cost ? $0 Monthly run rate × 12

Over fixed-seat model ? 0× Token cost vs fixed seats

Budget burn ? — Enter a monthly budget to compare

Formula ?

effective tokens = base_tokens × context_multiplier × agent_loop_multiplier

session cost = effective_input_tokens / 1,000,000 × input_price + effective_output_tokens / 1,000,000 × output_price

monthly cost = active_engineers × sessions_per_day × workdays × session_cost

Typical use cases

Safety-critical RTOS migration: converting FreeRTOS to Chinese Loongson CPUs as an ISO 26262-compliant RTOS where all code is MISRA C 2025 compliant. The preset assumes 50 engineers, a tight schedule, heavy codebase context, Loongson/LoongArch porting, memory-protection design, temporal-independence evidence, safety-case documentation, static-analysis remediation, and repeated verification loops.

Enterprise SaaS platform: building a large SaaS platform: multi-tenant backend, web UI, billing, identity, audit logs, observability, CI/CD, data migration, security reviews, and integrations. The preset models 90 engineers working under product-delivery pressure with many cross-team agent sessions.

Save JSON stores the current run in this browser and downloads a reusable file. Save HTML downloads a standalone report. Load saved session imports a JSON file. Diff compares all browser-saved JSON snapshots. Clean old data wipes browser-saved snapshots so the next diff session starts fresh.