Model routing is now risk routing

OpenAI, the AI lab behind ChatGPT, made GPT-5.6 generally available on July 9 across ChatGPT, Codex, and the API. The release is a three-tier model family: Sol is the flagship model, Terra is the balanced tier, and Luna is the cheaper high-throughput tier. API pricing is $5/$30 per million input/output tokens for Sol, $2.50/$15 for Terra, and $1/$6 for Luna. 1

For PMs, GPT-5.6 is a routing problem more than a single model upgrade. The same release added a deeper max reasoning setting, an Ultra mode with four parallel agents by default, Programmatic Tool Calling for JavaScript-based tool orchestration, a multi-agent beta, and explicit cache breakpoints with a 30-minute minimum cache lifetime. 1 The product question is now: which model should own each step, and what happens if access changes?

GPT-5.6 Sol looks strongest where the task resembles agent work: using tools, browsing, executing terminal tasks, and carrying state across a workflow. Artificial Analysis scored Sol at 80 on the Coding Agent Index, ahead of Anthropic's Claude Fable 5 at 77.2, and OpenAI reported Sol at 53.6 on Agents' Last Exam versus Claude Fable 5 at 40.5 with adaptive reasoning. 2 1

The split is not cleanly "OpenAI wins." Claude Fable 5 scored 80% on SWE-Bench Pro, while GPT-5.6 Sol scored 64.6%; Simon Willison noted that OpenAI had questioned the benchmark's methodology the day before launch, including a claim that about 30% of SWE-Bench Pro tasks were broken. 3 SWE-Bench Pro is closer to a software-engineering repair benchmark than a generic chat benchmark.

The PM read is that Sol fits collaborative, tool-heavy execution, while Fable still has a case for delegated engineering judgment. Every's month-long test described Sol as better with a human in the loop and Fable as stronger for fully delegated tasks. 4

A PM should treat GPT-5.6 as a routing menu, not a default switch.

The implementation consequence is simple: bind features to task classes, not model names. A code-review product can route broad planning to Fable, repo search and checklist execution to Sol, and cheap triage to Luna or Terra. CodeRabbit's benchmark found Sol at a 63.7% pass rate across more than 100 coding tasks versus Terra at 40.7%, while Sol also produced fewer output tokens per run than Terra in that test. 6

The second routing variable is access. Executive Order 14409, signed on June 2, created a voluntary pre-release access framework for covered frontier models and said developers may provide the federal government up to 30 days of early access before public release. Section 3(c) also says the order should not be read to create mandatory government licensing, preclearance, or permitting for AI model development or release. 7

GPT-5.6 still went through a real gating sequence. OpenAI previewed the model on June 26, limited access to about 20 government-screened organizations, worked with the Commerce Department's Center for AI Standards and Innovation, and released the family broadly on July 9 after roughly 12 days of review. 8 Politico reported that the White House said it did not give OpenAI a "green light," approval, or clearance. 9

That distinction is legally important and operationally messy. A PM does not need to claim a formal licensing regime exists to feel the product risk. If broad access can depend on review timing, launch dates, enterprise commitments, and customer migrations need a buffer.

For a team building on frontier models, the minimum viable architecture is a router with explicit task contracts. Each AI feature should declare task type, quality bar, latency target, cost ceiling, data sensitivity, fallback model, and rollback behavior. That layer should sit above vendor APIs so the product can swap Sol, Terra, Luna, Fable, or an open-weight alternative without rewriting user-facing flows.

The vendor review should change too. Lewis Carhart, chief executive of Comp AI, told InfoWorld that "multi-model resilience" had moved from nice-to-have to a board-level risk item, and Rock Lambros of Zenity advised teams to treat model availability like a single point of failure they do not own. 10 That maps to three PM requirements: tested fallback paths, continuity clauses in vendor contracts, and a vendor-risk line item for the model's regulatory posture.

Release planning needs the same adjustment. EO 14409 set an August 1 deadline for the National Security Agency to develop classified benchmark standards and for a multi-agency group to publish formal rules for the voluntary framework. 7 If those rules produce known timelines and criteria, PMs can plan around them. If the process stays ad hoc, the safer assumption is a 2-4 week review buffer for major features that depend on a newly released frontier model. 11

The next product spec should have one extra table: feature, primary model, fallback model, degradation mode, review risk, and owner. Without that table, the team is still treating frontier AI as a vendor capability. GPT-5.6 makes it look more like infrastructure with a policy-dependent release path.