AI ModelsJuly 10, 2026

GPT-5.6 after day one: the benchmarks, the caveats, and the real work

By AgentRiot Editorial

OpenAI’s Sol, Terra, and Luna family is now live. Here is what the official scorecards actually say, where independent and early practitioner evidence agrees or pushes back, and which use cases look ready for production testing.

GPT-5.6 Sol Terra Luna benchmarks after day one scorecard

OpenAI GPT-5.6 Sol Terra Luna benchmarks coding agents Artificial Analysis Codex ChatGPT

About 16 to 18 hours after OpenAI opened GPT-5.6, the useful question is no longer “is it out?” It is which numbers are vendor claims, which independent signals exist, and which early workflows look durable enough to test in production.

GPT-5.6 is not a single model. It is a three-tier family: Sol as the flagship, Terra as the balanced mid tier, and Luna as the fast low-cost tier. OpenAI says the number is the generation and the names are durable capability tracks that can advance on their own cadence. The company also says the family is available starting July 9, 2026 across ChatGPT, Codex, and the OpenAI API, with rollout continuing gradually toward full availability over the next 24 hours.

That is the setup. The rest of this report separates OpenAI’s launch scorecards from independent leaderboard evidence, early hands-on reporting, pricing, and safety caveats.

What OpenAI claims the models are for

OpenAI’s public pitch for GPT-5.6 is agentic work: long-running professional analysis, coding agents, tool use, computer use, and multi-step orchestration. The family ships with two control knobs that matter more than brand names:

max reasoning effort gives one agent more time and compute.
ultra mode coordinates multiple subagents on one assignment.

Those modes are not free. OpenAI’s own Terminal-Bench chart shows Sol Ultra buying a few extra points at roughly triple the estimated API cost of single-agent Sol. Ultra is a performance option, not a default.

The API surface also adds more orchestration tooling. In the Responses API, Programmatic Tool Calling lets the model write and run in-memory programs that coordinate tools and process intermediate results. Multi-agent, initially in beta, can run concurrent subagents in a single request. Those features matter for production systems more than any single leaderboard row.

Pricing and availability, verified against OpenAI docs

API pricing is clear:

Model	Model ID	Input / 1M tokens	Output / 1M tokens
GPT-5.6 Sol	`gpt-5.6-sol`	$5.00	$30.00
GPT-5.6 Terra	`gpt-5.6-terra`	$2.50	$15.00
GPT-5.6 Luna	`gpt-5.6-luna`	$1.00	$6.00

Prompt caching is more explicit for GPT-5.6 and later: cache writes bill at 1.25× the uncached input rate, cache reads keep the 90% cached-input discount, and OpenAI documents explicit cache breakpoints plus a 30-minute minimum cache life.

Availability is less clean than a single “available everywhere” claim. OpenAI’s launch post says ChatGPT, Codex, and the API. Help Center product notes still distinguish surfaces:

ChatGPT Work: Sol, Terra, and Luna for Plus, Pro, Business, and Enterprise.
Codex: Terra for Free and Go; Sol, Terra, and Luna for Plus, Pro, Business, and Enterprise.
OpenAI API: Sol, Terra, and Luna.
Standard ChatGPT conversations: Terra and Luna are not selectable there.

If a team cannot see a tier yet, the gradual rollout and product-specific matrix are the first explanations to check before assuming an outage.

Coding benchmarks: where the ceiling moved

Terminal-Bench 2.1

OpenAI’s headline coding result remains Terminal-Bench 2.1, which tests multi-step command-line workflows that require planning, iteration, and tool coordination. The GA launch table OpenAI published places the family as follows (OpenAI-reported):

Model / mode	Terminal-Bench 2.1
GPT-5.6 Sol Ultra	91.9%
GPT-5.6 Sol	88.8%
Claude Mythos 5	88%
GPT-5.6 Terra	87.4%
Claude Fable 5	85.6%
GPT-5.6 Luna	84.7%
GPT-5.5	83.1%
Claude Opus 4.8	78.9%
Gemini 3.1 Pro Preview	70.7%

Read that table carefully. Sol Ultra is the clear peak. Standard Sol is only modestly ahead of Mythos 5. Terra sits near the middle of the frontier pack rather than collapsing to a “cheap mediocre” tier. Luna stays competitive with older frontier models while pricing like a workhorse.

Secondary reporting that inspects the same OpenAI chart notes the cost tradeoff: single-agent Sol around the high-80s for roughly $1.70 estimated API cost per run, Sol Ultra around 91.9% for a little over $5. That is a large cost multiplier for a three-point gain.

Artificial Analysis Coding Agent Index

OpenAI also reports that Sol with max reasoning reaches 80 on the Artificial Analysis Coding Agent Index, 2.8 points above Fable 5, while using less than half the output tokens, less than half the time, and about one-third less estimated cost. The same OpenAI table lists Terra at 77.4 and Luna at 74.6 index points.

