Cheap model fallback for Claude Code 核验日期 2026-09-17

Most agent tokens don't need a frontier model. Plan-then-execute keeps the flagship for architecture and hands the grunt work — file edits, boilerplate, docstrings, test scaffolding — to a flash-tier model that costs1–3% as much. Here's how to wire a cheap second model into each agent.

Why fallback saves so much

Interactive coding agents burn a large share of tokens on mechanical steps. Claude Code in particular also runs a background "small fast" model for sub-tasks — and it happily inherits your expensive default if you don't override it. Verified price spread for the same "smart enough" work (per 1M tokens, in/out):

Provider	Model	Input $/1M	Cache read $/1M	Output $/1M	Good as
DeepSeek	DeepSeek V4.1 Flash	$0.3	$0.006	$1.2	Fallback or heavy lifting
Z.ai (Zhipu GLM)	GLM-5.3	$1.4	$0.26	$4.4	Fallback or heavy lifting
Z.ai (Zhipu GLM)	GLM-5.3-Flash	$0.15	$0.03	$0.5	Primary daily driver / fallback
Z.ai (Zhipu GLM)	GLM-4.7	$0.6	$0.11	$2.2	Fallback or heavy lifting
Z.ai (Zhipu GLM)	GLM-4.7-Flash	Free	Free	Free	Primary daily driver / fallback
Moonshot AI (Kimi)	Kimi K2.6	$0.95	$0.16	$4	Fallback or heavy lifting
SiliconFlow	DeepSeek-V4-Flash	$0.13	$0.028	$0.28	Primary daily driver / fallback
SiliconFlow	GLM-5.3	$1.4	$0.26	$4.4	Fallback or heavy lifting
SiliconFlow	GLM-5.3-Flash	$0.15	$0.03	$0.5	Primary daily driver / fallback
SiliconFlow	Kimi-K2.6	$0.77	$0.14	$3.4	Fallback or heavy lifting
SiliconFlow	MiniMax-M2.5	$0.3	$0.03	$1.2	Fallback or heavy lifting
OpenRouter	DeepSeek V4 Flash (deepseek-v4-flash)	$0.07	$0.014	$0.14	Primary daily driver / fallback
Anthropic	Claude Sonnet 5 (baseline)	$2.00	$0.20	$10.00	Architecture, reviews, hard debugging

Pattern 1 — Claude Code: two shells, two models

Claude Code reads its endpoint from env vars at launch, so the simplest robust fallback is two profiles:

# ~/.bashrc — flagship profile (default)
alias cc-max='ANTHROPIC_BASE_URL="https://api.z.ai/api/anthropic" \
  ANTHROPIC_AUTH_TOKEN="YOUR_ZAI_KEY" ANTHROPIC_MODEL="glm-5.3" claude'

# ~/.bashrc — cheap fallback profile
alias cc-cheap='ANTHROPIC_BASE_URL="https://api.siliconflow.com/" \
  ANTHROPIC_AUTH_TOKEN="YOUR_SF_KEY" ANTHROPIC_MODEL="DeepSeek-V4-Flash" \
  ANTHROPIC_SMALL_FAST_MODEL="GLM-5.3-Flash" claude'

Within one session you can also steer manually: keep the cheap endpoint as default and promote hard problems to the flagship profile only when the first attempt stalls. The split typically lands 70–90% of tokens on the cheap model.

Pattern 2 — Codex CLI: provider profiles in config.toml

# ~/.codex/config.toml — switch with: codex --profile cheap
[profiles.cheap]
model = "DeepSeek-V4-Flash"
model_provider = "siliconflow"

[profiles.max]
model = "glm-5.3"
model_provider = "zai"

model_catalog_json = "~/.codex/models.json"   # GLM metadata, see Z.ai Codex guide

[model_providers.siliconflow]
name = "SiliconFlow"
base_url = "https://api.siliconflow.com/v1"
env_key = "SILICONFLOW_API_KEY"
wire_api = "responses"

[model_providers.zai]
name = "Z.ai GLM"
base_url = "https://api.z.ai/api/v1"
env_key = "ZAI_API_KEY"
wire_api = "responses"

Pattern 3 — Cline: per-task model override

Cline lets you switch the model in the settings panel between tasks. Keep two saved configurations — SiliconFlow DeepSeek-V4-Flash for “Act” mode bulk edits, GLM-5.3 for “Plan” mode reasoning (values in /configs).

Verdict

Default tier: GLM-5.3-Flash ($0.15/$0.50) or DeepSeek-V4-Flash on SiliconFlow ($0.13/$0.28) — both handle routine agent work at 1–3% of Sonnet 5's cost.
Escape hatch: keep a flagship profile one alias away for reviews and gnarly bugs.
Free tier exists: GLM-4.7-Flash is listed at $0 on Z.ai — usable as an ultra-fallback for non-critical sub-tasks.
Mind the ceiling: fallback models fail differently (weaker tool-calling, worse long-context recall). Route work by risk, not just cost.

FAQ

Does Claude Code have a native "fallback model" setting?: Not a cost-based one. It exposes ANTHROPIC_MODEL and ANTHROPIC_SMALL_FAST_MODEL (plus per-tier defaults like ANTHROPIC_DEFAULT_SONNET_MODEL) — which is enough to run flagship + cheap in parallel via profiles.
Will a flash model break tool calling?: No — GLM-5.3-Flash and DeepSeek-V4-Flash both support function/tool calls on their official endpoints. Quality on complex multi-step tool chains is where you'll notice the difference, not single calls.
How do I know which share of tokens went to the cheap model?: Check each provider's usage dashboard; if you split by profile alias, the per-key usage makes the split obvious.
Are these numbers first-party?: Yes — all prices on this page were read from official provider pricing pages on 2026-09-17. Unverifiable ones are marked 待核验.