GPT-6 Sol and LunaMore useful work. Half the token price.

The names describe durable tiers; the number describes the generation. GPT-6 Sol and Luna bring Astra-era training, factuality, collaboration style, and caching improvements to the places previously occupied by GPT-5.6 Sol and Luna. The most visible change is price.

The lower rate is not the entire story. OpenAI reports that GPT-6 Sol makes roughly half as many factual errors as GPT-5.6 Sol in its flagged-error evaluation. At higher effort, GPT-6 Luna matches GPT-5.6 Sol's factuality at about one hundredth of the task cost. Both GPT-6 models also show lower rates of misleading claims about completed coding work in OpenAI's alignment tests.

The launch evidence supports a narrower conclusion than “every score went up.” GPT-6 improves merge-ready coding, factuality, agent workflows, and cost per accepted task. A historical DeepSWE number for GPT-5.6 Sol was higher than the new Sol launch result, so cross-launch tables must not be presented as a clean universal capability win.

The best reading is operational: the new family reduces the cost of reaching an acceptable answer. Luna high improves by 5.4 points on AutomationBench while costing 58% less per task than its predecessor. On OSWorld, Luna max exceeds GPT-5.6 Sol medium at one tenth of the task cost.

On the supplied FrontierCode 1.1 Main snapshot, Claude Opus 5.5 beats GPT-6 Sol at every matched effort. The smallest gap is 3.0 percentage points at xhigh; the best observed scores are 54.6% for Opus and 49.3% for Sol. That benchmark grades whether a patch is mergeable, including tests, scope discipline, style, and repository standards.

Opus 5.5 is the stronger FrontierCode model in this snapshot. GPT-6 Sol is the efficiency play, not the benchmark winner.

Two other rows from the comparison should not be used as clean head-to-head claims. AutomationBench shows Sol xhigh at 33.2% and Opus max at 40.0%, but the efforts differ. OSWorld shows 60.5% versus 81.8%, but the evaluation setups differ. We record them as directional context, not proof of a model-wide gap.

Existing GPT-5.6 users do not need a new mental model. Move complex Codex work from 5.6 Sol to GPT-6 Sol; move focused, high-volume work from 5.6 Luna to GPT-6 Luna. Then tune effort and routing against your own accepted-task rate.

Luna is cheap when verification is cheap. Sol is cheap when one correct decision prevents several failed implementation loops.

A cost-aware agent stack does not repeat one model and reasoning effort across every worker. Sol can own decomposition, constraints, and synthesis. Luna workers can answer narrow questions and return evidence that code or a reviewer can verify.

Each worker needs one question, a limited search surface, a required artifact, and a verification command. The savings disappear when several workers duplicate a repository scan or return prose that the orchestrator must rediscover. The existing Supercode for Codex workflow helps keep the shared context layer bounded; it does not assign models automatically.

Luna costs one twentieth of Sol at standard API rates, but token price is only the first term. A missed requirement can trigger another repository scan, another test cycle, and another review.

The GPT-6 release improves default cache-hit rates in addition to preserving the 90% discount on cached input reads. The OpenAI prompt-caching guide says reasoning effort and tool availability can now change without invalidating earlier reusable context, while explicit breakpoints let API builders choose where the cached prefix ends.

OpenAI reports that GitHub reduced the share of prompt tokens requiring fresh processing by more than 50% across billions of requests. That is a deployment observation, not a guaranteed hit rate. Supercode's Context Boosting and Token Optimization target the same operational problem from the Codex side: bounded project context and less repetitive tool output. Actual savings depend on the task.

OpenAI continues to describe GPT-6 Astra as its best model across the board. Sol and Luna move the cost–intelligence curve; they do not erase the frontier tier.

Escalate when a task crosses systems, lacks reliable acceptance tests, carries material security or migration risk, or repeatedly fails under Sol. Do not use model choice to relax approval, sandbox, ownership, or verification rules.

Run a frozen repository task set with identical state, instructions, tools, reasoning effort, and acceptance checks. Change one routing rule at a time.

• Record the model, effort, input, output, cache reads, cache writes, latency, and tool calls.
• Count retries, reviewer intervention, failed checks, and rework before comparing task cost.
• Start Luna only where failure is cheap, visible, and reversible.
• Escalate when evidence conflicts, a check fails twice, or the task widens beyond its boundary.
• Keep approval, sandbox, ownership, and deployment authority independent from the model.
• Recheck model snapshots, pricing, and routing thresholds when OpenAI changes the service.