Disclosure: independent site. Links to AgentRouter are referral links — we may earn a reward if you sign up, at no extra cost to you. We do not publish pricing or promo numbers we have not verified.
Free Coding AI Lab

Comparing your options

You will not find a price table on this page. Provider pricing, promotional credits, and model availability change faster than any table stays honest, and a stale number costs you real money. What follows is the method for checking any endpoint yourself — it takes about ten minutes and stays correct.

Run the Setup Doctor →Error fixes

Why there's no table here

Comparison tables for model providers are the most confidently wrong content on the internet. They are usually assembled once, ranked well, and then never corrected — so they keep recommending a promotional rate that ended months ago or a model slug that has since been retired.

We would rather teach you a check that stays true. Run the sequence below against any endpoint you are considering, including one we link to. If a provider fails these checks, that tells you more than any table would.

The five checks that actually matter

In order. Each one eliminates providers cheaply before you spend time on the next.

  1. Does it serve the model you specifically need, right now? Not the family — the exact id. Request the model list and look for it. If it isn't there, nothing else about the provider matters for your use case.
  2. Is it compatible with the tool you already use? An OpenAI-compatible endpoint means base URL plus model id and you are done. Anything requiring a custom adapter has a hidden cost in your time that dwarfs the price difference.
  3. What is the real per-request cost at your context size? Headline per-million-token rates are close to meaningless for coding agents, which resend large contexts on every turn. Send one representative request and read the actual usage figures back.
  4. How does it behave at your rate? Send a realistic burst rather than a single request. Latency under concurrency and whether limits are per-minute or per-day matter more than the sticker price for interactive coding work.
  5. What happens when it fails? Check whether errors are specific and actionable. A provider that returns a bare 401 for six different causes will cost you hours — that is a real cost, just not one on the pricing page.

Checks 1 and 3 in one pass — run this against any candidate endpoint

# 1. Does it serve the exact model id you need?
curl -s "$BASE_URL/v1/models" -H "Authorization: Bearer $TOKEN" \
  | python3 -c "import sys,json;[print(m['id']) for m in json.load(sys.stdin).get('data',[])]" \
  | grep -i "$MODEL_SUBSTRING"

# 3. What does one realistic request actually consume?
curl -s "$BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer $TOKEN" -H 'Content-Type: application/json' \
  -d '{"model":"PASTE_EXACT_ID","messages":[{"role":"user","content":"Refactor this function"}],"max_tokens":200}' \
  | python3 -c "import sys,json;d=json.load(sys.stdin);print('served by:',d.get('model'));print('usage:',d.get('usage'))"

# Multiply the usage figures by the provider's published rates yourself.
# That number is your real cost. The headline rate is not.
Costs that never appear on a pricing page
  • Output tokens are typically billed several times higher than input tokens, and coding agents produce a lot of output.
  • Context resend dominates long sessions. The same task costs wildly different amounts depending on session hygiene, which no comparison table can capture.
  • Auxiliary calls — summarizing, titling, compression — are invisible in the chat and easy to leave pointed at an expensive model.
  • Retries on transient failures are billed. A flaky endpoint costs more than its rate suggests.
  • Your own debugging time. A provider with vague error messages and thin docs is expensive in the currency you cannot buy more of.

Before you switch anything

Switching providers is the most common answer to a problem that a free configuration change would have solved. Before you evaluate anything, confirm you actually have a capacity problem rather than a hygiene problem — the usage-limits page covers the free steps, and they compound in a way that a cheaper rate does not.

If you have done those and still need more capacity, then the checks above are the right way to choose. Run them against every candidate, including any provider this site links to.

Running the five checks against AgentRouter

Here is how it scores on the checks above, so you can skip straight to the ones that matter to you. Check 2 (tool compatibility): it is OpenAI-compatible and documents setup for 15 tools, so no adapter. Check 1 (exact model): the default group serves gpt-5.6-sol, claude-opus-4-8 and claude-opus-5, published openly. Check 3 (real cost): claude-opus-5 is listed at $2.000 per 1M prompt tokens and $10.000 per 1M completion tokens, pay-as-you-go, so you can multiply it against your own usage figures. Checks 4 and 5 — behaviour under your real load, and whether its errors are actionable — you should run yourself with the commands above, because we cannot measure your workload for you.

Run the checks against it →

Referral link — we may earn a reward if you sign up, at no extra cost to you. This is the same advice we would give with no link at all. Check their current pricing and model list on their own site; we deliberately do not restate numbers that change.

Note what that does and does not claim: those are the figures its pricing page published on 20 August 2026, not a promise that it is cheapest for you. Rates change and your workload is not ours — which is why checks 3, 4 and 5 are commands you run rather than numbers we print. Re-run them the day you decide.