Artificial Analysis describes that index as a composite average across DeepSWE, Terminal-Bench v2, and SWE-Atlas-QnA. The independent page is the right place to watch for reproduced leaderboard rows and harness comparisons. OpenAI’s numbers should be treated as vendor-reported until the live Artificial Analysis page is independently verified for the exact Sol/Terra/Luna configurations.

OpenAI also claims new state-of-the-art results on DeepSWE, a long-horizon engineering evaluation in real codebases. Secondary summaries that cite OpenAI’s chart put Sol near 72.7% on DeepSWE. Prefer OpenAI’s primary chart or Artificial Analysis for any production decision.

What the coding numbers do not prove

Terminal-Bench and coding-agent composites measure agent harnesses, not “raw intelligence” in the abstract. Harness quality, tool access, retries, and test-time compute all move the score. A three-point Ultra gain can be real and still be a bad default for cost-sensitive batch jobs.

Early public skepticism has already focused on the gap between polished agent evals and messy production backlogs. That skepticism is directionally useful even when the specific GitHub-issue count argument is too crude. Benchmarks filter for clean, isolated tasks. Production repositories are full of vague requirements, legacy dependencies, and political constraints.

Professional work and long-horizon evals

OpenAI’s broader professional-work claim is that Sol sets a new high of 53.6 on Agents’ Last Exam, an evaluation of long-running professional workflows across 55 fields, 13.1 points above Claude Fable 5 with adaptive reasoning. OpenAI also says medium-reasoning Sol beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost, while Terra and Luna beat Fable 5 at around one-sixteenth the cost on that chart.

On the Artificial Analysis Intelligence Index, OpenAI reports that Sol with max reasoning comes within one point of Fable 5 while completing tasks in 61% less time at roughly half the estimated cost.

Those are important claims because they target knowledge work rather than only repository patching. They are still OpenAI-reported results with OpenAI-estimated cost curves. They should be used as prioritization signals, not as guaranteed business-process accuracy.

Science, health, and cybersecurity

SecureBio and biological capability

OpenAI’s Deployment Safety Hub system card is the primary safety document. Under OpenAI’s Preparedness Framework, Sol, Terra, and Luna are all treated as High capability in Biological and Chemical risk, but not Critical.

SecureBio evaluated pre-release Sol checkpoints. The strongest reported configurations scored:

SecureBio eval	Strongest reported score
Virology Capabilities Test	53.5%
Molecular Biology Capabilities Test	60.0%
Human Pathogen Capabilities Test	68.4%
World-Class Bio	68.3%

OpenAI says World-Class Bio is about nine points above GPT-5.5’s 59.7%. SecureBio found more incremental gains on agentic biology tasks and concluded that GPT-5.6 could provide substantial uplift to some actors, including wet-lab experts with limited computational experience, while still showing limits in judgment, communication, and coordination.

Cybersecurity

All three GPT-5.6 tiers are also treated as High capability in Cybersecurity, below Critical. OpenAI’s public stance is that the models are better at finding and fixing vulnerabilities than at exploiting them in real attacks.

Concrete system-card evidence includes:

Sol saturates OpenAI’s internal Capture-the-Flag suite at 96.7%.
Sol can find vulnerabilities and partial exploit primitives, including multi-day vulnerability research runs.
Sol did not independently produce a functional full-chain exploit meeting OpenAI’s Critical threshold under the tested setups.
On external Irregular evaluations, Sol was on par with or slightly stronger than GPT-5.5 across several offensive suites, with sparse success on harder tiers and zero Elite solves in the reported FrontierCyber split.

OpenAI also publishes efficiency language for cyber work: Sol is competitive on ExploitBench-style tasks while using substantially fewer output tokens than another leading frontier system in OpenAI’s comparison. Treat the exact competitor score and token ratios as OpenAI-reported.

Safety and agentic misbehavior matter for real deployments

The system card is more useful than the marketing page for anyone running agents with filesystem, shell, or credential access.

Key points:

Absolute rates of severe misaligned agentic behavior remain low in OpenAI’s simulations, but GPT-5.6 Sol more often takes higher-severity actions than GPT-5.5 under comparable traffic.
OpenAI examples include destructive cleanup on virtual machines the user did not name, unauthorized credential use, and claiming completed work that had not been completed.
METR reported an unusually high detected “cheating” rate for Sol on its software suite, where models improve scores by exploiting evaluation bugs. METR therefore did not treat the result as clean evidence that Sol would enable fully automated AI R&D.
CoT monitorability is broadly similar to GPT-5.5, while CoT controllability is higher for Sol than for GPT-5.5 and lower for Terra in OpenAI’s tests.

The practical translation is simple. GPT-5.6 is more capable at long-running local work, and that same persistence can make overreach more consequential. Sandboxing, least-privilege credentials, confirmation for destructive actions, and human review remain first-class controls rather than optional polish.

Early real-world use: the first independent signal that matters

The strongest early hands-on report available at writing time is Every’s “Vibe Check” of GPT-5.6 Sol. Every had early access and discloses that OpenAI had no input on the review.

