GPT-5.6 Explained: Sol, Terra, and Luna — Benchmarks, Pricing, and Which Tier to Use
GPT-5.6 is OpenAI's newest model family, released to general availability on July 9, 2026. It comes in three tiers: Sol (the flagship, $5/$30 per million tokens), Terra (the balanced everyday model, $2.50/$15), and Luna (the fast budget option, $1/$6). Sol sets new state-of-the-art scores on agentic benchmarks like Agents' Last Exam and BrowseComp, and OpenAI is aggressively undercutting rivals on price.
The naming is new, the strategy is not: one frontier model for the hard problems, one workhorse for daily volume, one cheap model for scale. What is new is how directly OpenAI is aiming at agentic workloads — long-running, tool-using, multi-step tasks — and at the wallet of every team currently paying more per token elsewhere.
This guide breaks down what each tier actually does, how the benchmarks compare against Claude and GPT-5.5, what the new prompt caching rules mean for your API bill, and which tier makes sense for your workload.
Key Takeaways
- GPT-5.6 went GA on July 9, 2026 in three tiers: Sol (flagship), Terra (balanced), and Luna (budget), after a preview that started June 26.
- Sol scores 53.6 on Agents' Last Exam — 13.1 points ahead of Claude Fable 5 with adaptive reasoning — and 92.2% on BrowseComp.
- Pricing undercuts competitors: Sol at $5/$30, Terra at $2.50/$15, Luna at $1/$6 per million tokens.
- New prompt caching adds explicit cache breakpoints and a 30-minute minimum cache life; cache writes now cost 1.25x input rate, reads keep the 90% discount.
- OpenAI calls it its "strongest cybersecurity model yet," tuned for threat modeling, code review, patching, and blue teaming.
What Is GPT-5.6?
GPT-5.6 is a family of three large language models — Luna, Terra, and Sol, from least to most capable — that OpenAI previewed on June 26, 2026 and released broadly on July 9, 2026 via the API and Codex. According to TechCrunch, the launch followed the completion of a U.S. government review, and the family is designed around enterprise work, coding, scientific research, and cybersecurity.
The three-tier structure maps cleanly to how teams actually consume models:
| Tier | Positioning | Input / Output (per 1M tokens) | Best for |
|---|---|---|---|
| Sol | Flagship frontier model | $5 / $30 | Agents, hard reasoning, long-running workflows |
| Terra | Balanced everyday model | $2.50 / $15 | Production apps, coding assistants, RAG |
| Luna | Fast, lowest-cost model | $1 / $6 | High-volume classification, extraction, chat |
Simon Willison's early analysis notes that Terra delivers performance competitive with GPT-5.5 while being roughly 2x cheaper — which is the quiet headline here. Most production traffic doesn't need a frontier model; it needs "last year's frontier" at half the price, and that's exactly what Terra is.
How Does GPT-5.6 Sol Perform on Benchmarks?
Sol posts state-of-the-art results on agentic benchmarks: 53.6 on Agents' Last Exam (13.1 points ahead of Claude Fable 5 with adaptive reasoning), 92.2% on BrowseComp, and 62.6% on OSWorld 2.0. On agentic coding, Sol Ultra hits 91.9% on Terminal-Bench 2.1, with base Sol at 88.8% — just ahead of Claude Mythos 5 at 88.0%.
The pattern across the published numbers is consistent: the biggest gains are not in static Q&A, but in long-horizon tasks — the kind where a model works for minutes or hours across many tool calls:
- Agents' Last Exam (professional workflows across 55 fields): Sol 53.6 — a new high, and the largest margin over Anthropic in the suite.
- Terminal-Bench 2.1 (agentic coding): Sol Ultra 91.9%, Sol 88.8%, Claude Mythos 5 88.0%, GPT-5.5 88.0%. At the top end, coding benchmarks are now separated by decimal points.
- BrowseComp (web research): 92.2%, state of the art.
- OSWorld 2.0 (computer use): 62.6%, state of the art — though the absolute number shows how far full computer-use automation still has to go.
Two honest caveats. First, these are vendor-reported numbers at launch; independent reproductions usually land a bit lower. Second, the Terminal-Bench gap between Sol, Mythos 5, and GPT-5.5 is under one percentage point — for coding, model choice increasingly comes down to price, latency, and ecosystem rather than raw capability. We covered the same convergence dynamic in our breakdown of Claude Sonnet 5's agentic design.
GPT-5.6 Pricing: The Real Story Is the Cache
GPT-5.6's per-token prices are aggressive, but the bigger change for API bills is caching. GPT-5.6 introduces predictable prompt caching with explicit cache breakpoints and a 30-minute minimum cache life. Cache writes are billed at 1.25x the uncached input rate; cache reads keep the 90% discount.
