The model is half the bill. The harness is the other half — and most teams never measure it.
On July 12, 2026, Systima published a controlled comparison: same machine, same pinned model (claude-sonnet-4-5), logging proxy at the API boundary, 185 tamper-evident request/response records. Hacker News sent it to 456 points — not because anyone doubted models cost money, but because the meter climbs faster under Claude Code than OpenCode for identical "Reply: OK" tasks.
This is the explainx.ai read on what was measured, what HN argued about, and what enterprises should log after Nadella's trust-boundary framing.
TL;DR — floor vs whole task
| Metric | Claude Code | OpenCode |
|---|---|---|
| T1 floor ("Reply OK") | ~32,800 tokens first turn | ~6,900 tokens |
| Tool count (default) | 27 (+ orchestration suite) | 10 (coding core) |
| T2 file summarise | 6 requests, ~199k cumulative input | 4 requests, ~41k |
| T3 write-test-fix | 3 requests, ~121k (wins) | 9 requests, ~132k |
| Cache prefix stability | 3 classes; mid-session re-writes | Byte-identical every run |
| 2 subagents on T3 | ~513k (4.2× direct) | Leaner subagent profile |
Headline: Claude Code starts ~4.7× hungrier on Sonnet 4.5 floor; ~3.3× on Fable 5. Whole-task totals can invert when batching beats serialization.
Why measure at the API boundary
Systima's setup:
harness (Claude Code / OpenCode)
→ logging proxy (payload + usage block)
→ model endpoint
The proxy captures ground truth: exact JSON bodies (system blocks, tool schemas, messages) plus API usage (input, cache write, cache read, output).
That mirrors what we tell enterprise clients: treat inference logs like an audit trail, not a dashboard vanity metric. Systima chained records with @systima/aiact-audit-log — same integrity pattern as regulated AI logging. A token benchmark is a toy; a credit-decisioning agent is not.
If your traffic passes through a gateway (Systima used Meridian), measure the envelope with a bare request first. Otherwise you are grading your proxy, not the harness.
Part I — The floor (22-character prompt)
Task T1: "Reply with exactly: OK" — three runs per harness, fresh config, no MCP, no memory, empty workspace.
| Component | Claude Code | OpenCode |
|---|---|---|
| System prompt | 27,344 chars (3 blocks) | 9,324 chars (1 block) |
| Tool schemas | 27 tools, 99,778 chars | 10 tools, 20,856 chars |
| First-message scaffolding | 7,997 chars <system-reminder> | none |
| Calibrated first-turn payload | ~32,800 tokens | ~6,900 tokens |
Zero-tools variant: Claude Code system prompt alone ~6.5k tokens; OpenCode ~2.0k — still more than 3× residual doctrine (tone, safety, task management).
explainx.ai note: Tool schemas dominate both harnesses. Progressive tool disclosure — project-scoped MCP, not global kitchen-sink — is the first cost lever most teams skip.
Where the gap closes — T3 inversion
Multi-step FizzBuzz write-run-test-fix inverted the narrative:
| Claude Code | OpenCode | |
|---|---|---|
| Model requests | 3 | 9 (+1 title Haiku call) |
| Tool style | parallel batch | one tool per turn |
| Cumulative input | ~121,000 | ~132,000 |
Formula that matters:
whole_task_input ≈ baseline × request_count + conversation_growth
Claude Code paid 33k × 3; OpenCode paid 7k × 9. Totals converged.
HN takeaway (PUSH_AX): "Contractor A quoted $33k, Contractor B $7k — are we measuring the right thing?" Systima acknowledged this and promised deeper tasks + qualitative comparison. Fair. Tokens are necessary, not sufficient. But on T1–T2, both harnesses completed correctly — the gap was cost for identical outcome, which is measurable.
Part II — Multipliers (real sessions)
1. Instruction file (+20k each)
72KB production AGENTS.md / CLAUDE.md → ~20,000 tokens per request on both harnesses.
Footgun: Claude Code 2.1.207 ignored AGENTS.md until renamed CLAUDE.md. OpenCode reads either. Silent ignore = zero value, full tax if you duplicate both.
2. MCP servers (~1–1.4k per server)
Five public servers: +4,900 tokens (CC payload) / +6,967 metered (OpenCode). Production APIs ship much larger schemas.
Footgun: project .mcp.json ignored in print mode until --mcp-config passed.
3. Framework templates (size × turns)
8,405-char slash-command template ≈ 2,100 tokens — re-carried every subsequent request. Nine-turn session → nine× template tax.
4. Subagents (largest multiplier)
Same T3 direct: 121k → two parallel subagents: 513k (4.2×).
Each subagent pays its own bootstrap (~3,554-char agent prompt + 24/27 tools); parent ingests transcripts. HN reports of 7, 41, even 415 Fable subagents on one prompt are not outliers — they're the same mechanism runaway.
