LLM API pricing comparison 2026: 12 models, two gateways

This is a like-for-like snapshot of what twelve current models cost on two OpenAI-compatible gateways, JZS Token and OpenRouter. It is not a list of every model either platform sells, and it is not a comparison against each vendor's first-party list price. It is the number a team actually pays when it sends a request to one of these two gateways.

Captured 2026-10-26. USD per 1M tokens, standard non-batch routes. A 3:1 input/output blend is used for the ranking; the raw input and output rates are in the table.

Methodology (so you can argue with it)

• Same model ID on both sides. Each row uses an exact ID both platforms list on the capture date. No family-name matching, no “closest equivalent”.
• Standard routes only. OpenRouter batch prices are excluded. Batch is often 50% cheaper and deserves its own comparison.
• No cache, no retries, no free tier. The blended number is a list-price calculation, not a bill. Cache hit rates and retry rates change the result.
• 3:1 input/output. This matches coding assistants and document Q&A better than 1:1. If your workload is output-heavy, the expensive column matters more than the blended rank.

How the blended number is calculated

The blended price is (3 × input + output) / 4. That is the cost of one million tokens where three quarters are input and one quarter are output. It is a single number for ranking, not a replacement for the two raw rates.

Worked example for GPT-5.5 on JZS: (3 × $0.3892 + $2.3353) / 4 = $0.876 per 1M. On OpenRouter: (3 × $5.00 + $30.00) / 4 = $11.25 per 1M. Same formula, same weights, same model ID.

• Input-heavy (9:1). Long documents and RAG pipelines land here. The input column dominates; a gateway with a large input advantage can win even if its output rate is higher.
• Balanced (1:1). Some chat and agent workloads. The simple average of input and output is closer than the 3:1 blend.
• Output-heavy (1:3). Generation, summaries, and reasoning traces. The output column dominates, which is where premium models get expensive fast.

Pull your own ratio from a week of usage logs before you use this table to make a decision. The ranking is stable for most coding workloads, but an output-heavy workload can move a row by several places.

The table, sorted by JZS blended price

USD per 1M tokens, standard (non-batch) routes, captured 2026-10-26. Blended assumes 3 input tokens per output token. OpenRouter batch routes are often half price and are not compared here.

Sources: JZS public pricing API and OpenRouter public model list, both read on 2026-10-26. Prices move; check both pages before you commit to a provider.

What stands out

• The budget tier is very cheap. GPT-5.6 Luna and GPT-6 Luna land around $0.16–$0.18 per 1M blended. They are appropriate for routing, extraction, and high-volume simple chat.
• The biggest relative gaps are on GPT-5.5, GPT-6 Sol, and Gemini 4 Argon. At the captured rates, JZS is roughly 81–92% below OpenRouter on those rows.
• OpenRouter wins three rows. DeepSeek V4.1 Flash, Qwen3.8 Flash, and GLM-5.3 Flash are cheaper on OpenRouter at a 3:1 mix. That is in the table on purpose.
• Claude Opus is the premium row. It is still about 70% cheaper on JZS than on OpenRouter, but it is the most expensive model here by a wide margin. Reserve it for tasks where a wrong answer costs more than the tokens.

The gap between the cheapest and most expensive rows is more than 10×. That is why “which gateway is cheaper” is a less useful question than “which model do I actually need for this task”. A team that moves classification from Opus to Luna can save more than a gateway switch.

Official price vs gateway price

A vendor's official price list and a gateway's price for the same model are different products. The official list is what the vendor charges for its own API, with its own rate card, priority tiers, batch discounts, and caching rules. A gateway publishes a single rate for a model ID and routes the request to an upstream channel it operates or contracts.

That means a gateway can be cheaper than the official list on one model and more expensive on another, and neither direction is evidence of anything by itself. It depends on the upstream deal, the routing mix, and the gateway's margin. The number that matters is the delivered cost for your workload, including retries and cache hits, not the spread between two published rates.

This page compares two gateways because that is the choice most readers are actually making. If you are deciding between a gateway and a direct vendor account, add the direct account's rate limits, billing minimums, and quota process to the comparison; those often outweigh a per-token difference.

