OpenAI released GPT-5.6 on July 9, 2026. The framing in the announcement is worth noting before any benchmark: the pitch is not "the smartest model we have made." It is "more intelligence from every token, stronger performance per dollar, and more capability on demand." That is a different sales pitch than the one that accompanied most prior frontier releases, and it points at how the decision to adopt AI is changing for the businesses buying it.
Here is what shipped, what is genuinely new, and how to read it without getting swept up in the launch-day numbers.
Three Tiers, One Generation
GPT-5.6 is not a single model. It is a family of three, and the naming carries a deliberate design decision:
- Sol is the flagship, tuned for the hardest work.
- Terra is the balanced everyday model, priced below Sol with performance OpenAI positions as competitive with the previous GPT-5.5 flagship.
- Luna is the fastest and most affordable.
The important part is how OpenAI describes the names. The number, 5.6, identifies the generation. Sol, Terra, and Luna are "durable capability tiers that can advance on their own cadence." In plain terms, the company is telling buyers to stop thinking in version numbers and start thinking in tiers. You are meant to pick a capability level for a job and let the underlying model improve underneath that label over time.
For anyone standardizing AI across a company, that is the more useful mental model anyway. Most organizations do not need their flagship model answering routine questions. They need the right tier on the right task.
The Real Headline: Performance Per Dollar
The claim OpenAI leans on hardest is efficiency, and the numbers behind it are the ones worth internalizing.
On the Artificial Analysis Intelligence Index, a broad measure spanning agentic work, coding, and reasoning, GPT-5.6 Sol at maximum reasoning lands within a point of Anthropic's Claude Fable 5 while, per OpenAI's own estimates, completing tasks in 61 percent less time and at roughly half the cost. On the Artificial Analysis Coding Agent Index, Sol sets a new high of 80 while using less than half the output tokens and about a third less cost than the model it edges out.
The pattern repeats down the family. OpenAI reports that Terra and Luna outperform larger competing models on coding at roughly a quarter of the estimated cost, and that on the OSWorld computer-use benchmark, Sol surpasses Anthropic's Opus 4.8 while using 85 percent fewer output tokens.
Strip away the leaderboard positioning and the message is this: the unit that matters is shifting from cost per token to cost per finished task. A model that reaches the same result in fewer tokens and less time changes the math on which workflows are worth automating, even when its peak intelligence is not the highest number on the chart.
More Capability On Demand: max and ultra
GPT-5.6 keeps an efficient default and lets you spend more when a problem justifies it. Two settings sit above the normal reasoning levels:
- max gives the model more time than the previous "xhigh" setting to reason, explore alternatives, run checks, and revise its approach.
- ultra goes further by coordinating four agents in parallel by default, trading higher token use for stronger results and faster time to a finished answer on demanding tasks.
This is the quiet structural shift in the release. Multi-agent work, several instances collaborating on one problem, is moving from research demo to a setting you toggle. In the API, OpenAI exposes it through a multi-agent beta in the Responses API. It also introduced Programmatic Tool Calling, which lets the model write and run small in-memory programs that coordinate tools and filter intermediate results, rather than passing every tool response back through the model. For tool-heavy automation, that means fewer round trips and less hand-holding.
Where It Improved
Beyond efficiency, OpenAI highlights three areas of substantive gain:
Design and computer use. OpenAI calls out a "step change in design judgment." More to the point for practical work, stronger computer-use ability lets the model inspect and refine a rendered result, not just generate the code or content behind it, so it can catch visual and functional issues before handing work back.
Knowledge work. The model is positioned to take messy context from documents and tools like Slack, Notion, Microsoft 365, and Google Drive and turn it into polished, shareable artifacts: editable presentations, documents, and spreadsheets that follow templates and reference formats more faithfully than the prior generation.
Cybersecurity and science. GPT-5.6 posts large gains on security benchmarks, scoring 73.5 percent on ExploitBench against GPT-5.5's 47.9 percent, and shows broad improvement across life-sciences evaluations. OpenAI is clear that these capabilities are dual-use and gates the most sensitive cyber work behind a verified-access program.