Mitigations from practitioners:
CRITICAL: Do NOT spawn sub-agents unless I explicitly request.
Use Explore (Haiku) subagents, not parent-model inheritance.
Or --disallowedTools Task / concurrency caps.
5. Extended thinking
Systima declined to publish numbers — gateway overrode harness toggles. Mechanism still compounds: reasoning blocks join history and re-send.
The "everything" cold start
| Config | First-request metered input |
|---|---|
| OpenCode: 11 MCP + 72KB instructions | 90,817 tokens (179 tools, 277KB schemas) |
| Claude Code: 4 MCP + plugins + instructions | ~75,000 tokens (118 tools, 311KB payload) |
~12× OpenCode's lean floor — your config sets the bill, not the harness alone.
Cache economics — stability beats size
Both set cache breakpoints. Discount does not erase:
- Write premium — TTL expiry (5 min meeting, lunch) re-primes full stack at 1.25×–2×
- Read × request count — subagents and serial loops multiply
- Context window — 85k bootstrap = 40%+ of 200k before code; compaction spends more tokens
Decisive difference: prefix stability.
| OpenCode | Claude Code | |
|---|---|---|
| Prefix across runs | Byte-identical | Varies by session |
| Mid-session re-write (T2) | 0 | 43k–54k tokens |
| Cache-write vs OpenCode | 1× | 5.9×–54× |
Re-writing a byte-identical prefix buys zero quality — pure waste. Prompt caching guide covers breakpoint design; Systima proves harness implementation matters as much as provider pricing.
Fable 5 floor — model-conditional bloat
Re-run on Fable 5: Claude Code system prompt 27,787 → 10,526 chars; tool schemas trimmed. OpenCode unchanged. Gap 4.7× → 3.3×.
Implication for July's limit-reset churn: upgrading models changes harness payload, not just capability.
HN debate — malice vs incompetence vs architecture
Three camps surfaced:
| Camp | Argument |
|---|---|
| Conflict of interest | Model vendor ships harness → token merchant incentives |
| Architecture | CC is orchestration platform (Cron, Task, skills, subagents); OpenCode is lean REPL |
| Measurement | Cache hits cheapen floor; quality on hard tasks untested |
explainx.ai synthesis:
- Vendor harness ≠ neutral infrastructure — aligns with Nadella's Control pillar: own traces, private evals, model-agnostic runners
- Subscription vs API changes incentives — flat-fee users hit rate limits; enterprise API users hit $/token × unstable cache
- Goodhart arrives fast — optimizing "tokens" without $/merged PR or human edit ratio is spec gaming
- Pi (~1k), maki.sh, custom 200-line proxies — competition is fierce; gains get copied
Practical floor from HN: --system-prompt "" on Claude Code CLI; verify in /context. Strips doctrine, not tools.
What enterprises should do Monday
| Action | Why |
|---|---|
| Deploy API-boundary logging | Payload + usage per request; SHA-chained if regulated |
| Baseline harness A vs B | Same golden tasks from private eval set |
| Score $/successful-task | Not tokens alone — include cache writes |
| Audit MCP + instruction files | One team reported $400k/yr zero-value system-prompt waste |
| Subagent policy in repo rules | Default-off beats post-hoc budget shock |
| Decouple model from harness | OpenCode, Pi, claudex |
explainx.ai read
Systima did what most "Claude Code vs OpenCode" threads skip: count bytes at the wire.
- Floor gap is real — 33k vs 7k is not folklore; it's tool schema + orchestration doctrine
- Multipliers are non-linear — subagents are the cigarette of agent economics
- Cache instability is a silent tax — bigger than model list price for some session shapes
- Whole-task math inverts — batching matters; procurement needs task-shaped benchmarks
- Logging is the product — token accounting falls out once you own the boundary
We teach harness literacy in workshops: proxy → baseline → loop → eval → cost dashboard. The model API is rented; what you send is yours.
Related on explainx.ai
- OpenCode Desktop tabs — session UI, worktrees gap (Jul 15)
- geohot — labs may not capture AI value
- How to build your own enterprise AI benchmark
- Nadella Reverse Information Paradox
- OpenCode complete guide
- Prompt caching decision framework
- Token spend governance — Ramp data
- GPT-5.6 claudex setup
- Agent harness engineering
- Specification gaming & Goodhart
- PxPipe — cut Claude Code image context
- Ploy GPT-5.6 migration — harness tuned to incumbent
Figures from Systima's July 12, 2026 study (Claude Code 2.1.207, OpenCode 1.17.18, Sonnet 4.5 + Fable 5 floor re-run). Harness versions change frequently — treat numbers as a snapshot; treat the measurement method as durable.