01
A three-model suite
Sol, Terra, and Luna span one family from flagship to fast-and-cheap. Same context window, same tool surface, same caching model. The only axis that changes is capability against price, which is exactly the axis you route on.
★ PREVIEW · LIMITED ACCESSOpenAIModel brief
OPENAI GPT-5.6 · SUITE: SOL / TERRA / LUNA · 26 JUN 2026GPT-5.6 Sol. Terminal-coding state of the art.
OpenAI's GPT-5.6 ships as a three-model suite. Sol is the flagship, and it sets a new state of the art on agentic terminal coding while pulling clear of GPT-5.5 across reasoning and tool use. We integrate it directly, eval-gated, behind a vendor-neutral abstraction. Preview now, general availability in the coming weeks.
LLM APIgpt-5.6-solFunction callingEval pipelines
Released
26 Jun 2026
Flagship model ID
gpt-5.6-sol
Sol input
$5 / 1M tokens
Sol output
$30 / 1M tokens
Context
400K context
Max output
128K max output
Modalities
Text + vision + audio
Knowledge
Mar 2026 cutoff
[TL;DR FOR CEO + CTO]
Five things to know.
- 01
It is a suite, not a single model.
GPT-5.6 ships in three sizes. Sol is the flagship at $5 / $30 per million tokens, Terra is the balanced tier at $2.50 / $15, and Luna is the fast, cheap tier at $1 / $6. The right answer is almost always a mix, routed by task difficulty.
- 02
Sol leads agentic terminal coding.
Sol sets a new state of the art on Terminal-Bench 2.1, ahead of GPT-5.5 and the current frontier from Anthropic and Google. For command-line agents and CI-style automation, this is the headline win.
- 03
Real gains over GPT-5.5, with fewer tokens.
On GeneBench v1 (long-horizon quantitative analysis) Sol beats GPT-5.5 while spending fewer tokens. On ExploitBench it is competitive with the strongest models at roughly a third of the output tokens. Quality and unit cost moved in the same direction.
- 04
Prompt caching is finally predictable.
GPT-5.6 adds explicit cache breakpoints and a 30-minute minimum cache life. Cache reads keep the 90% discount; cache writes bill at 1.25x the uncached input rate. For agents with a large shared preamble, this changes the cost model.
- 05
It is a preview. We treat the numbers as claims.
Access is limited and general availability is weeks out. We re-run our own eval suite on customer tasks before recommending a switch, and we hold the line on cost and quality, not the launch slide.
[VERIFIED PERFORMANCE]
How it stacks up.
From OpenAI's reported preview numbers, set against the current frontier. Sol leads on agentic terminal coding and financial agent tasks, posts solid gains over GPT-5.5 everywhere, and trails Claude Opus 4.8 on pure coding and multidisciplinary reasoning. We treat these as claims until our own evals confirm them on your tasks.
| Capability | GPT-5.6 Sol | GPT-5.5 | Claude Opus 4.8 | Gemini 3.1 Pro |
|---|---|---|---|---|
Agentic terminal coding Terminal-Bench 2.1 Terminus-2 public harness | 80.7% +2.5 pts vs 5.5 | 78.2% | 74.6% | 70.3% |
Agentic coding SWE-Bench Pro | 64.9% +6.3 pts vs 5.5 | 58.6% | 69.2% | 54.2% |
Multidisciplinary reasoning Humanity's Last Exam no tools / with tools | 45.6% / 55.8% +4.2 pts vs 5.5 | 41.4% / 52.2% | 49.8% / 57.9% | 44.4% / 51.4% |
Agentic computer use OSWorld-Verified | 81.0% +2.3 pts vs 5.5 | 78.7% | 83.4% | 76.2% |
Knowledge work GDPval-AA | 1824 +55 vs 5.5 | 1769 | 1890 | 1314 |
Agentic financial analysis Finance Agent v2 | 54.6% +2.8 pts vs 5.5 | 51.8% | 53.9% | 43.0% |
Numbers as reported by OpenAI in the 26 Jun 2026 preview. OpenAI also reports Sol beating GPT-5.5 on GeneBench v1 with fewer tokens, and matching the strongest models on ExploitBench at roughly a third of the output tokens. We re-run our own evals on customer tasks before recommending a switch, and a benchmark lead has to clear on your workload first.
[SOFTWARE DEVELOPMENT IMPACT]
What it changes for the team building with it.
