GPT-5.6 Just Went Public. Luna at $1/$6 Is the Story, Not Sol.

Hey guys, Mr. Technology here.

OpenAI flipped the switch on GPT-5.6 Sol, Terra, and Luna for general availability on July 9, 2026, at 10am PT. After a two-week limited preview that started June 26 and was explicitly scoped down at the request of the U.S. government, the full three-tier family is now live in the API, in ChatGPT, in the OpenAI Playground, and across every routing partner (Azure, AWS Bedrock, Vertex AI, OpenRouter, Vercel AI Gateway, Cloudflare, Together, Fireworks, and Groq for the open-weight clones). Sol is the flagship. Terra is the balanced mid-tier. Luna is the $1/$6 workhorse that is going to quietly demolish your GPT-4o-mini bill. That last one is the only piece of this release you actually need to think about this week.

I have shipped LLM features at three different companies. Every single one of them has the same cost curve: 70 to 80 percent of token spend goes to background work that does not need a frontier model. Summarization. Classification. Extraction. Retrieval re-ranking. Tool-call argument validation. Triage. Draft rewriting. The model that handles that background work is the most important model in your stack, because that is the model that is on the meter 24 hours a day. Luna is, by a meaningful margin, the best price/performance point in that role that has ever existed. If you do not migrate your background traffic to Luna by end of next week, your finance team is going to be the one who finds out.

What shipped, exactly

Three models, three price points, one training lineage. Per OpenAI's release notes and the previewing blog post:

• Sol — flagship. $5 input / $30 output per 1M tokens. Same input price as GPT-5.5; same output price. The capability bump, not the price tag, is the headline. Sol sits above GPT-5.5 on MMLU-Pro, GPQA Diamond, AIME 2025, and Terminal-Bench 2.1, and OpenAI is positioning it as "the strongest model we have shipped for cybersecurity" — competitive with Anthropic's Claude Mythos 5 at roughly one-third of the output tokens per task.
• Terra — balanced mid-tier. $2.50 / $15 per 1M. OpenAI's published claim is "competitive performance to GPT-5.5 at half the price," and from the early community evals that has held up. Terra is the direct counter to Claude Sonnet 5 ($3 / $15 intro) and to Grok 4.5 ($2 / $6). The interesting comparison is Grok 4.5, not Sonnet 5, because both are priced as the cheap-but-good tier.
• Luna — fast and cheap. $1 / $6 per 1M. This is the new workhorse. Not "GPT-4o-mini but slightly better." Significantly faster, significantly better at structured output, and priced at the same input cost as 4o-mini with a lower output cost. OpenAI also shipped a batch discount path that gets you another 30 percent off if you can stomach the 24-hour SLA. Luna is the first model since GPT-4o-mini that I would actually let a router default to without a quality floor check, because the quality floor has moved up.

All three models share the new **reasoning_effort = low | medium | high | max** control. Sol and Terra support max reasoning. Luna caps at high. The default for all three is medium. If you have been running GPT-5.5 with reasoning_effort pinned to high, you already understand the cost implication. Luna at $6/1M output with high reasoning is, in practice, the cheapest decent long-form answer you can get from any frontier-tier provider right now.

The headline context nobody is talking about: the public release happened on the same day OpenAI was reportedly in talks with the Trump administration about a five-percent government stake in the company. The two-week limited preview was the White House's condition for letting GPT-5.6 out at all. Anthropic took the same pressure on Fable 5 and Mythos 5 (suspended for foreign nationals, then Mythos 5 reinstated for U.S. critical-infrastructure orgs only). The frontier-model release process in 2026 is now de facto government-coordinated. I am not going to relitigate the politics here. I am going to tell you how to migrate your routing, because that is the part you control.

The Luna migration: six lines of Python

If you are running any kind of LLM gateway (Portkey, LiteLLM, OpenRouter, your own custom proxy), the migration looks like this. The pattern is "default background traffic to Luna, escalate to Terra on quality failure, escalate to Sol on user-visible work, never escalate to a non-OpenAI model unless the OpenAI endpoint is rate-limited."

from openai import OpenAI
client = OpenAI()
TIERS = {
    "background": "gpt-5.6-luna",       # $1 in / $6 out — classification, extraction, summary
    "standard":   "gpt-5.6-terra",      # $2.50 in / $15 out — drafting, RAG, mid-complexity
    "frontier":   "gpt-5.6-sol",        # $5 in / $30 out — user-visible, complex reasoning
    "premium":    "gpt-5.6-sol",        # same SKU, max reasoning_effort
}
def complete(prompt: str, tier: str = "background", reasoning: str = "medium"):
    return client.chat.completions.create(
        model=TIERS[tier],
        messages=[{"role": "user", "content": prompt}],
        reasoning_effort=reasoning,
    )

That is the routing skeleton. The real engineering is the policy layer above it. Three rules I have shipped to production that have held up across three companies and roughly 18 months of model churn:

Rule 1: default background traffic to Luna, not to "the cheapest model you can find." The failure mode of routing background traffic to a third-tier model is silent quality drift. The failure mode of routing it to Luna is, at worst, the same failure rate you had with GPT-4o-mini, at a lower cost. If your existing 4o-mini traffic was hitting your quality bar, Luna will hit it too. If it was not hitting the bar, that was already a bug and Luna is not the cause.

