OpenRouter vs Crazyrouter: Four Fields to Audit Your API Bill

·
5 min read
·
AI-assisted

AI tooling disclosure

Generative AI materially contributed to this article.

Human review confirmed by the author.

OpenRouter vs Crazyrouter: Four Fields to Audit Your API Bill
Article

OpenRouter vs Crazyrouter: Four Fields to Audit Your API Bill

Send the same Claude request to OpenRouter and to a second gateway and both answer 200. Both expose an Anthropic Messages endpoint, both accept x-api-key and a Bearer token, and both bill at the vendor's list price. The surface is a tie.

What is left to compare is engineering: what gets added on top of the token price, and what happens after the request leaves your code. Four things actually differ, and only one of them is about model count.

The parts that are identical

Both gateways give you one key and one balance across OpenAI, Anthropic, Google and the open-weight labs, prepaid credit, image, video and speech models under the same key, per-key spend limits, and card or Alipay top-ups. Both serve OpenAI Chat Completions and Anthropic Messages, so Claude Code runs on either without a client change.

That last point is worth stating plainly, because a lot of comparison tables still imply otherwise: "Anthropic-compatible" is not a differentiator here. It is table stakes.

Where they diverge

Numbers come from each vendor's public API and documentation; the main set was verified on 12 October 2026, with other dates marked inline.

OpenRouter The other gateway
Model entries (normalized unique IDs) 458 / 372 177 / 177
Protocols Chat, Responses, Messages Chat, Messages (both auth headers work)
Open-weight models Priced per hosting provider One price per model ID
Where limits and fallback live Request or @preset API key (allowlist + budget)
Bring your own provider keys No platform fee up to $25k/month, then 5% No documented equivalent
Batch and flex pricing About half price No batch endpoint documented
Cache miss reason None; you only see cached_tokens drop to 0 diagnostics.cache_miss_reason
Thinking token count Normalized into usage output_tokens_details.thinking_tokens
Zero data retention Available on every plan No ZDR option documented; images retained 48h
In-region routing EU or US on Business None

The response body is the real split

One side normalizes usage across vendors; the other passes Anthropic's response through untouched. Normalization buys you portability: move from Claude to Gemini to DeepSeek and your usage parsing does not change.

The cost shows up when something breaks. On the native side there are four fields that OpenAI-compatible shapes cannot express:

  • cache_creation.ephemeral_5m_input_tokens and cache_creation.ephemeral_1h_input_tokens — cache writes counted in two tiers
  • output_tokens_details.thinking_tokens — reasoning tokens listed separately from output
  • diagnostics.cache_miss_reason — why the cache did not hit (the test run returned previous_message_not_found)
  • safeguard_results — safety check results
# Two assertions that turn a silent failure into a loud one
diag = resp.get("diagnostics") or {}
assert resp["usage"]["output_tokens_details"]["thinking_tokens"] > 0, "no thinking block: check budget_tokens"
if reason := diag.get("cache_miss_reason"):
    print(f"cache miss: {reason}")

The first assertion matters more than it looks. A budget_tokens value that is too small does not raise an error; the API returns 200 with thinking_tokens at 0. Across four Claude models, 128 produced no thinking block and 1024 did.

Where the rules live

OpenRouter retries on another model using a models list in the request, or a saved preset referenced as @preset/<name>, plus an openrouter/auto router. The client has to send something OpenRouter-specific.

The other gateway puts the rules on the API key: model allowlist, token budget. The client keeps sending its usual model name; an unlisted model returns model not allowed (403) and an over-budget key returns 429.

Rate limits differ in scope too. OpenRouter states that extra keys do not raise your limit — capacity is managed globally. The other gateway's troubleshooting doc lists spreading traffic across multiple tokens as the fix for 429s, which means limits are counted per token.

The bill: markup is not the story

OpenRouter adds its fee at top-up: 5.5% on card payments for Standard (8% on Business), non-refundable, which lands at list price × 1.055 per token.

Model Vendor list price OpenRouter (with fee) The other gateway (discount)
Claude Opus 5.5 $4 / $20 $4.22 / $21.10 Group discount
Claude Sonnet 5.5 $2 / $10 $2.11 / $10.55 Group discount
GPT-6.1 Sol $2 / $10 $2.11 / $10.55 Group discount
GPT-6 Astra $10 / $50 $10.55 / $52.75 35% below list
Gemini 3.8 Flash $0.75 / $3.75 $0.79 / $3.96 Group discount
GLM-5.3 $1.40 / $4.40 $1.48 / $4.64 Group discount

The discount column states the discount only, not a converted price: group discounts differ per model and change with the price page. The 35% figure is a 13 October 2026 snapshot for one model ($6.50 / $32.50 against a $10 / $50 list).

At $1,000 a month of list-price Astra usage, OpenRouter comes to $12,660 a year; the discounted side projects to $7,800. A $4,860 gap.

But the larger number is the cache. Anthropic's rules — 1.25× on cache writes, 0.1× on reads, 2× for the 1-hour TTL — apply on both sides. A 60k-token context, 40 turns a day:

  • No cache hit: 0.06M × $2 = $0.12 per turn, $4.80 per day
  • Full cache hit: $0.15 to write, then 39 turns at $0.012, $0.62 per day

That is $4.18 a day, about $125 a month — a 10× spread on identical tokens. And only one side tells you why a miss happened.

Picking a side

Neither column is all green. The honest split looks like this:

Your requirement Side Why
Multi-model evaluation, long-tail open weights OpenRouter 372 vs 177; Qwen3 37 vs 12
Code that switches vendors often OpenRouter Normalized fields, no parser changes
Cross-provider redundancy OpenRouter Multiple hosts with Sticky Routing
Data residency, ZDR, sovereign cloud OpenRouter Documented ZDR and in-region routing
Implementing rate-limit policy from docs OpenRouter 402/403/429 split plus in-flight budget
Model pinned to Claude tooling Other gateway Native fields, thinking and tool_use intact
Debugging why the cache missed Other gateway cache_miss_reason in the response
Per-token budgets and environment isolation Other gateway Built in, no wrapper needed
One key for chat, images and speech Other gateway Separate endpoints

Claude Code switches with three environment variables:

export ANTHROPIC_BASE_URL="https://api.crazyrouter.com"
export ANTHROPIC_AUTH_TOKEN="your key"
export ANTHROPIC_MODEL="claude-sonnet-5-5"

What this comparison does not cover

Latency and error rates are not measured — neither side offers a stable basis for a single-machine comparison. The 429 threshold is inferred from a troubleshooting doc, not from load testing. The OpenRouter side is documented, not paid-tested, so actual billing behaviour is best confirmed against your own invoice. And the 35% discount is a one-day snapshot for one model; do not extrapolate it.

Pull the numbers yourself before committing:

curl -s https://openrouter.ai/api/v1/models | jq '.data | length'
curl -s https://crazyrouter.com/api/pricing | jq '.data | length'

Sources

Crazyrouter

Technical sources

OpenRouter (comparison sources)

Comments(0)

Join the discussion by logging into your account.

No comments yet. Be the first to comment!

xiaoying zhou

Passionate developer sharing knowledge about modern web technologies and best practices.

Subscribe to xiaoying zhou's Newsletter

Direct email dispatches when new stories are published. Zero algorithms.