What changes for the engineering team. Two comparisons that matter: Sol against the model it replaces (GPT-5.5), and against the cheaper tiers in its own suite (Terra and Luna), where most production volume should actually run.
| Dimension | vs GPT-5.5 | vs Terra + Luna |
|---|---|---|
Coding workflows | +2.5 pts on Terminal-Bench 2.1 and +6.3 pts on SWE-Bench Pro over GPT-5.5. The terminal-coding lead is the reason to reach for Sol on command-line agents and CI automation. | Terra handles most local edits and review at half the token price. Reserve Sol for the hard, multi-file, agentic runs where its lead actually shows up. |
Cost and latency | Same headline price as GPT-5.5 ($5 / $30 per million), but more predictable caching and better token efficiency on long-horizon tasks. Effective cost per finished task drops. | Sol is 2x Terra and 5x Luna on input. Routing easy, high-volume steps down to Terra or Luna is where the suite economics work. |
Caching + context | Explicit cache breakpoints and a 30-minute minimum cache life replace the old implicit, short-lived cache. Large shared preambles start paying back as cache reads sooner. | All three tiers share the 400K context and the same caching model, so a prompt built for Sol drops onto Terra or Luna without a rewrite. |
Migration risk | Behind a vendor-neutral abstraction, the switch from 5.5 to 5.6 is a config change plus an eval pass. Most prompts work identically. Preview access and weeks-out GA are the real gating factors. | Tiering is a routing decision, not three integrations. We wire the suite behind one interface and let the agent pick the tier at runtime by difficulty. |
Inside a Kensink build, picking Sol over Terra or Luna is a routing decision the agent makes at runtime by task difficulty, not a vendor commitment frozen at design time.
[MODEL SPEC + WHAT IS NEW]
The features that ship with it.
02
Predictable prompt caching
Explicit cache breakpoints let you mark where a cacheable prefix ends, and a 30-minute minimum cache life replaces the old short, implicit window. Cache reads keep the 90% discount; cache writes bill at 1.25x the uncached input rate from GPT-5.6 onward.
03
Token efficiency as a headline
OpenAI is reporting equal-or-better quality at fewer output tokens: stronger than GPT-5.5 on GeneBench v1 with fewer tokens, and competitive with the strongest models on ExploitBench at about a third of the output. For agentic loops, fewer tokens per step compounds.
04
Terminal-Bench 2.1 state of the art
Sol posts the top reported score on Terminal-Bench 2.1, the command-line agent benchmark. If your product runs shell-driven agents, this is the single most relevant number in the release.
05
Preview gating
GPT-5.6 is a limited preview as of 26 Jun 2026, with general availability promised in the coming weeks. Final GA pricing and context limits can still move, which is one more reason we eval before we commit a default.
[USE CASES · ROUTING]
Which tier for which job.
The point of a suite is that you do not pick one model. You route. Here is how we map the three tiers onto real workloads inside a build.
Sol
gpt-5.6-sol- Command-line and shell-driven agents (its terminal-coding lead)
- Hard, multi-file refactors and migrations
- Long-horizon reasoning where a wrong early step is expensive
The flagship. Route to it for the few genuinely hard steps in a workflow, not for everything.
Terra
gpt-5.6-terra- High-volume production generation and extraction
- Most local code edits and review
- Structured-output workflows where Sol is overkill
The balanced default. Most tokens in a well-built system should run here, at half Sol's input price.
Luna
gpt-5.6-luna- Classification, routing, and tagging
- Cheap retrieval and reranking steps
- The fast, simple sub-steps inside a larger agentic loop
The fast, cheap tier. Right whenever a smaller model clears the eval bar, which is more often than teams expect.
[PRICING · THE SUITE]
What it costs.
SolFlagship
$5 input
$30 output
$30 output
gpt-5.6-sol
Flagship. Frontier reasoning and agentic coding.
Terra
$2.5 input
$15 output
$15 output
gpt-5.6-terra
Balanced. The production default for most volume.
Luna
$1 input
$6 output
$6 output
gpt-5.6-luna
Fast and cheap. Classification, routing, simple steps.
Per million tokens. Cache reads keep the 90% discount; cache writes bill at 1.25x the uncached input rate from GPT-5.6 onward. Preview pricing, may shift at general availability.
[ALIGNMENT + SAFETY]
What the safety story says.
Preview safeguards, limited access.
OpenAI is gating GPT-5.6 behind a limited preview and staging general availability. The most capable tier carries the most scrutiny, and refusal behaviour on high-risk topics tends to be tighter on a flagship than on the smaller tiers.
Token efficiency is a safety-adjacent win.
Fewer tokens to reach the same answer means fewer chances to drift on a long agentic run, and a smaller surface for prompt injection to ride in on. We still treat it as a capability claim until our evals confirm it on your data.
Model alignment does not replace your evals.
