Skip to content

cost-ladder.md

Latest commit

 

History

History
234 lines (191 loc) · 12 KB

cost-ladder.md

File metadata and controls

234 lines (191 loc) · 12 KB

Cost Ladder

Concrete tier choices for opencode-sdlc-wizard users — what the SDLC loop actually costs at three monthly budgets, what trade-offs each budget forces, and which models clear the SDLC capability floor in each slot.

This is the document that complements the four privacy tiers in PRIVACY.md. Privacy is "where does my data go?" Cost is "how much am I paying for ceiling vs floor?" Most teams need to balance both.

SDLC capability floor (the unmoving constraint)

Before picking a model, know the floor. The SDLC protocol (plan → TDD → self-review → optional cross-model review) requires:

  1. Decent instruction-following — multi-step tool calls without hallucinated arguments.
  2. Long-context reasoning — tracking a plan across 20-50 turns.
  3. Tool-use reliability — handling structured edits, bash output, schema validation.
Class Examples Holds the protocol?
7B–13B Mistral 7B, Llama 3.2 8B, Qwen-Coder-7B No. Fails partway.
14B–22B Phi-4 14B, DeepSeek-Coder-V2-16B, Codestral 22B Borderline. Routine fixes OK; complex refactors slip.
30B–70B code-tuned Qwen2.5-Coder-32B, Qwen3-Coder, Llama 3.3 70B Yes. The local sweet spot.
100B+ open-weight DeepSeek-V3.1 (671B MoE), GPT-OSS 120B Yes. Strongest OSS reasoning.
Frontier proprietary Claude Opus, GPT-5.5 xhigh, Gemini 2.5 Pro Yes. Ceiling.

Below the 30B-code-tuned bar, the wizard installs and runs but the protocol degrades. That's a capability outcome, not a port bug.

$0/mo path — pure free tiers

For hobbyists, students, or anyone evaluating the wizard before committing.

# Detect what's reachable, biased toward free providers
bash .opencode/scripts/detect-backends.sh --free-tier-first

# Pick the highest-leverage free option (verify the exact model ID
# against the provider's current catalog — see "Verifying current
# model IDs" below; these change faster than this doc ships)
bash .opencode/scripts/configure-backend.sh \
  --tier hosted_oss --provider cerebras \
  --model qwen-3-235b-a22b-instruct-2507
Slot Provider Example model (verify before use) Why Cost
Coder Cerebras free qwen-3-235b-a22b-instruct-2507 or gpt-oss-120b Fastest inference (~2000 tok/s), generous daily quota $0
Coder (alt) Groq free llama-3.3-70b-versatile Sub-second, free tier resets daily $0
Coder (alt) Google AI Studio gemini-2.5-flash (1M context) Generous daily request quota — verify current limits $0
Coder (alt) OpenRouter deepseek/deepseek-chat:free Aggregator routing, OSS models $0 (rate-limited)
Reviewer NVIDIA NIM deepseek-ai/deepseek-r1 Reasoning model, free credits at build.nvidia.com $0
Reviewer (alt) OpenRouter qwen/qwen-3-coder:free Code-tuned, free routing $0 (rate-limited)

Verifying current model IDs. Provider catalogs change monthly. Before pinning a model, sanity-check the ID against the provider's live docs — Cerebras: https://inference-docs.cerebras.ai/models/overview, Google AI Studio: https://ai.google.dev/gemini-api/docs/models, NVIDIA NIM: https://build.nvidia.com, Groq: https://console.groq.com/docs/models. The IDs in this table were calibrated 2026-05-05; codex round-1 review caught llama-3.3-70b as no longer in Cerebras's catalog and gemini-2.0-flash as deprecated (shutdown June 2026). If configure-backend writes a stale ID, opencode run will surface a clear model not found error — fixable in seconds, not a footgun.

Real constraints at $0:

  • Daily quotas reset, but agentic loops burn fast — one complex task can exhaust Groq's daily cap.
  • OpenRouter :free variants are aggressively throttled under load.
  • Free tiers may log your prompts for training (read each provider's policy). Privileged-data work belongs in private_local, not free hosted.
  • No SLA. Outages happen.

