OpenAI now tells developers that trimming repeated rules from system prompts raised internal coding evaluation scores by up to 15% while cutting token use by as much as 66%.
Key Points:
- The GPT-5.6 prompting guide asks developers to define the outcome and the stopping conditions, then let the model choose its own path.
- Internal coding-agent runs showed leaner system prompts improved evaluation scores by 10% to 15%, cut tokens by 41% to 66% and lowered cost by 33% to 67%.
- The document adds sections on Programmatic Tool Calling and a text.verbosity setting, neither of which appeared in the older GPT-5 playbook.
OpenAI Guide Rewrites The GPT-5 Prompting Playbook
OpenAI published the guidance alongside the GPT-5.6 model family, which reached general availability on Jul. 9, and aimed it squarely at API developers and teams running automated agents.
The document tells engineers to state the user-visible outcome, the constraints, the available evidence and the completion bar, then leave the model room to choose an efficient path on its own. It calls that approach outcome-first prompting.
The advice reverses much of the GPT-5 playbook from August 2025, which pushed XML persistence blocks, detailed context-gathering templates and tool preamble scripts that narrated each step out loud. Those rails now count as noise.
OpenAI also warns against absolute rules such as always and never, reserving them for true invariants like safety limits, required fields and actions that must never happen. Repeating instructions such as ask first or wait for approval can trigger needless approval requests for safe, expected actions, the guide says. Conflicting rules create more instability than missing detail.
Also Read: Dodgers Bets Hit $68M As Polymarket And Kalshi Turn To Baseball's Playoff Race
Simon Willison Weighs GPT-5.6 Tool Calling
Independent developer Simon Willison flagged Programmatic Tool Calling and multi-agent support as the most interesting additions in the release, noting the feature lets models compose and run JavaScript that orchestrates tool calls. He also said Sol felt competent without clearly beating Anthropic's Claude Fable 5 on the complex coding tasks he had been running.
Cost is the second reason the guidance matters. Token use fell 41% to 66% in the internal runs, and spending dropped 33% to 67%, figures that reshape the arithmetic for any team running agents at scale.
OpenAI cautions that results vary by workload, calls the ranges directional, and asks developers to validate them on representative tasks from their own applications.
GPT-5.6 Family Follows A Fast Release Cadence
GPT-5.6 shipped in three sizes. Luna, Terra and Sol carry input prices of $1, $2.50 and $5 per million tokens, with output priced at $6, $15 and $30, and the guide adds a text.verbosity setting to control response length.
The reversal follows a steady pattern. OpenAI's GPT-5.5 guidance in April already told teams to rebuild prompts rather than port old ones forward, while the GPT-5 guide from August 2025 leaned the other way and asked for explicit rails around eagerness. Each cycle has pushed developers to write less and trust the model with more of the route.
Read Next: Ethereum Gains On Bitcoin, Testing Tom Lee's 2026 Bull Case