GPT-5.2 vs Claude Opus 4.6 pricing comparison: what is actually live

Quick answer: GPT-5.2 is retired, Claude Opus 4.6 is not

Searches for "GPT-5.2 vs Claude Opus 4.6" usually assume both names are still active model IDs. Only one is. GPT-5.2 was OpenAI's flagship for a period after it shipped, but the GPT-5 line kept moving through point releases, and gpt-5.2 is not a model you can call on most OpenAI-compatible catalogs today, ours included. Claude Opus 4.6 stayed put: Anthropic still lists it at $5.00 per million input tokens and $25.00 per million output tokens, and it is carried in our catalog at $4.00 / $20.00, a plain 20% discount off that official rate. That asymmetry is the honest starting point here. If your code still targets gpt-5.2, the useful question is not which model is cheaper today, it is which current model replaces it. The sections below work through GPT-5.2's reported list pricing at the time next to Claude Opus 4.6's live numbers, a worked cost scenario for a typical workload, and the current flagship options that actually sit in an active catalog right now. None of that means the GPT-5.2 number is worthless. Teams that budgeted against it, or that are migrating an old integration off that model ID, still need an honest anchor to compare against, and pretending the comparison never existed would make this page less useful, not more accurate.

Where GPT-5.2 went, and how Claude Opus 4.6 is priced

OpenAI's GPT-5 line did not stop at 5.2. Point releases followed in sequence: gpt-5.3-codex-spark, gpt-5.4 and its mini variant, gpt-5.5, and the current gpt-5.6 family, with Sol, Terra, and Luna variants, each becoming the new default in the models list as the previous one aged out. That is normal for API-only model IDs: providers retire older point releases once traffic has migrated, and a name that was a live flagship not long ago can start returning a plain model_not_found error with no warning. GPT-5.2's reported list pricing at the time put it at $15.00 per million input tokens and $60.00 per million output tokens with a 128K context window; treat that figure as a historical reference, not a rate you can still buy at. Claude Opus 4.6 prices differently on the axis that actually moves a bill: the output rate. Anthropic's official $25.00 per million output tokens runs five times its $5.00 input rate, so a request's cost depends heavily on how much of the total is prompt versus generation. A short question with a long answer costs far more than a long prompt with a short answer, even at identical total token counts. Our catalog carries the same model at $4.00 input / $20.00 output, which keeps that same 5x input-to-output ratio at a lower absolute rate, with a 1M-token context window.

Worked example: pricing a real workload on both

Sticker rates are easy to compare; a realistic workload is more useful. Take a support-ticket triage service sending roughly 2,000 input tokens (ticket text, prior messages, a system prompt) and receiving about 500 output tokens (a categorized reply) per request, a fairly typical shape for a classification or summarization task. Pricing that per 1,000 requests, then scaling to a modest 50,000 requests a month, shows how far apart the numbers land once volume replaces eyeballing. The shape of the workload matters as much as the rate card. This scenario is output-heavy relative to its input, which is exactly where Claude Opus 4.6's 5x input-to-output ratio shows up in the total. Flip the ratio, a long document summarized into a short answer, and the gap between these three rows narrows considerably, because the cheaper side of each rate card carries more of the total.

Why the sticker price rarely matches the invoice

None of this shows up in a two-column price table, which is exactly why a workload-level comparison, not just the headline rate, is the number worth planning around.

• Output tokens carry the real weight. At a 5x input-to-output rate, a chatty completion costs far more than the input side suggests, so max_tokens and verbosity matter more than the headline per-token number.
• Retired model IDs fail silently until they do not. Code that hardcodes gpt-5.2 keeps working right up until the provider removes it from the catalog, and then every request returns a flat model_not_found with no notice.
• Context window size changes call count, not just price per token. A document that overflows a 128K window has to be split into multiple calls, and each split re-bills the overlapping input tokens used to stitch the pieces back together.
• Batch and caching programs are not universal. Anthropic, for one, documents a 50% discount on its Batch API for non-real-time jobs; whether a given model on a given provider offers a batch or cached-input discount, and at what rate, has to be checked per provider rather than assumed to carry over.
• Rate limits cap the bill from the other direction. A cheaper per-token rate does not help if your account's requests-per-minute ceiling throttles throughput before your budget does; tier and price are separate constraints.

Current model options, compared

GPT-5.2 is not a line item you can add to a live comparison table, so the more useful exercise is putting Claude Opus 4.6 next to the current models actually available today. All six rows below are live catalog entries with both a discounted and an official rate, spanning flagship reasoning down to high-volume budget work.

How to keep the bill predictable

None of these steps require changing your code beyond configuration; they are billing hygiene, not architecture work.

• Keep the model ID in a config value or environment variable, never a hardcoded string. That is what turns a provider retiring a model like gpt-5.2 into a one-line change instead of an incident.
• Route by task, not by habit. Reserve Claude Opus 4.6 for the reasoning, long-document, or accuracy-sensitive work it is actually priced for, and send routine drafts, extraction, and classification to a mid or budget-tier model.
• Set max_tokens deliberately on output-priced models. Given the 5x input-to-output ratio, an unbounded completion is where the budget actually breaks, not the prompt side.
• Pull real usage logs before trusting any percentage discount. Multiply your actual input/output token mix against the rate card; a workload that is mostly output behaves very differently from one that is mostly input, even on the same model.
• Route through a gateway priced below official list when the model you need is still live. APIsRouter carries claude-opus-4-6 at $4.00 input / $20.00 output per million tokens against Anthropic's own $5.00 / $25.00, pay-as-you-go with no subscription.

Calling Claude Opus 4.6 through an OpenAI-compatible endpoint

Anthropic's own SDK is not the only way in. Claude Opus 4.6 is reachable through the same OpenAI-compatible chat completions shape used for every other model in this catalog, so switching between it and any other carried model is a model string change, not a rewrite. Confirm the model ID is actually live on your key before wiring it into production, and while you are at it, confirm the old one is really gone:

curl https://api.apisrouter.com/v1/chat/completions \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-opus-4-6",
    "messages": [{"role": "user", "content": "Summarize this ticket in two sentences."}],
    "max_tokens": 200
  }'

FAQ