跪拜 Guibai
← All articles
Artificial Intelligence · LLM · ChatGPT

OpenAI Slashes Luna Pricing 80%, but Agent Costs Barely Budge

By 武子康 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

An 80% model price cut creates an illusion of proportional savings that collapses the moment tool calls, retries, and human review enter the ledger. Teams that migrate on token price alone will over-provision Luna, under-budget for Web Search and long-context step jumps, and discover the real cost per successful task barely moved.

Summary

OpenAI's July 2026 pricing update reduces Luna Standard input, cache read, cache write, and output rates to exactly 20% of their previous levels. For a homogeneous token bill, the math is clean: costs drop 80%. But agent workloads are never just tokens. Tool calls, loop retries, long-context step jumps at 272K tokens, cache mismatches, and human review all add independent costs that the model price cut leaves untouched. A task that once cost $0.20 might only fall to $0.12 — a 40% reduction, not 80%.

The real metric to track is cost per success, not cost per token. Web Search fixed fees now exceed the model token cost for a typical request, making tools the new cost center. The 272K context threshold triggers a discrete rate jump to 2× input and 1.5× output pricing, so average context figures hide dangerous tail behavior. Cache architecture matters too: writing a prefix that never gets read is pure waste at 1.25× the normal input rate.

A practical migration path runs 100–500 real tasks through a Luna-first routing tier with conditional escalation to stronger models. Low-uncertainty, auto-verifiable tasks go straight to Luna. High-stakes, hard-to-verify tasks stay on Terra or Sol with human review. The 30-day plan starts with baseline exports, moves through small-traffic A/B tests, and ends with hardened upgrade and rollback thresholds.

Takeaways
Luna Standard short-context rates dropped to exactly 20% of old prices: input $0.20, cache read $0.02, cache write $0.25, output $1.20 per million tokens.
The 80% equation holds only when model, service tier, context interval, and token composition all stay identical.
A typical agent task costing $0.20 (model $0.10, tools $0.03, human review $0.07) drops to $0.12 after the cut — a 40% total reduction, not 80%.
One Web Search fixed call costs $0.01, which already exceeds the $0.0064 model token cost of a 20K-input, 2K-output Luna request.
Input exceeding 272K tokens triggers a discrete rate jump: 2× input and 1.5× output pricing for the entire request.
Cache read costs 10% of normal input, but hits require stable prefixes; putting timestamps or dynamic state first breaks reuse.
Cache write costs 1.25× normal input and is wasted if the prefix is never read again.
Cost per success sums all model fees, tool fees, external services, and human review amortization, divided by successful tasks.
Migration A/B testing needs 100–500 real tasks covering low-risk, medium-uncertainty, high-error-cost, long-context, and historical tail failures.
A Luna-first routing strategy sends auto-verifiable, low-uncertainty tasks to Luna, escalates medium-uncertainty failures, and keeps high-stakes tasks on Terra or Sol with human review.
Small task volumes make multi-model routing engineering more expensive than the savings; a single strong model remains the simpler choice.
The 1.05M token context window is not free capacity — the 272K step jump and separate cache and tool billing make every token above the line cost more.
Conclusions

Token price cuts shift the cost bottleneck from the model to the tooling layer. When a single Web Search call costs more than the model inference for a typical request, agent architecture must optimize tool call frequency, not token count.

The 272K threshold is a billing cliff, not a gradual slope. A request at 273K tokens pays double input rates on the entire request, so teams that only monitor average context length will miss catastrophic tail costs.

Cache write pricing at 1.25× normal input inverts the usual intuition that caching is always cheaper. Writing a prefix that never gets a cache hit is strictly more expensive than sending it as uncached input.

Cost per success exposes a fundamental tension: cheaper models may require more retries, more tool calls, and more human review, erasing the per-token savings. The metric forces teams to optimize the whole pipeline, not just the model choice.

Benchmark scores are useless for migration decisions because they exclude tool loops, long-context step jumps, and error costs — the very factors that dominate real agent task economics.

Concepts & terms
cost_per_success
The total cost of all attempts (model fees, tool calls, external services, human review amortization) divided by the number of successfully completed tasks. It replaces per-token cost as the primary metric for evaluating agent pipeline economics.
272K threshold
A billing boundary in GPT-5.6 Luna where any request exceeding 272,000 input tokens triggers a discrete rate jump: the entire request is charged at 2× the standard input rate and 1.5× the standard output rate, regardless of how little it exceeds the threshold.
Luna-first routing
A conditional execution strategy where low-uncertainty, auto-verifiable tasks default to the cheaper Luna model, medium-uncertainty tasks run A/B tests with escalation on failure, and high-stakes tasks bypass Luna entirely for stronger models with human review.
cache write/read billing
OpenAI charges separately for writing to and reading from the prompt cache. Cache reads cost 10% of normal input, but cache writes cost 1.25× normal input. Writing is only cost-effective when the cached prefix is actually reused in subsequent requests.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