Pricing and Availability
GPT-5.6 is available starting July 9 across ChatGPT, Codex, and the OpenAI API, rolling out globally over roughly 24 hours. API pricing, per million tokens:
- Sol: $5 input / $30 output
- Terra: $2.50 input / $15 output
- Luna: $1 input / $6 output
The family also introduces more predictable prompt caching, with explicit cache breakpoints and a 30-minute minimum cache life. In ChatGPT, paid tiers reach Sol through medium and higher effort settings, with a Sol Pro option for the highest-quality results; free and Go users get Terra. In Codex and ChatGPT Work, users can pick among the three tiers and set an effort level for each.
Read the Tables, Not Just the Prose
A discipline worth keeping on every model launch: the announcement prose and the benchmark tables do not always tell the same story, and the tables are the primary source.
GPT-5.6's efficiency lead is real and well documented. Its lead on raw intelligence is narrower than the headline suggests. In OpenAI's own comparison tables, Claude Fable 5 still edges Sol on the Artificial Analysis Intelligence Index, 59.9 to 58.9, and on the GDPval professional-work benchmark, 1,759.6 to 1,747.8 Elo. On SWE-Bench Pro, a test of difficult software engineering, Anthropic's Claude Mythos 5 leads at 80.3 percent against Sol's 64.6 percent.
None of this undercuts the release. It sharpens what the release actually is: a strong, efficient family that wins decisively on cost and speed at a given capability level, and trades the very top of the raw-intelligence chart for that efficiency. That distinction matters when a single benchmark number is about to land in a strategy deck.
Safety Scaled With Capability
OpenAI describes its most robust safety system to date for this release, calibrated per model and backed by more compute than before, including roughly 700,000 GPU hours of automated red teaming. The company states that GPT-5.6, while more capable in biology and cybersecurity than earlier models, does not cross its "Critical" risk threshold in either domain.
The notable design choice is a reasoning monitor that reviews a conversation for potential harm rather than relying on classifier flags alone, paired with the ability to update protections quickly without retraining from scratch. OpenAI reports its GPT-5.6 cyber safeguards block roughly ten times more potentially harmful activity than prior models, and, acknowledging the friction that creates for legitimate work, provides an option to retry a blocked prompt on a lower-capability model. The most sensitive capabilities are reserved for verified users.
What Honra Sees in This
The durable lesson in GPT-5.6 is not a benchmark. It is the buying posture the release quietly assumes.
Match the tier to the task. The flagship is rarely the right default for every seat and every job. A three-tier family makes that explicit, and the same logic applies whether you are choosing among GPT-5.6 models or across providers: decide deliberately which work needs the top tier and which is served just as well, and far more cheaply, by a smaller one. That deliberate matching of model to workflow is the kind of adoption decision we help clients get right.
Measure cost per finished task, not per token. When a model reaches the same result in fewer tokens and less time, the economics of automation move. Workflows that were too expensive to hand to AI a year ago deserve a fresh look on those terms, not on sticker price alone.
Go to the primary source before the decision. Launch coverage compresses a nuanced benchmark table into a single number, and the number is often the flattering one. Reading the tables directly, as we did above, is the difference between adopting a capability and adopting a headline.
GPT-5.6 is a meaningful release. The more useful thing it signals is that the frontier is no longer a single race to the smartest model. It is a set of choices about capability, cost, and control, and those choices are yours to make well.
Sources
GPT-5.6: Frontier intelligence that scales with your ambition - OpenAI Official Announcement (primary source for all benchmark figures)
GPT-5.6 System Card - OpenAI Deployment Safety Hub (safety evaluations and risk thresholds)
Programmatic Tool Calling - OpenAI API Documentation
OpenAI Releases GPT-5.6 (Sol, Terra, Luna): A Three-Tier Model Family - MarkTechPost
OpenAI unveils ChatGPT Work agent, GPT-5.6 models now available - 9to5Mac
