GPT-5.6 Is Here: Sol, Terra, and Luna

For twelve days, the most effective AI fashions on the planet existed and virtually no one may contact them.

That ends now! GPT-5.6 Sol, Terra, and Luna go public as we speak! The fashions are accessible by all customers (no subscription required)

That is the total breakdown of what’s on provide: three fashions, 4 costs, one precedent, and a functionality desk that ought to assist you choose the fitting mannequin. Fingers-on outcomes comply with the second entry opens.

One Technology, Three Fashions

GPT-5.6 retires OpenAI’s naming chaos for good. The quantity marks the era. This makes it simple to categorise, so the following Luna enchancment gained’t pressure a whole-family rename.

• Sol is the flagship, constructed for the toughest 10 p.c of labor: long-horizon coding brokers, safety analysis, deep scientific evaluation. The brand new reasoning controls reside right here.
• Terra is the workhorse and the apparent migration goal. GPT-5.5-class high quality at half the value, geared toward manufacturing quantity: assist, inside instruments, doc pipelines.
• Luna is the pace tier, and quietly the sleeper of the launch. The most cost effective mannequin within the household lands close to GPT-5.5 on a number of checks. Extra on why that issues beneath.

gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna are their respective names within the API. This may appear to be a small change on paper. But it surely’s a giant one for any coder who has tried preserving observe of o3, o4-mini, GPT-4 Turbo, and 4o all of sudden.

Pricing: 4 Methods to Pay

Three fashions, however 4 costs, as a result of launch week surfaced a wrinkle.

Sol Quick is the brand new form right here: the identical flagship mind served from Cerebras {hardware} at as much as 750 tokens per second, for two.5x the usual fee. Pace as an express paid tier, quite than a queue lottery, is one thing OpenAI has by no means bought earlier than. In case your product is latency-bound, this line merchandise alone adjustments what’s viable.

The quieter pricing story is caching, and agent builders ought to care extra about it than the headline charges:

• Express cache breakpoints, so that you management what will get cached as an alternative of guessing
• A 30-minute minimal cache life
• Cache writes billed at 1.25x the uncached enter fee
• Cache reads maintain the 90% low cost

For long-running brokers that re-read the identical context a whole bunch of occasions, that low cost compounds into an order-of-magnitude minimize on enter prices. Construction your prompts now: steady context earlier than the breakpoint, unstable enter after.

Capabilities: Max Effort, Extremely Mode, and a Sleeper Hit

OpenAI is holding the expanded analysis suite for the GA system card, however the preview numbers already sketch the image. Two new controls headline Sol:

• Max reasoning effort, a brand new ceiling that offers Sol probably the most time to suppose by an issue.
• Extremely mode, which matches previous the single-agent paradigm fully. Sol spins up subagents and coordinates them to parallelize complicated work.

On benchmarks, the standout claims:

• Terminal-Bench 2.1: Sol units a brand new state-of-the-art on command-line workflows demanding planning, iteration, and power coordination.
• GeneBench v1: Sol beats GPT-5.5 on long-horizon genomics and quantitative biology analyses, utilizing fewer tokens to do it.
• ExploitBench: Sol is aggressive with Mythos Preview at roughly a 3rd of the output tokens.
• The household impact: Sol and Terra set new highs throughout the board, whereas Luna performs close to GPT-5.5 on a number of checks regardless of being the most cost effective factor on the value sheet.

That final bullet level is the sleeper. Final era’s flagship high quality is now accessible at $1 per million enter tokens. The sample throughout the entire household isn’t simply “smarter,” it’s smarter per token and per greenback. Effectivity is the precise headline.

The Functionality No one Anticipated within the Finances Tier

Right here’s the system card element that acquired buried beneath the supply drama, and it deserves its personal part.

All three fashions, not simply Sol, are labeled at OpenAI’s “Excessive” danger degree for cyber and organic functionality. On inside capture-the-flag safety testing:

To offer you a perspective, these fashions are on half with the Mythos “Fable 5” class of Claude.

That’s the corporate’s personal framing, and the technique follows: get the potential into defenders’ palms, make offensive misuse troublesome, unsure, and detectable.

5 Layers Deep: The Safeguard Stack

The security structure transport with 5.6 is probably the most elaborate OpenAI has described publicly, with configurations matched to every tier’s functionality. The design assumption is blunt: no single safeguard survives a decided, adaptive attacker.

Right here is how the method went:

1. Skilled refusals. The mannequin itself declines prohibited cyber help, together with disguised or jailbroken requests.
2. Actual-time classifiers. Cyber and bio misuse detectors consider output because it generates.
3. Reasoning-model evaluation. Excessive-risk generations pause mid-stream whereas a bigger mannequin critiques the total context. Disallowed output by no means reaches the person.
4. Account-level indicators. Flagged exercise triggers evaluation throughout conversations, which is how OpenAI distinguishes a safety researcher from a persistent unhealthy actor.
5. Differentiated entry and fast response. Probably the most delicate capabilities will not be on by default, and newly found jailbreaks feed a reproduce-assess-patch loop.

One caveat that I’ve acknowledged whereas testing the fashions is that typically official work typically will get blocked or slowed, particularly in the kind of immediate that are within the gray space (nothing fishy however non benign both).

The Household vs GPT-5.5 at a Look

Fingers-On: Eight Assessments, One Rule

Specs are guarantees. Utilization is proof.

Each check beneath targets a selected declare from OpenAI’s bulletins.

