Three GPT-5.6 tiers drop a $7,500 agent bill to $1,830

Run an agent system’s 300,000 monthly calls through GPT-5.6 Sol and the bill comes to $7,500. Split the same calls across Sol, Terra and Luna and it drops to $1,830, according to the worked example in a write-up on agent economics published on GitHub on August 19. That figure rests on the post’s Terra and Luna prices, which are lower than the launch rates Tech Journal reported. At Tech Journal’s prices, the same three-tier split comes to about $2,775 a month.
The post’s argument is that most calls in an agent workflow don’t need frontier reasoning, and in a multi-agent setup those calls add up quickly. One user request can fan out to planners, workers, reviewers, retrievers, classifiers and tool-using sub-agents, and retries and validation add more calls on top. “Once that starts happening, model cost stops looking like a per-request expense and starts looking like infrastructure,” the author writes.
The worked example
The scenario assumes 10,000 model calls a day, about 300,000 over a 30-day month. Each call averages 2,000 input tokens and 500 output tokens. To avoid comparing unrelated products from different vendors, the author uses the standard short-context API prices of three models in one family:
- GPT-5.6 Sol (flagship): $5.00 input and $30.00 output per million tokens, $0.025 per average call
- GPT-5.6 Terra (main): $2.00 and $12.00, $0.010 per call
- GPT-5.6 Luna (economy): $0.20 and $1.20, $0.001 per call
All on Sol, the system costs $250 a day. In the second scenario, 80% of calls go to Terra and only the 20% that justify the flagship stay on Sol. That comes to $130 a day, or $3,900 a month, 48% below the all-Sol bill. The third scenario sends 60% of calls (routine, predictable and easy to verify) to Luna, 30% (moderately demanding) to Terra and 10% (difficult or high-value) to Sol. That costs $61 a day, or $1,830 a month: 75.6% below the all-flagship setup and $5,670 cheaper each month.
The author is explicit about what the 75.6% doesn’t mean. It isn’t a general prediction, and it “certainly does not imply accepting 75.6% lower quality”. The assumption is that cheaper models only receive tasks where they already meet the required quality bar. The percentages aren’t a recipe either. A coding agent, a research assistant, a customer-service system and a financial-analysis workflow will each split their calls differently, and one system might need the flagship for half its calls while another needs it for a few percent. Prompt caching, context length, reasoning tokens, retries, batch processing, tool calls and provider discounts can all shift the real numbers considerably, the post adds.
Two price lists for Terra and Luna
The Sol figures are the same in both sources. Tech Journal, covering the July 9 public release of GPT-5.6, gave Sol’s API price as $5 input and $30 output per million tokens, unchanged from GPT-5.5. Its numbers for the other two tiers are higher than the post’s: Terra at $2.50 and $15, which it described as half of Sol’s price, and Luna at $1 and $6. The post gives $2 and $12 for Terra and $0.20 and $1.20 for Luna. Its prices are labelled “at the time of writing”, and it doesn’t explain the difference.
Under the post’s own token assumptions, the gap matters most for Luna, which carries 60% of traffic in the cheapest scenario. Re-run with the launch prices Tech Journal reported, the 80/20 Terra/Sol split comes to about $4,500 a month instead of $3,900. The 60/30/10 split comes to about $2,775 instead of $1,830. The direction doesn’t change, since both versions land well below the $7,500 all-Sol bill, but the size of the saving depends on whose Luna price you use.
Caching, one of the variables the post lists as moving the real bill, also changed with this release. Tech Journal reported that cache writes now bill at 1.25 times the normal input rate, while cached reads keep their 90% discount.
Which calls can drop a tier?
According to the post, work suited to cheaper models has recognisable traits. The problem is well defined, the context is bounded and the output format predictable. The task is extraction, transformation, classification, summarisation or routine execution. The output can be checked against deterministic rules or reviewed later by another agent or a human. Failures are limited, recoverable or cheap, and the task can be retried or escalated. The examples given are a support-ticket classifier, an invoice extractor, a meeting summariser, a short-message translator, a structured-record comparator and an agent calling a well-defined API. A frontier model might handle unusual edge cases somewhat better on these, the author says, but if a cheaper model reliably clears the bar, the extra quality has little practical value.
The flagship cases the post names are a major architectural decision, a subtle security review, an ambiguous legal document, a complex scientific problem and an important business decision. Its working rule is to choose the least expensive model that meets the required quality and reliability threshold, then escalate difficult, ambiguous or high-risk cases to more capable models.
The post also explains why coding benchmarks don’t settle the question. OpenAI promotes GPT-5.6 Sol for coding and long-horizon engineering workflows, and Anthropic positions Claude Fable 5 as its most capable model for ambitious coding projects and long-running agentic work. Most of the workload list in the post is unrelated to writing software: summarising documents, extracting fields, translating, updating business records, routing messages, checking rules. Strong coding performance, the author argues, “does not imply that every non-coding task benefits proportionally from the same level of capability.”
OpenAI’s own tier descriptions follow the same division. Per Tech Journal, Sol is meant for the hardest work, such as complex coding, long agentic tasks, research and security. Terra is pitched for everyday production work, with OpenAI calling it competitive with GPT-5.5 at half the cost. Luna is aimed at high-volume, speed-sensitive tasks such as summarising, drafting and routing.
Routing or cascading
The post doesn’t say how to implement escalation, and there are two common approaches. Routing classifies each request up front and runs a single model, so latency stays predictable, but the router’s accuracy limits the quality of the whole system. Cascading runs the cheap model first, judges the output and escalates only when the answer falls short. That lowers average cost, but in the worst case a request goes through several models plus the judge. Nepex Group warns that an expensive classifier wipes out the savings. It recommends a fine-tuned small model, a few embedding comparisons or plain heuristics, and estimates that 70-80% of requests in most applications are simple enough for a small model.
Independent figures for routing in practice cover a range that includes the post’s results. Digital Applied says teams with a tuned routing layer report bill reductions of 40-85%. It says an 80% cheap, 20% frontier split has been reported to cut cost per query toward 72%, and that the peer-reviewed RouteLLM work reached 85% cost savings in controlled evals while keeping 95% of GPT-4’s quality.
The author frames the case as one about defaults, not against flagship models: “You do not need weaker models everywhere. You just need to stop treating the strongest model as the default.”
Up next
Forced rm detection tightened in Codex
A small number of user reports sent OpenAI’s Codex team hunting for forced rm commands. Detection now catches them in pipelines, traps, nested shells and behind sudo or env wrappers, per the team’s GitHub post.
Related stories
- A day of OpenAI’s dot cost 307,000 tokens, all unprompted
- OpenAI drops lottery for free ChatGPT access for scientists
- ChatGPT users in the EU report getting OpenAI’s dots
- Grok Bot paid plans get 7% more usage
- US Pro users get ChatGPT Wallet for agent payments
- OpenAI’s dots can now read email codes to pass 2FA
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

