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

An analysis of agent-system economics prices a workflow of 10,000 model calls a day at $7,500 a month when every call goes to a flagship, against $1,830 when the same volume is split across three tiers of one model family. The worked example, published on Github, assumes an average call of 2,000 input tokens and 500 output tokens.

At a glance

• Per-call pricing drives the gap: GPT-5.6 Sol costs $0.025 for the example call, GPT-5.6 Terra $0.010 and GPT-5.6 Luna $0.001, on short-context API rates at the time of writing.
• An intermediate split, 80% of calls to Terra and 20% to Sol, costs $130 a day or about $3,900 a month, a 48% reduction against the all-flagship arrangement.
• The author's criterion is to pick the least expensive model that clears the required quality and reliability threshold, escalating ambiguous, high-risk or long-horizon work to the stronger tiers.

Model selection reads less like a configuration default and more like an architectural decision once a single user request fans out into planners, workers, reviewers, retrievers and retries. At a fraction of a cent per call the difference is invisible; across 300,000 calls a month it becomes infrastructure spending. The saving depends entirely on the cheaper tiers already clearing the quality bar for the work they are handed, not on accepting weaker results.

Three allocations move the same 300,000 monthly calls from $7,500 to $1,830

The first scenario sends all 10,000 daily calls to GPT-5.6 Sol at $0.025 each: $250 a day and $7,500 over 30 days. The second routes 80% to GPT-5.6 Terra and keeps 20% on Sol, giving 8,000 Terra calls at $80 and 2,000 Sol calls at $50, or $130 a day.

The third allocation divides the work more finely: 60% routine, easily verifiable tasks to GPT-5.6 Luna, 30% moderately demanding tasks to Terra and 10% difficult or high-value tasks to Sol. That comes to $6, $30 and $25 respectively, $61 a day and roughly $1,830 a month, about $5,670 less than the all-flagship arrangement and 75.6% below it.

Anthropic positions Claude Fable 5 for ambitious coding projects and long-running agentic work

OpenAI highlights GPT-5.6 Sol's performance on coding and long-horizon engineering workflows involving planning, iteration and tool coordination, while Anthropic presents Claude Fable 5 as its most capable model for ambitious coding projects and long-running agentic work.

The source attributes that emphasis to structure and economics: code is highly structured, engineering work is economically valuable, and outputs can often be checked automatically with compilers, test suites, static analysis and CI systems. Dedicated coding agents have also become substantial products in their own right.

Only part of a typical agent workload is software. The same list covers classifying and routing tickets, extracting structured fields from unstructured text, summarizing meetings, translating short messages, updating CRM systems and issue trackers, comparing records against predefined criteria, and dispatching subtasks to specialized sub-agents.

Routine calls share bounded context, predictable output formats and automatic verification

The markers for routine work are properties rather than task names: the problem is well defined, the context is bounded, the required output format is predictable, mistakes can be detected automatically, and the output can be verified against deterministic rules or reviewed later by another agent or a human.

Failure in those cases is limited, recoverable or inexpensive, the task can be retried or escalated, and throughput matters more than small gains in reasoning quality. Running all of that on a flagship is compared to assigning the most senior engineer to rename files, sort support tickets, copy values between spreadsheets and check for missing fields.

What the price table leaves out Real bills move with prompt caching, context length, reasoning tokens, retries, batch processing, tool calls and provider discounts, none of which the example models. Prices are the standard short-context API rates at the time of writing and will change. The allocation percentages are not a recipe: a coding agent, a research assistant and a customer-service system carry different task distributions, and no fixed split is given for any of them.