OpenRouter pricing 2026: real 30-day bill vs direct providers
OpenRouter takes 5.5% on credit purchases with a $0.80 minimum, then passes provider rates through. Here is the real 30-day bill across Anthropic, OpenAI, and Mistral at four workload mixes, with the anti-patterns where the routing tax stops being worth it.
Updated on July 2, 2026
On this page
On a quiet Sunday in June 2026 I bought $200 of OpenRouter credits to settle a question my freelance clients keep asking. Is the 5.5 percent OpenRouter takes on top of every credit purchase a tax worth paying for the routing layer, or a tax that quietly bankrupts solo practices once you scale past hobby use? I priced four real workloads at June 2026 provider rates to find out.
Quick answer
As of June 2026, OpenRouter adds a 5.5% fee with a $0.80 minimum when you buy credits, then passes the underlying provider rate through to you with no per-token markup. For a heavy code agent on Claude Sonnet 4.6, the OpenRouter bill is $196.23/mo vs $186.00 direct, a $10.23 routing tax. For a low-volume Haiku chatbot, the $0.80 minimum makes OpenRouter 20.2% more expensive on a $3.96 base. OpenRouter wins when you need multi-provider routing, free failover on failed attempts, and a single bill; direct provider keys win on raw price for any single-model workload above roughly $50/mo. (Source: OpenRouter's pay-as-you-go pricing page, June 2026.)
The four providers in this teardown:
The actual 30-day math, no marketing spin
Four real workloads, four token mixes, two columns of bills. Every line item maps to a real shop I have priced for in 2026. Numbers come from the published pricing pages, not vibes.
Scroll to see more
| Workload | Model | Direct bill / mo | OpenRouter bill / mo | OpenRouter premium |
|---|---|---|---|---|
| Heavy code agent | Claude Sonnet 4.6 | $186.00 | $196.23 | +$10.23 (5.5%) |
| RAG retrieval | GPT-5.4-mini | $62.70 | $66.15 | +$3.45 (5.5%) |
| Light chat | Claude Haiku 4.5 | $3.96 | $4.76 | +$0.80 (20.2%) |
| Structured output | Mistral Medium 3.5 | $60.00 | $63.30 | +$3.30 (5.5%) |
| Network total | mixed | $312.66 | $330.44 | +$17.78 |
The last row is the number people argue about on Reddit and never actually compute. Across a realistic four-workload month, the OpenRouter tax is $17.78. The right question is not whether $17.78 is a lot. The right question is whether the routing layer earns it back in time saved or reliability gained. Below, the math row by row.
Mix A: Heavy code agent on Claude Sonnet 4.6
Profile: a solo dev or two-person shop running 4 to 6 hours per day of autonomous Claude Code style sessions across 30 days. Realistic monthly token volume: 40M input (about 50% cache reads from sticky system prompts and project context), 8M output. Model: Claude Sonnet 4.6 at the published $3 input / $15 output / $0.30 cache-hit June 2026 rate, per Anthropic's published model pricing.
Direct Anthropic math
- Uncached input: 20M × $3 / 1M = $60.00
- Cache reads (5-min cache hit at 0.1x): 20M × $0.30 / 1M = $6.00
- Output: 8M × $15 / 1M = $120.00
- Total: $186.00/mo
Via OpenRouter
You buy credits. Stripe charges OpenRouter, OpenRouter takes 5.5% off the top, then you spend the rest at the provider's posted rate.
- Spend $186 at the provider rate via OpenRouter: $186 + ($186 × 0.055) = $196.23/mo of card outflow (or equivalently you buy $196.83 of credits to net $186 of usable spend).
- Direct = $186.00, OpenRouter = $196.23, premium = +$10.23 (5.5%)
For a single shop on a single model the $10.23 buys you nothing direct does not already give you. Verdict for Mix A: go direct unless you specifically need failover routing to a fallback model on the same API.
Mix B: RAG retrieval on GPT-5.4-mini
Profile: a knowledge-base assistant doing ~2,500 doc lookups per day with small structured answers. Realistic volume: 80M input tokens (cache reuse rare, maybe 5%), 1.2M output. Model: GPT-5.4-mini at the published $0.75 input / $0.075 cached / $4.50 output rate per OpenAI's per-model rates.
Direct OpenAI
- Uncached input: 76M × $0.75 / 1M = $57.00
- Cached input: 4M × $0.075 / 1M = $0.30
- Output: 1.2M × $4.50 / 1M = $5.40
- Total: $62.70/mo
Via OpenRouter
- $62.70 + 5.5% = $66.15/mo, premium +$3.45.
Same story. Nothing OpenRouter adds for $3.45/mo on a single model on a single provider. Go direct.
Mix C: Light chat on Claude Haiku 4.5
Profile: a small SaaS support chatbot, ~20 customer chats per day, 2k tokens average per chat including system prompt. Realistic volume: 1.5M input (40% cached because the system prompt is sticky), 600k output. Model: Claude Haiku 4.5 at $1 input / $0.10 cached / $5 output (June 2026).
