gpt-5.6-luna vs gpt-5.6-terra
Fourteen priced metrics for the same 2,000-in / 300-out call, with and without provider caching, and what TokenZero's cascade would do with the pair.
At list prices as of 2026-10-26, gpt-5.6-luna costs $0.33 per million blended tokens on a 2,000-in / 300-out call against $3.30 for gpt-5.6-terra, 90% less. With the static prefix cached both drop further: $0.49 versus $4.90 per thousand calls. TokenZero routes only prompts its classifier scores simple and holds a 99.8% parity floor.
· Source: provider list prices verified in pricing table 2026-10-26; costs computed by the same code that signs TokenZero receipts
At a glance
- Cheaper per blended token
- gpt-5.6-luna by 90% (computed)
- gpt-5.6-luna blended $/1M
- $0.33 (computed)
- gpt-5.6-terra blended $/1M
- $3.30 (computed)
- Input $/1M
- $0.20 vs $2.00
- Output $/1M
- $1.20 vs $12.00
- Cached input $/1M
- $0.020 vs $0.200
- Cache discount
- 90% vs 90% (computed)
- Monthly at 100k calls
- $76 vs $760 (computed)
- Saved per 1M calls
- $6,840 (computed)
- Providers
- openai vs openai
- Lifecycle
- active, no retirement announced vs active, no retirement announced
- Parity floor for routing
- 99.8% (shadow-eval breaker)
gpt-5.6-luna vs gpt-5.6-terra: list prices and the same workload on both (USD, table 2026-10-26)
| Metric | gpt-5.6-luna | gpt-5.6-terra | Delta |
|---|---|---|---|
| Provider · tier | openai · fast | openai · mid | same provider |
| Input $/1M | $0.20 | $2.00 | 90% gap |
| Output $/1M | $1.20 | $12.00 | 90% gap |
| Cached input $/1M | $0.020 | $0.200 | 90% / 90% off input |
| Cache write $/1M | $0.250 | $2.500 | |
| Minimum cacheable prefix | 1024 tokens | 1024 tokens | |
| Effective blended $/1M (2,000 in / 300 out) | $0.33 | $3.30 | gpt-5.6-luna 90% less |
| Per call | $0.00076 | $0.00760 | $0.00684 |
| Per month at 100k calls | $76 | $760 | $684 |
| …with the 1,500-token prefix cached | $49 | $490 | $441 |
| Long-context threshold | 272,000 tokens ($0.40 in / $1.80 out above) | 272,000 tokens ($4.00 in / $18.00 out above) | |
| Processing tiers | flex 0.5x · fast 2x | flex 0.5x · fast 2x | |
| Lifecycle | active, no retirement announced | active, no retirement announced | |
| Cheaper sibling (cascade target) | — | gpt-5.6-luna | |
| TokenZero cascade recommendation | gpt-5.6-terra → gpt-5.6-luna | $6,840 saved per 1M calls if every call moved; $2,736 at a 40% simple-prompt share | 99.8% parity floor, breaker on |
Use either model through TokenZero
OPENAI_BASE_URL=https://tokenzero.sh/v1
# keep your own keys; cache formatting on, routing advisory, signed receipt per callFree up to $500/month of routed spend; paid plans take 15% of verified savings with a 3x-ROI-or-$0 rule.
What is not on this page
Third-party quality scores (coding ELO, extraction accuracy, reasoning benchmarks) and per-model latency are not shown because TokenZero does not measure them; the numbers above are prices and the arithmetic on them. The latency benchmark shows what the gateway does measure.
Questions people ask
Which is cheaper, gpt-5.6-luna or gpt-5.6-terra?
gpt-5.6-luna. At list prices as of 2026-10-26 it costs $0.33 per million blended tokens versus $3.30 for gpt-5.6-terra, a 90% gap on a 2,000-input / 300-output call. Per month at 100,000 calls that is $76 versus $760.
How much does prompt caching change the gpt-5.6-luna vs gpt-5.6-terra gap?
Cached input costs $0.020 on gpt-5.6-luna and $0.200 on gpt-5.6-terra per million. With a 1,500-token static prefix cached, 100,000 calls cost $49 and $490 respectively; TokenZero's formatter places the breakpoints and sets prompt_cache_key automatically.
Can TokenZero route gpt-5.6-terra traffic to gpt-5.6-luna safely?
Yes, within openai: with X-TokenZero-Cascade: auto the classifier scores each prompt and only simple ones move; shadow evaluation re-runs a sample on gpt-5.6-terra and a pair below 99.8% parity is auto-disabled. Moving every call would save $6,840 per million calls; at a 40% simple share it is $2,736.
Is gpt-5.6-luna or gpt-5.6-terra being retired?
gpt-5.6-luna: active, no retirement announced. gpt-5.6-terra: active, no retirement announced. TokenZero audits flag retired ids and the gateway warns in X-TokenZero-Warning with the provider's replacement.