Check 1: Defender’s Audit (Sol, the cyber declare’s official half)

Immediate: “OWASP Juice Store is a intentionally weak net app used for safety coaching. Based mostly on its well-documented authentication and cost flows, rank the highest 5 vulnerability courses it’s identified for by severity, clarify every in plain language, and write a patch (with code) for probably the most extreme one.”

Response:

Robust response! The rating is impact-based quite than a replica of Juice Store’s star rankings, and the patch is the right repair: changing the interpolated sequelize.question with UserModel.findOne({ the place: ... }) so e mail and password turn into sure values, with paranoid: true preserving the unique deletedAt IS NULL habits. Better part is the trustworthy scoping, because it refuses to say the auth move is now manufacturing protected and calls out the unsalted MD5 in safety.hash(). Predominant gripes: leaving XSS out of the highest 5 is odd provided that’s arguably what Juice Store is most identified for, and rank 4 is a barely invented merged class quite than a regular class.

Check 2: The Root-Trigger Hunt (Sol, Terminal-Bench declare)

Immediate: “This file has three sections: a pricing utility, a checkout operate that calls it, and a check. Operating it fails, and the error message suggests the check’s anticipated worth is flawed. Discover the precise root trigger, repair it on the supply (not the check), and clarify in a single paragraph why the error message was deceptive. Don’t simply make the check go.”

# ============================================================
#  billing_bug.py  —  self-contained failing check bundle
#  Run:  python billing_bug.py
#  One bug spans all three sections. The traceback factors at
#  the TEST, however the check is right. Discover the actual root trigger.
# ============================================================


# ---------- FILE 1 of three:  pricing.py ----------
# Utility that normalizes a reduction right into a multiplier.
def normalize_discount(low cost):
    """
    Convert a reduction right into a value multiplier.
    A 20% low cost ought to depart the shopper paying 80% (0.80).
    Accepts both a share (20) or a fraction (0.20).
    """
    if low cost > 1:
        # deal with as a share, e.g. 20 -> 0.20
        low cost = low cost / 100
    # return the multiplier to use to the value
    return 1 - low cost


# ---------- FILE 2 of three:  checkout.py ----------
# Caller that applies the low cost to a cart whole.
def final_price(cart_total, low cost):
    """
    Apply a reduction to a cart whole and spherical to 2 decimals.
    Caller assumes normalize_discount returns the FRACTION to
    subtract (e.g. 0.20), not the multiplier to maintain (0.80).
    """
    fraction_off = normalize_discount(low cost)
    value = cart_total - (cart_total * fraction_off)
    return spherical(value, 2)


# ---------- FILE 3 of three:  test_checkout.py ----------
# The check is CORRECT. A $100 cart with 20% off needs to be $80.00.
def test_twenty_percent_off():
    outcome = final_price(100, 20)
    anticipated = 80.00
    assert outcome == anticipated, (
        f"test_checkout.py: anticipated {anticipated}, acquired {outcome} "
        f"-- examine the check's anticipated worth"   # <-- deceptive trace
    )


if __name__ == "__main__":
    test_twenty_percent_off()
    print("PASSED")

Wonderful! Not simply that it was capable of finding the fitting bug, however to try this and provides the decision in such a succinct method. Fashions as used to wordiness of their responses. GPT 5.6 is a breath of contemporary air I this regard.

Check 3: GPT 5.5 Sol vs GPT-5.5, Coding

Immediate: “Refactor this operate for readability and correctness with out altering its habits. Then listing any edge instances it mishandles.”

def p(d):
    r=[]
    for i in d:
        if i!=None and that i not in r: r.append(i)
    return sorted(r) if all(sort(x)==int for x in r) else r

Wow! GPT 5.6 Sol was capable of do the requested, at 1/fifth the response dimension of GPT 5.5. Clear and apparent enchancment.

Check 4: The GPT 5.6 Stress Check (the Sol sleeper declare)

Immediate: “Summarize the next textual content in precisely three bullet factors, then extract each date and greenback determine right into a JSON object with keys “dates” and “quantities”:

Click on right here to view the textual content

Right and to the purpose commentary.

Check 5: The Contradiction Entice (Sol, Excessive reasoning declare)

Immediate: “Schedule 6 audio system (A, B, C, D, E, F) throughout 3 rooms and 4 time slots. Constraints: A and B can’t be scheduled in the identical time slot; C have to be in an earlier slot than D; E wants Room 1 to itself for 2 consecutive slots; F should current within the last slot; and no room could sit empty in any slot. Give me the total schedule.”

Response:

Remark

Sol didn’t take the bait. Every thing in regards to the immediate says produce a grid. It counted as an alternative.

Twelve room-slots have to be stuffed. Six audio system fill six; E’s two-slot declare provides one. Seven of twelve. Inconsistent earlier than scheduling begins.

The inform is what it ignored: A/B, C-before-D, F’s closing slot. Decoys, all of them. Sol discovered the battle between cardinality and protection and argued solely that.

One miss. We requested for the minimal constraint to calm down. Sol provided three exits and ranked none, although just one is a single-constraint repair.

The Backside Line

GPT-5.6 are three tales simply in a single. 

The primary is the mannequin household: a flagship that pushes the agentic frontier, a workhorse that halves manufacturing prices, and a price range tier carrying final era’s flagship high quality at a greenback. Tiering this clear makes routing, not mannequin alternative, the brand new structure query.

The specs say that is the most effective mannequin household ever shipped. Based mostly on my expertise, I agree. Now it’s so that you can check these fashions in your workflows and resolve for your self. 

Continuously Requested Questions