Direct Anthropic
- Uncached input: 0.9M × $1 / 1M = $0.90
- Cache reads: 0.6M × $0.10 / 1M = $0.06
- Output: 0.6M × $5 / 1M = $3.00
- Total: $3.96/mo
Via OpenRouter
The 5.5% credit fee on $3.96 is $0.22, which is below the $0.80 minimum (per OpenRouter's FAQ on credit purchase fees). So you eat the $0.80 floor.
- $3.96 + $0.80 = $4.76/mo, premium +$0.80 (20.2%)
This is where the minimum bites. If your monthly bill is under ~$15, the floor turns the 5.5% into a 5 to 25 percent effective tax. OpenRouter is bad economics for hobby toys. Either consolidate volume to clear $15/mo, or use the OpenRouter free tier (50 reqs/day on free models only) for that scale.
Mix D: Structured output extractor on Mistral Medium 3.5
Profile: an extraction agent pulling structured JSON from invoices, contracts, or scraped pages. Realistic volume: 20M input, 4M output (the JSON keeps growing as you add fields). Model: Mistral Medium 3.5 at $1.50 input / $7.50 output, per Mistral's published per-token rates (June 2026).
Direct Mistral
- Input: 20M × $1.50 / 1M = $30.00
- Output: 4M × $7.50 / 1M = $30.00
- Total: $60.00/mo
Via OpenRouter
- $60.00 + 5.5% = $63.30/mo, premium +$3.30.
Same pattern. The 5.5% is the routing tax for any single-provider workload. The picture changes only when you start routing across providers in a single product.
When OpenRouter actually loses
Five anti-patterns I have watched real teams walk into in 2026.
Scroll to see more
| Anti-pattern | Why OpenRouter loses | Direct fix |
|---|---|---|
| High-volume single-model | You pay 5.5% on every dollar with zero failover benefit | Direct API key, in-house monitor |
| Tier 3/4 eligibility | Anthropic and OpenAI negotiate discounts beyond 5.5% at volume | Talk to their sales team once you clear $5k/mo |
| Aggressive 1-hour cache writes | Anthropic exposes a 2x write 1-hour cache; provider-specific cache primitives are not consistently surfaced through routers | Direct Anthropic for cache-heavy reads |
| Latency-critical paths | The OpenRouter hop adds roughly 50 to 150 ms per call | Direct provider with same-region endpoint |
| Under-$15/mo bills | The $0.80 minimum credit fee becomes a 5 to 25 percent tax | Stay on free tier or go direct |
When OpenRouter actually wins
Three patterns where the 5.5% pays for itself, easily.
- Multi-provider routing as a feature, not a fallback. If your product calls Claude for reasoning, GPT-5.4 for tool use, and Mistral Small 4 for cheap classification in the same request flow, OpenRouter saves you three sets of keys, three billing dashboards, three rate-limit emails. The 5.5% is the price of skipping that ops layer.
- Free failover on failed attempts. Per the OpenRouter pricing page, failed or fallback attempts incur no charge when routing is enabled. For a production agent that retries up to 3 times across providers when one is degraded, that math compounds.
- The 1M free requests/month threshold. Pay-as-you-go gives you 1M free requests per month before a 5% per-request fee kicks in. A high-RPS, low-token workload (small-prompt classifier, sentiment scorer, query router) can ride that free band for a long time before the meter starts.
The 5.5% math: when it matters
The 5.5% is a flat tax on credit purchases. It compounds like a subscription, not a transaction fee. The thresholds I care about:
- Under $50/mo total. The 5.5% is small in absolute terms ($2.75) but the $0.80 minimum dominates if your purchases are frequent and small. Buy credits in bigger chunks if you stay on OpenRouter at this scale.
- $50 to $500/mo. The 5.5% is $2.75 to $27.50. Roughly the price of a coffee a week. Worth it if you actually use the routing layer.
- $500 to $5,000/mo. The 5.5% is $27.50 to $275/mo. Now it is a real line item. Direct keys plus a thin in-house router (30 lines of TypeScript) will pay back inside 2 weeks of saved fees.
- Over $5,000/mo. You should be on Anthropic Tier 4 or OpenAI's enterprise tier negotiated below the OpenRouter fee. At this scale the 5.5% is structural waste.
If you have ever priced the true cost of running an LLM workflow in 2026 on BudgetForge, you already know token cost is roughly 25 percent of the all-in number. Adding 5.5 percent on top of that 25 percent is a 1.4 percent tax on the all-in cost. Often worth it when the routing layer earns it back; never worth it when it does not.
How we routed a real $300/mo budget across these providers in 2026
When a client asked me to design a $300/mo AI app budget in May 2026, here is the split that actually worked across 30 days.