Provider notes

OpenAI

GPT-5.5 is the clearest example of the gap: $0.876 blended on JZS against $11.25 on OpenRouter, a 92% difference. The Luna models are cheap on both sides, but JZS is still below OpenRouter at the captured rates. One mapping caveat: GPT-4o is not in the JZS catalog, so a workload pinned to that ID needs a different route or a model change before migrating.

Anthropic

Claude Sonnet 5.5 is $1.557 blended on JZS versus $4.00 on OpenRouter. Claude Opus 5.5 is $2.335 versus $8.00. The same model IDs are used by Claude Code when it points at an Anthropic-compatible base URL, which is why the Claude Code guide uses the Sonnet row for its cost example.

DeepSeek

DeepSeek V4.1 Flash is the row where OpenRouter wins at a 3:1 mix: $0.265 versus $0.299 on JZS. The gap is small in absolute terms. DeepSeek V4 Pro goes the other way: $0.898 blended on JZS versus $1.194 on OpenRouter. Check the exact ID; the two DeepSeek tiers do not rank the same way.

Qwen

Qwen3.8 Flash is cheaper on OpenRouter at $0.230 versus $0.311. Qwen3.7 Max is cheaper on JZS at $1.168 versus $2.213. This is the clearest illustration that “the Qwen row” is not one decision; the tier matters more than the family name.

Google

Gemini 4 Argon is $0.778 blended on JZS against $4.00 on OpenRouter. It is also the row with the most important non-price caveat: upstream availability has been intermittent, so test it before putting production traffic on it. A cheap model that returns errors is not cheap.

Z.ai

GLM-5.3 Flash is cheaper on OpenRouter at $0.238 versus $0.311 on JZS. The absolute difference is small at low volume; at hundreds of millions of tokens it is worth splitting the workload.

Who wins by workload

Where the largest gaps are

The three rows with the largest JZS advantage at the captured prices are:

• GPT-5.5 — 92.2% cheaper on JZS ($0.876 vs $11.250 per 1M blended).
• GPT-6 Sol — 82.5% cheaper on JZS ($0.701 vs $4.000 per 1M blended).
• Gemini 4 Argon — 80.5% cheaper on JZS ($0.778 vs $4.000 per 1M blended).

A large percentage on an expensive model is a large absolute number. A large percentage on a cheap model is pocket change. Multiply by your actual token volume before you let a percentage drive a migration.

What this comparison does not include

• OpenRouter batch routes (often half price) and free endpoints.
• Provider-direct prices, enterprise discounts, committed-use deals, and regional taxes.
• Cache hit rates: cached input is priced separately on both gateways.
• Retries, failed requests, and output-length differences between models.
• Latency and rate limits, which can matter more than a few percent of price.

If you need any of those, compare them directly. A pricing table is a starting point, not a procurement decision.

How to use this in a real migration

• Pick the exact model IDs you already call; map them to the JZS IDs in the migration guide.
• Calculate your own input/output ratio from a representative week of usage.
• Run the formula on both gateways, then run 1% of traffic through the new one for a few days.
• Compare cost per successful request, not just the per-token list price.

The model-ID mapping and a safe dual-run setup are in How to migrate from OpenRouter. For the 100M-token savings math, see Cheaper than OpenRouter. For the cheapest options by category, see Cheapest LLM API.

FAQ

Why not compare the vendors' official prices?

Because a team buying through a gateway pays the gateway's price, not the vendor's list price. We compare the two gateways on the same model ID. If you buy directly from a vendor, use that vendor's current price list instead.

Why are some rows cheaper on OpenRouter?

Routing, provider mix, and each gateway's own margin differ by model. There is no rule that one gateway is cheaper on every ID. The table shows the actual result rather than hiding the rows where we lose.

How do I estimate my own bill from this table?

Export a representative week of usage: input tokens, output tokens, retries, and the model ID for each request. Multiply input tokens by the input rate and output tokens by the output rate, add the retries, and divide by the number of successful tasks. Do that for both gateways using the same week of traffic. The result is your delivered cost per task — the only number that should decide a migration.

How often do prices change?

Both gateways update prices as upstream costs move. Treat this as a dated snapshot and check the live pricing pages before committing. The page will be updated when the snapshot is refreshed.

Compare models on the live catalog

The catalog shows input, output, and cache prices per 1M tokens for every JZS model. Start with the IDs you already use.