If your engineers are calling "gpt-5.2" and you don't know which SKU suffix follows the dot, you have a procurement problem hiding in a config string. As of this week, OpenAI ships four distinct gpt-5.2 SKUs, all sharing quality score 96, with output prices spanning $14/M to $168/M — a 12× delta inside one model family at identical quality.
Last month we wrote about the 720× spread across OpenAI's full catalog. This is the sharper, more actionable cut: the spread inside the family your team is probably already calling.
chat $14/M → codex $40/M → pro $168/M. Same quality. 12× delta. The full SKU string is the contract; the model name is not.Live snapshot from CostGoat's LLM pricing tracker, May 8, 2026:
| SKU | Input ($/M) | Output ($/M) | Quality | Value Score |
|---|---|---|---|---|
gpt-5.2-chat | $1.75 | $14.00 | 96 | 6.9 |
gpt-5.2 (base) | $1.75 | $14.00 | 96 | 6.9 |
gpt-5.2-codex | $5.00 | $40.00 | 96 | 2.4 |
gpt-5.2-pro | $21.00 | $168.00 | 96 | 0.6 |
gpt-5.2-chat ($14/M) and gpt-5.2-pro ($168/M). Same model family. Same quality score. The full SKU string is the contract; the model name isn't.Read the columns carefully. Same quality score across all four. Cost-per-completed-task differs by an order of magnitude. The value scores tell the same story from the other direction: gpt-5.2-chat returns 6.9 quality points per dollar; gpt-5.2-pro returns 0.6.
For comparison, claude-opus-4.6 at quality 100 (the highest score in the catalog) costs $25/M output. gpt-5.2-pro is 6.7× more expensive than Opus at lower quality.
"Quality 96 across all four" deserves a beat of skepticism. Quality scores are normalized benchmarks — they're a useful single-number proxy, not a complete characterization of model behavior. Two models scoring 96 on a benchmark are not interchangeable for every workload.
What "same quality" does mean in practice:
gpt-5.2-codex is tuned for code; gpt-5.2-pro is tuned for higher-stakes reasoning chains where a single wrong inference is expensive.gpt-5.2-pro at $168/M output. None.That last category is where the 12× delta lives. If you're paying it without an explicit quality argument, you're paying a SKU-suffix tax.
gpt-5.2-pro? If "we don't track at SKU granularity," that's the first project. If it's 20%+ for non-reasoning workloads, that's the routing change.A finance leader can ask one question this week and get an immediate answer:
"For every API key billing OpenAI traffic, which exact SKU string is being called, and what fraction of our output tokens are running through
gpt-5.2-pro?"
If the answer is "we don't track at SKU granularity," that's the first project. SKU-level routing visibility is now table-stakes for any AI cost stack — without it, the 12× delta is invisible until the bill arrives.
If the answer comes back with gpt-5.2-pro consuming 20%+ of output tokens for non-reasoning workloads (anything that isn't deep multi-turn agentic reasoning), the second project is the routing change. Map the workload back to whichever team owns it. Have them justify the Pro tier with a quality-difference benchmark. If they can't, route them to gpt-5.2-chat and watch the bill.
The cross-vendor spread is the more famous number — 442× from gpt-5.2-pro ($168/M) to deepseek-v3.2 ($0.38/M) (live pricing snapshot). It's the spread that gets quoted in board decks and analyst reports.
The intra-vendor spread is the easier spread to capture, for two reasons:
1. No vendor migration friction. Switching from OpenAI to DeepSeek means new API contracts, new billing relationships, new compliance reviews, new failover handling. Switching from gpt-5.2-pro to gpt-5.2-chat is a one-character config change.
2. No quality argument to defend. A switch from a frontier model to an open-weight challenger requires defending "quality 79 is enough for this workload." A switch from gpt-5.2-pro to gpt-5.2-chat requires defending nothing — the quality score is identical. The procurement conversation goes from "are we taking quality risk?" to "are we paying 12× more for nothing?"
The intra-vendor spread is the one that converts in the next 30 days. The cross-vendor spread is the multi-quarter project.
gpt-5.2) tells you generation; the suffix is the only reliable pricing signal. SKU resolution belongs at the gateway, not in application code — it's the only way to maintain cost discipline across teams.A general lesson from the gpt-5.2 lineup: the SKU suffix is now the only reliable pricing signal. The model family name (gpt-5.2) tells you generation; everything that determines cost — chat vs codex vs pro — lives in the suffix.
This is not a unique OpenAI quirk. Anthropic ships Sonnet/Opus/Haiku at a similar pattern. Google ships Flash/Pro/Ultra. The vendors are all converging on multi-SKU families at one generation level, with intra-family price spreads larger than the inter-vendor spread used to be.
For procurement, this means:
gpt-5.2-pro traffic that isn't deep agentic reasoning, and move SKU resolution from app code to the gateway. The audit is a 30-minute conversation; the routing change is a config-line edit.Three actions for finance, FinOps, and AI program leaders:
gpt-5.2-pro that isn't deep multi-turn reasoning. Pro is justifiable for the workloads it's tuned for. It's not justifiable as a default. The audit is a 30-minute conversation with whichever team owns the calling code.-mini to -pro and grows the bill 30× overnight.OpenAI's gpt-5.2 family ships four SKUs at identical quality with a 12× output cost spread. That spread is invisible from the model name alone — it lives entirely in the SKU suffix. A finance leader can get the audit number this week, and the routing change is one config-line edit, not a vendor migration. The 12× is the easiest cost discipline available in 2026 because it's hiding inside a contract you've already signed.
Trimio is the LLM API gateway built for AI cost governance — including SKU-level routing visibility and per-workload SKU enforcement, so the pricing signal in the model name is matched by the cost discipline at the call site. See how it works.