GPT-5.6: The cost and latency worksheet

Rank GPT-5.6 by verified output per dollar and per second, not by the number at the top of the pricing page, and skip the marketing. That's the whole rule. The rest of this piece just earns it.

OpenAI documents three tiers: Sol, Terra, and Luna, each with a 1.05M context window and its own price-performance slot, running frontier reasoning through balanced work to cost-sensitive volume. Test gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna directly, and check the docs before shipping: aliases and prices move faster than blog posts.

None of that tells you whether the model is worth routing traffic to. What does is deciding, before the first prompt, what counts as an accepted result: a schema-valid record for extraction, a patch that clears tests and review for coding, a sourced claim set for research. Fluent prose isn't on that list.

the bill hides in retries and cache misses

Once you know what a win looks like, log every request: cached and uncached input, reasoning tokens, visible output, tool charges, retries, p50/p95 latency, and human repair by workload. Keep the whole request: instructions, tool schemas, files, sampling settings, output limits. A model looks brilliant or useless depending on whether the harness gives it native tools or makes it parse pasted terminal output, so hold the workflow constant.

model id + dated api version
prompt and tool-schema revision
input / cached input / reasoning / output tokens
queue time, first token, total completion
pass, retry, abstain, human repair
provider error, fallback route taken

The expensive mistake is comparing input-token list prices while output tokens, cache writes, tool calls, and failed attempts run up the real bill. Frontier output is hard to catch because it almost always reads fine. Measure what code can check, have a reviewer tag concrete defects, and keep the wrong answers on purpose: failures teach you more about routing than another folder of clean transcripts.

Long context is the part people trust most and should trust least. A bigger window doesn't shrink prefill time, filter retrieval noise, or decide which text is instruction versus data someone handed you. Put stable material first if the API caches prefixes, keep volatile metadata off the cache boundary, and send a structured checkpoint instead of full history when possible. Keep output short: a decision and its evidence beat a model that decides to teach you something.

Tool use changes the risk profile in a way the vendor can't fix. GPT-5.6 might pick the right function more often, but your application still validates arguments, checks who's allowed to call it, caps the blast radius, and logs what happened. Web pages, files, emails, and tool output are data, not instructions. A sharper model isn't a replacement for a narrower interface.

turn the rule into a config file, not a preference

The durable version of the rule belongs in configuration, not a Slack message about which model is good this week. Log model identity, phase timings, token categories, tool outcomes, and pass/fail results, skipping sensitive prompt content without a reason.

Keep a cheap, fast lane for routine traffic and a clear trigger for escalating to Sol. The frontier tier should see requests that benefit from more reasoning, not everything the gateway can send it: the same instinct behind turning the effort dial down for easy calls. The reverse holds: a bargain tier stops being a bargain once retries and review outrun what it saved.

Treat every version bump as normal operations, not a one-time upgrade. Pin dated versions where stability matters, watch deprecation notices, canary a new alias before it touches all traffic, and keep a known-good fallback wired in, the same posture model migration reality argues for elsewhere. Re-run your evaluation set whenever the model, prompt, tool contract, or reasoning default changes: any one can move the numbers alone.

All of this assumes you have the volume to make the measurement worth doing. Send a few hundred calls a day and the spreadsheet costs more time than it saves. Picking Terra and moving on is probably the sensible call.