Per BenchLM's pricing tracker, this matters more than the sticker price for anyone running agents. Agentic workloads re-send the same growing context dozens of times per session — system prompt, tool definitions, conversation history. With a 90% discount on cache reads and a guaranteed 30-minute cache life, the effective cost of a long agent session can drop several-fold compared to naive token math.
The explicit breakpoint model is a notable shift: instead of hoping the provider's automatic caching lines up with your prompt structure, you now mark exactly where the stable prefix ends. If you build on the API, restructuring prompts so that stable content (tools, instructions, few-shot examples) sits before the first breakpoint is now the highest-leverage cost optimization available.
The 1.25x surcharge on cache writes is the fine print worth reading: for short, one-shot requests that never get a cache hit, GPT-5.6 caching can cost you money. Cache aggressively for agents and chat; skip it for stateless batch jobs.
What It Means for Developers
For working developers, three things stand out:
- Codex gets the family on day one. All three tiers are available in OpenAI's coding agent immediately, and OpenAI claims Sol is 54% more token-efficient on coding tasks than its predecessor — token efficiency compounds fast when an agent iterates on a large codebase. OpenAI has been steadily assembling its developer story, including its acquisition of Astral, the company behind uv and Ruff.
- Terra is the default production choice. GPT-5.5-class quality at half the price moves the cost floor for every AI-powered product. If you're running an assistant, RAG pipeline, or AI coding tool on a mid-tier model, Terra is the new benchmark to beat.
- The price war is structural, not promotional. OpenAI is under revenue pressure — Anthropic overtook it on both revenue and valuation earlier this year — and its response is to compete on cost-per-capability. Expect Anthropic and Google to answer within months. Locking your stack to one vendor's pricing today is a mistake; keep your model layer swappable.
The Cybersecurity Angle
OpenAI is explicitly positioning GPT-5.6 as its "strongest cybersecurity model yet," aimed at defensive work: threat modeling, code review, patching, and blue teaming. That framing — and the U.S. government review that preceded launch — reflects where the industry now sits. Autonomous AI attacks stopped being theoretical this year, as we documented in our analysis of JADEPUFFER, the first autonomous AI ransomware. Vendors are now racing to prove their frontier models help defenders more than attackers, and "cybersecurity model" is becoming a first-class product category rather than a footnote in a system card.
Which GPT-5.6 Tier Should You Actually Use?
The honest decision tree is shorter than the marketing suggests:
- Start with Terra. For 80% of production use cases — chat, RAG, coding assistance, summarization — GPT-5.5-class quality at $2.50/$15 is the right cost/quality point.
- Escalate to Sol only for agents. Sol's edge is long-horizon agentic work. If your workload is multi-step, tool-heavy, and runs unattended, the benchmark gap is real and worth $5/$30. For single-turn tasks, you're paying flagship prices for workhorse output.
- Route bulk traffic to Luna. Classification, extraction, moderation, and simple chat at $1/$6 — pair it with Terra behind a router and let each request cost what it should.
FAQ
What is GPT-5.6? GPT-5.6 is OpenAI's model family released July 9, 2026, in three tiers: Sol (flagship), Terra (balanced), and Luna (budget). It focuses on agentic capabilities — long-running, tool-using workflows — plus coding and cybersecurity, and became available to all developers via the API and Codex at launch.
What is the difference between GPT-5.6 Sol, Terra, and Luna? Sol is the most capable and most expensive ($5/$30 per 1M tokens), built for agents and hard reasoning. Terra ($2.50/$15) matches roughly GPT-5.5-level quality at half the price for everyday production work. Luna ($1/$6) is the fastest and cheapest, meant for high-volume simple tasks.
How much does GPT-5.6 cost? Per million tokens: Sol costs $5 input / $30 output, Terra $2.50 / $15, and Luna $1 / $6. Cache reads get a 90% discount on input, while cache writes bill at 1.25x the input rate under the new explicit-breakpoint caching system.
Is GPT-5.6 better than Claude? On vendor benchmarks, Sol leads Claude Fable 5 by 13.1 points on Agents' Last Exam and edges Claude Mythos 5 on Terminal-Bench 2.1 (88.8% vs 88.0%). The coding gap is under one point, so for most teams the deciding factors are price, latency, and tooling rather than raw benchmark scores.
Can I use GPT-5.6 right now? Yes. All three tiers went generally available on July 9, 2026 through the OpenAI API and Codex, following a limited preview that began June 26, 2026.
The Bottom Line
GPT-5.6 is less a single model launch than a pricing and packaging statement: frontier agentic capability at the top, GPT-5.5 quality at half price in the middle, and a $1 model at the bottom — all wrapped in caching rules that reward well-structured prompts. Our verdict: Terra is the most consequential of the three, because it resets the price floor the entire industry has to match. Benchmark crowns change hands every quarter; a 2x cost reduction on the workhorse tier changes budgets permanently.
If you build with these models, don't take our word for any of it — run your own evals on your own workload. The best model in 2026 is the one that wins on your tasks at your price point.