What you give up vs paid: ceiling on novel architecture decisions, long-context refactors, and "this is gnarly, throw the strongest model at it" moments. For routine fixes + TDD loops, the gap is small.

$0/mo path — local-only (no egress)

Same budget, different trade-off: hardware capex instead of token spend. For privileged-data work, NDA contracts, or anyone who wants zero egress.

Slot Tool Model Hardware floor
Coder Ollama qwen3-coder:30b 24GB VRAM (RTX 4090, M-series 32GB+)
Coder Ollama deepseek-coder-v2:16b 16GB VRAM (RTX 4070 Ti, M Pro 16GB)
Coder LM Studio Qwen3-Coder-30B-Instruct Same as above
Coder MLX (Apple Silicon) mlx-community/Qwen2.5-Coder-32B-Instruct-4bit 32GB unified memory; fastest on M-series
Reviewer Ollama qwen3-coder:30b (same instance) Use sequentially 
# Local + privacy-first
bash .opencode/scripts/configure-backend.sh \
  --tier private_local --provider ollama \
  --model qwen3-coder:30b

Constraints:

  • Hardware is the floor. <16GB VRAM = no serious local work for SDLC.
  • Inference is slower than hosted (token/sec on a single consumer GPU vs Cerebras's wafer-scale farm).
  • One model at a time means no parallel coder+reviewer (or you swap contexts).

$20/mo path — one premium subscription, harvest free for the rest

Most professional engineers' sweet spot. One ceiling model where ceiling matters, free for the rest.

Slot Provider Model Cost Why
Coder Anthropic Claude (Pro) claude-sonnet-4.6 $20/mo Pro sub or ~$3/M in via API Best instruction-following + tool-use
Coder (alt) OpenAI Plus gpt-5.3-codex $20/mo Plus sub Strong reasoning, especially with xhigh effort
Coder (alt) Z.AI GLM Coding Plan glm-5.1 $10/mo (or $30/quarter, $80/year — quarterly restructure May 2026; no flat-$18 SKU anymore) Post-Anthropic-OAuth-ban migration target; verify at https://z.ai/subscribe
Reviewer Cerebras free / Groq free gpt-oss-120b $0 Cheap second opinion; both Cerebras + Groq host gpt-oss-120b
Reviewer (alt) DeepSeek direct deepseek-v4-flash ~$0.14/M in cache-miss (pennies/review) — verify current pricing Strong OSS reasoning, cheapest hosted, V4 family shipped Apr 2026
Reviewer (high-stakes) Codex via API gpt-5.3-codex xhigh ~$1-3 per review at xhigh When release-critical 
# v0.10.0: Mixed-Mode in one shot via `pick` — coder + reviewer split
# the work, OpenCode auto-routes review tasks via agent.review.model
npx opencode-sdlc-wizard pick \
  --tier proprietary --provider anthropic \
  --reviewer-tier hosted_oss --reviewer-provider cerebras

# Equivalent low-level invocation (what `pick` orchestrates under it):
bash .opencode/scripts/configure-backend.sh \
  --tier proprietary --provider anthropic --model claude-sonnet-4.6 \
  --reviewer-tier hosted_oss --reviewer-provider cerebras --reviewer-model gpt-oss-120b

The hybrid pattern that maximizes this budget:

  • Coder = ceiling (Claude / GPT-5.5) for new code
  • Reviewer = cheap (DeepSeek direct or Cerebras free) for the "second pair of eyes" pass
  • Saves ~80% of review tokens vs running Claude on both sides without losing rigor

v0.10.0 ships Mixed-Mode in pick and configure-backend.sh — emits both provider blocks plus agent.review.model so OpenCode routes review tasks to the reviewer without any wrapper script.

$200/mo path — agentic-grade, multi-agent loops

For shops running CI gates on every PR, daily multi-hour agent loops, or anyone whose time is more valuable than tokens.