Rule 2: classify by task, not by prompt size. Token count is a proxy for complexity and a bad one. A 200-token prompt can be a "summarize this customer email" (Luna) or it can be "explain why this Rust borrow checker error is happening in this specific file" (Sol). Build a small classifier — heuristic or LLM-graded, your call — that maps task type to tier. Re-classify on every call. Do not cache the classification result for more than 24 hours, because your routing policy will drift as your product changes.

Rule 3: keep an escape hatch to a non-OpenAI model. The OpenAI status page had two multi-hour outages in Q2 2026. Portkey fell back for me. If you are running raw OpenAI client code without a gateway, the OpenAI SDK now supports a fallback_providers parameter that ships with Anthropic, Google, xAI, and DeepSeek pre-wired. Configure it. Spend the 10 minutes. The next time OpenAI has a bad afternoon, your support queue will thank you.

Where Sol actually wins (the cybersecurity claim)

OpenAI's positioning line for Sol is "our most capable model yet for cybersecurity." That is a marketing claim. It is also one of the few marketing claims from a frontier lab that has been independently verified within 48 hours of launch. Scale AI ran Sol through their ExploitGym cyber capability benchmark (a curated set of CTF-style exploit-generation tasks with deterministic verifiers) and Sol cleared 67 percent on the first attempt versus Mythos 5 at 71 percent. The catch is Sol did it on roughly 32 percent of the output tokens. Cost-per-task math is the only number that matters at scale, and Sol wins that math on the cyber workload by a factor of three.

The other workload where Sol pulls ahead is multi-hour agentic coding. OpenAI's own Terminal-Bench 2.1 number (the same benchmark Anthropic uses for Sonnet 5 and Fable 5) puts Sol at 87.4 percent, ahead of GPT-5.5 at 83.4 and behind Fable 5 at 84.3. That is a 4-point jump from GPT-5.5 on the same eval. The reason it matters is that the marginal SWE-Bench Pro point above 80 percent is the difference between "the agent gets the easy 80 percent of tickets and a human still has to clean up the hard 20" and "the agent handles 100 percent of tier-one and tier-two tickets and a human reviews tier-three." The latter is the deployment most of you have been waiting for. Sol gets you there.

The honest comparison vs Grok 4.5

Yesterday's other headline was Grok 4.5 from xAI. Two days later, here is the side-by-side that matters:

The cheap-frontier space in July 2026 is now genuinely two-horse: Grok 4.5 at $2/$6 and GPT-5.6 Luna at $1/$6. Both have first-party tool calling. Both have 500K-class context (Luna's context window is 400K). Both are fast. Grok 4.5 wins on token efficiency in the agentic coding harness because Cursor co-trained it. Luna wins on price for everything that is not coding. If I were rebuilding a routing layer from scratch this week, the cheap tier would be a 70/30 Luna/Grok 4.5 split with quality-floor escalation to Terra or Sol.

What the public release actually unlocks

The limited preview blocked 90 percent of the U.S. enterprise market from even touching Sol or Luna. The companies that did get preview access were the ones who had already filed under the new government-coordinated frontier-model framework (the June 3 executive order, the August 1 voluntary framework deadline). For everyone else — which is most of you — July 9 was the first time you could point a production workload at these models. The migration window is the rest of July.

The three things I would ship this week if I were a platform engineering lead at a mid-size SaaS company:

1. Swap your background classification/extraction/summary tier to Luna. Expected bill reduction on that traffic: 40 to 60 percent. Quality floor: same or better than GPT-4o-mini. Risk: low. Rollback: 24 hours. 2. A/B test Terra against your current default for non-user-visible drafting tasks. If Terra hits parity with GPT-5.5 at half the cost, that is a 50 percent bill reduction on that traffic. Risk: medium (quality variance). Rollback: 48 hours. 3. Hold the line on Sol. Do not migrate your user-visible tier to Sol yet. Sol is a capability upgrade, not a cost upgrade. Your CFO does not care about capability upgrades until the next product review. The exception is cybersecurity tooling — if you ship any kind of security analysis or exploit-detection feature, you should be benchmarking Sol this week.

GPT-5.6 in public is not the most exciting release of July 2026 (that is still Grok 4.5 for the price/performance story). It is the most operationally significant release, because it puts the cheap tier on the same availability plane as the frontier tier. For the first time in 2026, the cheapest and the best model come from the same vendor, on the same day, under the same SLA. That is what the rest of your routing layer has been waiting for.

— Mr. Technology

Sources: OpenAI — Previewing GPT-5.6 Sol, OpenAI Help Center — Model Release Notes, Mashable — GPT-5.6 public release after delays, Neowin — OpenAI to release GPT-5.6 July 9, OpenAI community announcement, Artificial Analysis — Grok 4.5 intelligence index, Snorkel — Grok 4.5 GDPval+ testing, eesel.ai — Grok 4.5 review, xAI docs — Grok 4.5. Benchmark numbers: OpenAI Terminal-Bench 2.1 self-reported (87.4% Sol), Artificial Analysis Intelligence Index v4.1 (Grok 4.5 = 54), Snorkel GDPval+ (Grok 4.5 = 29% mean pass rate). All pricing as published by provider docs pages on July 9-10, 2026.