Scroll to see more
| Provider | Use | $/mo | % of budget |
|---|---|---|---|
| Anthropic direct | Reasoning + code (Sonnet 4.6) | $150 | 50% |
| OpenAI direct | Mini-model classification (GPT-5.4-mini) | $60 | 20% |
| Mistral direct | Structured extraction (Medium 3.5) | $50 | 17% |
| OpenRouter | Failover route for the 3 above + experimentation budget | $40 | 13% |
| Total | $300 | 100% |
We kept OpenRouter at the failover tier (13% of budget, ~$2.20 in routing tax) because that is where its value compounds. Reasoning and primary inference went direct. The deeper unit-economics math behind each line lives in our full cost-of-running-an-LLM-workflow calculator; drop your $/day token spend in and see the all-in number that includes retries, embeddings, tool calls, and platform margin.
The honest verdict
OpenRouter is well priced for what it is: a routing convenience layer with a 5.5% credit purchase fee, a $0.80 minimum, and pass-through provider rates. It does not save you money on token cost. It saves you operational overhead (key management, failover wiring, one bill instead of three).
If your workload is one model on one provider, you are paying 5.5% for nothing. Go direct.
If your workload is three or more models across providers, or if you need free fallback on retries, or if writing a custom router would cost you more in your effective hourly rate than the 5.5% costs you in fees, OpenRouter earns the tax.
The Cursor vs Claude Code 30-day bill teardown from June 20, 2026 makes the same point at the tool-subscription level: sticker price is never the all-in number. Token-aware routing is the lever; understanding when not to add a layer is the discipline.
Math check: the routing tax is exactly 5.5% above ~$15/mo and a $0.80 floor below. At $300/mo total spend, the tax is $16.50. Break-even is one hour of router maintenance time at your effective hourly rate.
Written by
Camille ForsterCamille Forster writes the freelance and small-studio economics column at BudgetForge. She has shipped pricing experiments for 60+ solo practices since 2022. Her work focuses on the math nobody publishes: effective hourly rate, real all-in tool costs, and the routing decisions that quietly compound.
Frequently asked questions
Is OpenRouter cheaper than going direct to Anthropic in 2026?
No. OpenRouter passes the underlying Anthropic per-token rate through to you and adds a 5.5% fee on credit purchases (with a $0.80 minimum). On a $186/mo Claude Sonnet 4.6 heavy code agent bill, the OpenRouter version is $196.23. The 5.5% is the routing tax. OpenRouter is cheaper than direct only when its routing layer (multi-provider, failover, single bill) saves more operational time than $10.23/mo of your effective hourly rate.
Does OpenRouter charge a per-token markup on Anthropic, OpenAI, or Mistral models?
No. Per the OpenRouter pricing page (June 2026), 'We pass through the pricing of the underlying providers without any markup, so you pay the same rate as you would directly with the provider.' The 5.5% fee is on credit purchases, not on individual token use. Pay-as-you-go also gets 1M free requests per month before a 5% per-request fee kicks in.
What is the OpenRouter minimum fee?
$0.80 per credit purchase. The 5.5% credit fee never goes below $0.80. For a hobby Haiku 4.5 chatbot with a $3.96/mo bill, the $0.80 minimum becomes a 20.2% effective tax. If your monthly bill is under ~$15, the minimum dominates. Buy credits in bigger, less frequent chunks to dilute the floor.
When should I use OpenRouter instead of direct provider keys?
Three real use cases in 2026: (1) you call three or more models across providers and want one API + one bill; (2) you need failover to a backup model and per the OpenRouter docs failed or fallback attempts do not incur charges when routing is enabled; (3) your high-RPS / low-token workload fits inside the 1M free requests/month threshold. For a single-model single-provider workload above $50/mo, direct keys are cheaper.
How does OpenRouter handle failed or fallback requests?
Per the OpenRouter pricing page, 'failed or fallback attempts' incur no charges when routing is enabled. So if your primary model fails and OpenRouter falls back to a backup, you only pay for the successful call. For agents that retry 3 times across providers, that math compounds in OpenRouter's favor and is the single strongest case for paying the 5.5%.
Is OpenRouter cheaper than going direct to OpenAI for GPT-5.4-mini?
No. GPT-5.4-mini is $0.75/MTok input and $4.50/MTok output direct (June 2026). On an 80M input / 1.2M output RAG workload, that is $62.70/mo direct vs $66.15/mo through OpenRouter, a $3.45 routing tax. The 5.5% is constant; the savings only show up when you actually use the routing layer for multi-provider calls.
Related reading
True cost of running an LLM workflow in 2026
The sticker price per token is the smallest line in your LLM bill. Once you add retries, embeddings, vector reads, orchestration, and the platform margin, a "simple" agent workflow lands around $0.42 per run, roughly 4× the raw model cost. This breakdown shows where the money actually goes and which three levers cut a workflow bill the fastest without touching quality.
Cursor vs Claude Code: the real 30-day bill (2026)
Sticker is $20 for both. At a 5-engineer team shipping 100 pull requests a month in June 2026, the all-in 30-day bill diverges. Here is the honest math.
Effective hourly rate 2026: the only freelance number that matters
Your billable rate is a sticker price. Your effective hourly rate is the truth. A 2026 cohort breakdown of billable vs effective across five real freelance brackets, plus the math.