Slot Provider Model Approx monthly Why
Coder Anthropic claude-opus-4.7 ~$100-150 Highest ceiling, 1M context tier
Coder (alt) OpenAI gpt-5.3-codex xhigh ~$80-130 Strongest reasoning at xhigh effort
Reviewer Codex CLI (xhigh) gpt-5.3-codex ~$30-50 Cross-model review on every release
Reviewer (parallel) DeepSeek direct deepseek-r1 ~$10-20 Cheap second-reviewer for triangulation
CI gate Groq free / Cerebras free gpt-oss-120b (both providers) $0 Fast PR-review loops, free tier; v0.10.3 picker default

The pattern that earns this budget:

  • Three reviewers: codex (xhigh), DeepSeek-R1, plus the originating coder's self-review. Triangulated findings catch what any single reviewer misses.
  • CI gates on every PR using a free fast reviewer (Groq) — the expensive reviewers fire only on release branches.
  • 1M-context tier on the coder for whole-monorepo refactors.

Picking by job, not by tier

Forget the budget bracket for a sec — pick by what the job is.

Job Best tier Best model Notes
Routine fix / typo / small CSS hosted_oss free Cerebras gpt-oss-120b or qwen-3-235b-a22b-instruct-2507 Fast + free + clears the bar
TDD on new feature, mid-stakes $20 path Claude Sonnet 4.6 Tool-use is its strength
Long-context refactor $200 path Opus 4.7 (1M context) Floor is "fits in context"
Security audit / privileged code private_local Qwen2.5-Coder-32B local Zero egress
CI gating on every PR hosted_oss free Groq llama-3.3-70b-versatile Speed + cost dominate (Groq still ships Llama 3.3 70B; Cerebras dropped it)
Architecture decision (novel) $200 path Opus 4.7 + Codex xhigh review Ceiling matters
Bulk doc / refactor sweeps hosted_oss cheap-paid DeepSeek-V3.1 direct Cheap per token, decent quality
Air-gapped / compliance-locked private_local Whatever fits VRAM Hardware floor matters more than ceiling

Where this wizard helps regardless of tier

The SDLC protocol (plan → TDD → self-review → optional cross-model review) is model-agnostic. The hooks fire under any backend OpenCode supports. The skills work with any model that clears the capability floor.

That's the actual point: the wizard is not about a specific tier winning. It's about the discipline holding regardless of what model you're paying (or not paying) for.

Updating this doc

This is a snapshot — pricing and free-tier generosity shift faster than the wizard ships. If a number is stale by more than a few months or a provider has changed their tier structure, open an issue. The ratios (DeepSeek roughly an order of magnitude cheaper than Opus, Cerebras roughly an order of magnitude faster than hosted alternatives) tend to hold longer than the absolute prices.

Specific calibration notes for the prices/quotas in this doc:

  • DeepSeek deepseek-chat ≈ $0.14/M in cache-miss (was $0.27/M in v0.8.0 — codex round-1 F4 caught the stale figure). Cache-hit pricing is meaningfully cheaper. Verify at https://api-docs.deepseek.com/quick_start/pricing.
  • Gemini 2.0 Flash is deprecated with shutdown 2026-06-01 — use gemini-2.5-flash instead. Free quota is per-model and changes; some quotas apply only to grounded prompts. Verify at https://ai.google.dev/gemini-api/docs/rate-limits.
  • Cerebras catalog rotates faster than other providers — llama-3.3-70b was in v0.8.0 of this doc, but as of 2026-05-05 the public production models are llama3.1-8b, gpt-oss-120b, qwen-3-235b-a22b-instruct-2507, and zai-glm-4.7. Verify at https://inference-docs.cerebras.ai/models/overview.
  • Together initial credit and Groq daily quota change frequently — always recheck.

Last calibrated: 2026-05-18 — v0.10.6 sweep. Refreshed gpt-5.5 → gpt-5.3-codex (Feb 2026 release, most-pinned reviewer in surveyed configs), deepseek-chat → deepseek-v4-flash (V4 family April 2026), removed "Cerebras dropped llama-3.3-70b" qualifier from CI gate slot since both Cerebras and Groq host gpt-oss-120b (the v0.10.3 picker default). Added Z.AI GLM Coding Plan to $20/mo path with current quarterly pricing ($10/$30/$80 — no flat $18/mo SKU since May 2026 restructure). Anthropic Sonnet 4.6 / Opus 4.7 unchanged.

Prior calibration: 2026-05-05 (codex round-1 corrections for v0.8.1).