A stronger, better-behaved model does not relax the need for task-specific evals on your prompts, your data, and your guardrails. We run the customer eval suite on every model change, every release, and a preview model raises the bar for what those evals must catch.
[OUR TAKE]
What this means for the build.
01
We are evaluating, not switching, while it is in preview.
Limited access and weeks-out GA mean we do not move a production default onto Sol yet. We do wire it into the eval harness now, so the day GA lands we already know whether the gain clears on each customer's tasks.
02
The suite is the story, not just Sol.
The interesting design move is three tiers behind one interface. In our builds the agent already routes by difficulty, so adopting the suite is mostly a routing-table change. Most traffic should land on Terra and Luna, with Sol reserved for the hard steps.
03
Terminal coding is the clearest reason to care.
If your product runs shell-driven or CI-style agents, Sol's Terminal-Bench 2.1 lead is the most concrete, testable claim in the release. That is the first eval we run, because it is the one most likely to change a routing decision.
04
Caching changes the cost model for agents.
Explicit breakpoints and a 30-minute cache life make a large shared preamble genuinely cheap to reuse. For multi-step agents that replay a big system prompt every turn, this is a real line-item saving, and it is worth re-costing existing GPT-5.5 workloads against it.
[METHODOLOGY · K-FRAMEWORK]Integrated through the
Integrated through the
K-Framework.
Every model we integrate runs through the same operating system. Three pillars, sixteen layers, one Compound Growth Loop. The methodology that keeps AI work from rotting after the first ship.
Read the K-Framework01
Foundations
Direct API integration with the model. No LangChain, no orchestration vendor, no agent framework built on quicksand. Typed contracts, the same way we wire up Postgres.
02
Amplification
An eval suite built from your real tasks gates every prompt and model change. Quality is measured before it ships, not vibed in a demo.
03
Judgment
Governance, audit, and oversight wired in from day one. Who called what, with which prompt version, at what cost. Your auditors get answers, not screenshots.
[OBSERVABILITY]
Observability your team can read.
A model in production without observability is roulette. We instrument every integration so engineering and finance can see the same numbers, and so a regression at 3am surfaces before a customer opens a ticket.
Instrumented
Cost per call
Tokens in, tokens out, dollars spent. Sliced by feature, tenant, and route. Budgets enforced where it matters.
Instrumented
Latency p50 / p95 / p99
Real distributions, not averages. We know which routes are slow, and why.
Instrumented
Eval pass rates
The same eval suite that gates a release runs continuously in production. A regression on real traffic surfaces fast.
Instrumented
Prompt + completion logs
PII scrubbed at the proxy, shipped to your SIEM. Retention controls match your compliance window.
Dashboards your team owns, not ours. At handoff you get the queries, the alerts, and the runbook. We are not in the path to read your metrics.
[COMMON QUESTIONS]
Questions we are getting asked.
- Can we use GPT-5.6 in production today?
- Not as a default yet. As of 26 Jun 2026 it is a limited preview with general availability promised in the coming weeks. We wire it into the eval harness now so we can switch the moment GA lands and the numbers clear on your tasks, but we do not move a production default onto a preview model.
- Which tier should we use, Sol, Terra, or Luna?
- Usually all three, routed by difficulty. Luna for classification, routing, and the cheap steps; Terra as the balanced default for most production volume; Sol for the few genuinely hard, agentic, or long-horizon steps. We prove the routing with evals rather than defaulting everything to the flagship.
- What is the migration cost from GPT-5.5?
- For projects built behind a vendor-neutral abstraction, the migration is a config change plus an eval pass. Most prompts work identically. If your team built directly against the SDK with no abstraction, budget one engineering day to add the seam, then the same eval pass. The real gating factor right now is preview access, not code.
- How much does GPT-5.6 cost?
- Per million tokens in the preview: Sol is $5 input and $30 output, Terra is $2.50 and $15, Luna is $1 and $6. Cache reads keep the 90% discount and cache writes bill at 1.25x the uncached input rate. Pricing can still move at general availability, so we model cost on your real traffic before committing.
- Is it actually better than Claude Opus 4.8 or Gemini 3.1 Pro?
- It depends on the task. Sol leads on agentic terminal coding and financial agent tasks in OpenAI's reported numbers, and trails Opus 4.8 on pure coding and multidisciplinary reasoning. That is exactly why we integrate every frontier model behind one abstraction and route by task, rather than betting the product on a single vendor.
[RELATED]
Worth a look next.
DIRECT INTEGRATION · NO FRAMEWORKWant GPT-5.6 Sol
Want GPT-5.6 Sol
in your product?
Eval suite at handoff, full source ownership. We integrate against the model API the same way we integrate against Postgres, and route the suite by task difficulty. Sized to your scope.