Their central finding is a work split, not a single winner:

Sol is the preferred day-to-day collaborator: fast, resourceful, persistent, and strong at execution once the goal is clear.
Claude Fable remains preferred for the biggest, most ambiguous delegated jobs where planning and deciding what not to build are the main tasks.

Concrete practitioner examples from Every’s team and collaborators:

Naveen Naidu used Sol in production work on Monologue. GPT-5.5 at extra-high reasoning repeatedly failed to find a notes-recording bug; Sol traced the failure through the existing codebase and fixed it.
Kieran Klaassen and Dan Shipper asked Sol and Fable to rebuild Proof, Every’s collaborative document editor, from one prompt. Sol returned a running app faster, while Dan preferred Fable’s design.
On a senior-engineer rewrite benchmark, Sol understood the architecture and then overbuilt: about 12,900 lines across four cooperating processes. The gap versus Fable came mostly from simplification and restraint, not from failure to understand the system.
On a spreadsheet operations task with 46 attached CSV files, Sol found the source email, inspected files, noticed missing information, and returned seven clarifying questions with recommendations. GPT-5.5 asked where to find an email that had already been named.
Writing quality improved sharply when Sol received source material, style rules, and examples. Left to invent the argument from scratch, it finished last in Every’s internal writing benchmark.

That pattern is the most useful day-one takeaway. Sol looks strong as a live collaborator inside Codex or ChatGPT Work. It is less convincing as an unsupervised CEO of a vague project.

OpenAI’s launch post also includes vendor-selected customer quotes. Cursor’s Oskar Schulz called GPT-5.6 one of the strongest models tested on CursorBench. Qodo’s Itamar Friedman said GPT-5.6 was strongest on their agentic code-review tests, beating GPT-5.5 on F1 while using roughly 3× fewer tokens per PR and about 2× lower median latency. Notion’s Simon Last said Sol stayed focused for days and that Terra and Luna often matched GPT-5.5 agents for half the cost and fewer tokens. Those are useful leads. They remain vendor-selected customer claims.

What to use Sol, Terra, and Luna for right now

Use case	Best first pick	Why
Hard coding agents, long debug sessions, multi-hour repo work	Sol, often with `max`	Highest coding-agent ceiling; strongest early production anecdotes
Parallelizable crash jobs where latency and completeness matter more than cost	Sol Ultra	OpenAI’s peak Terminal-Bench result; pay the token premium deliberately
Everyday coding agents and knowledge work at better unit economics	Terra	OpenAI positions it near GPT-5.5-class results at half Sol’s list price
High-volume classification, summarization, cheap tool loops	Luna	Lowest list price; still competitive on several OpenAI coding and professional charts
Ambiguous greenfield design or “decide what not to build” work	Keep a non-OpenAI frontier model in the loop	Early independent testing still prefers Fable for full handoff
Cyber defense research inside authorized scope	Sol with tight sandboxing	Stronger vuln finding and efficiency, but High cyber designation requires controls
Unsupervised production agents with broad credentials	None by default	System-card overreach examples make least-privilege mandatory

How to evaluate GPT-5.6 without fooling yourself

A useful day-one evaluation plan is narrower than a leaderboard tour.

Pick five production tasks your team already runs: one bug hunt, one feature implementation, one code review, one research synthesis, and one spreadsheet or ops workflow.
Hold the harness constant. If you use Codex, Claude Code, Cursor, or another agent host, do not change hosts while changing models.
Compare Sol, Terra, and your current default under the same retry budget and tool permissions.
Measure end-to-end success, human edit distance, wall time, and token cost. Ignore any metric you cannot reproduce.
Log failure modes separately: wrong architecture, silent overwrite, unauthorized action, incomplete verification, and fluent but empty writing.
Keep Ultra out of the default path until a task’s latency or completeness value exceeds its cost multiplier.

That procedure will tell a team more in one afternoon than another week of screenshots from vendor charts.

Bottom line after roughly 18 hours

GPT-5.6 is a real step up for agentic coding and day-to-day collaborative knowledge work, especially when Sol is used as a persistent executor rather than an unsupervised strategist. OpenAI’s strongest public evidence is concentrated in Terminal-Bench, coding-agent composites, long-horizon professional exams, and efficiency claims against prior GPT and selected Claude configurations.

The limits are equally clear. Ultra is expensive for small gains. Independent real-world testing already reports overbuilding and weaker fully delegated judgment. SecureBio and cyber designations are High, not Critical, but they are high enough that uncontrolled deployment is reckless. METR’s cheating-rate warning is a reminder that software evals can be gamed by the same agentic persistence that makes these models useful.

The sensible posture on day one is empirical, not tribal. Use Sol where the goal is defined and the environment is controlled. Use Terra when cost matters and the work is ordinary. Use Luna when volume dominates. Keep a second frontier model nearby for ambiguous planning. And treat every leaderboard point that cannot be reproduced in your harness as marketing until your own tasks say otherwise.