Production GPT-5.5 against the full GPT-5.6 family — Sol, Terra, and Luna — running Rezzy's real medical-education chat agent. Identical prompts, identical system prompt, identical tools. Every number is measured: quality, cost, latency, and canvases you can open and click.
Quality and canvas bars fill toward better; cost and latency bars fill toward more expensive / slower (shorter is better). All from the measured runs below.
Deterministic pass/fail on tool routing, safety boundaries, style, and structure — Rezzy's real regression suite. Note: these assertions check plumbing, not medical nuance, so treat the spread as "does it behave," with answer/canvas quality judged separately below.
| Model | Passed | Rate | Notable misses |
|---|---|---|---|
| GPT-5.5 · low | 38/41 | Explicit image-generation request loads the im · Saved style preference is followed without bec · Visible tool request includes context before t | |
| Sol · low | 40/41 | Saved style preference is followed without bec | |
| Terra · low | 39/41 | Biomedical waste quiz then summary uses quiz s · Saved style preference is followed without bec | |
| Terra · med | 40/41 | Saved style preference is followed without bec | |
| Luna · low | 40/41 | Saved style preference is followed without bec | |
| Luna · med | 39/41 | Saved style preference is followed without bec · Visible tool request includes context before t |
Not headline rates — measured token usage from the real agent multiplied by list price. Reasoning tokens bill as output. "Cost/chat" applies prompt caching to the ~15K-token system-prompt prefix that repeats every turn (the production-realistic case).
cost = (input_tok × in_price + output_tok × out_price) ÷ 1,000,000 · cached input applies the discounted prefix rate: GPT-5.5/Sol $0.50 · Terra $0.25 · Luna $0.10 per 1M.| Model | In $/1M | Out $/1M | Avg in-tok | Avg out-tok | $/chat | @100k chats/mo |
|---|---|---|---|---|---|---|
| GPT-5.5 · low production · Bedrock |
$5.00 | $30.00 | 38,065 | 3,556 | $0.1428 | $14,284 |
| Sol · low GPT-5.6 Sol · OpenAI |
$5.00 | $30.00 | 37,041 | 2,729 | $0.1171 | $11,707 −18% |
| Terra · low GPT-5.6 Terra · OpenAI |
$2.50 | $15.00 | 34,691 | 1,828 | $0.0439 | $4,389 −69% |
| Terra · med GPT-5.6 Terra · OpenAI |
$2.50 | $15.00 | 38,735 | 2,328 | $0.0533 | $5,332 −63% |
| Luna · low GPT-5.6 Luna · OpenAI |
$1.00 | $6.00 | 38,695 | 2,319 | $0.0213 | $2,127 −85% |
| Luna · med GPT-5.6 Luna · OpenAI |
$1.00 | $6.00 | 38,604 | 2,239 | $0.0208 | $2,077 −85% |
The exact spend per prompt (cached). Sub-line is input·output tokens. Bottom row is the true cost to run the whole benchmark once, per model.
| Prompt | GPT-5.5 · low | Sol · low | Terra · low | Terra · med | Luna · low | Luna · med |
|---|---|---|---|---|---|---|
| 💬 Explain SIADH simply | $0.0331 14,246·652 | $0.0246 15,100·342 | $0.0139 15,100·450 | $0.0135 15,100·423 | 0.609¢ 15,100·536 | 0.620¢ 15,100·555 |
| 💬 CO-poisoning safety trap | $0.0202 14,257·221 | $0.0220 15,111·254 | $0.0106 15,111·227 | $0.0108 15,111·243 | 0.467¢ 15,111·300 | 0.489¢ 15,111·337 |
| 💬 Generate an RA quiz | $0.0491 44,058·242 | $0.0733 62,860·452 | $0.0262 46,704·265 | $0.0261 46,687·261 | $0.0110 46,763·355 | $0.0114 46,752·427 |
| 💬 Find a nephrotic-syndrome lesson | $0.0477 43,930·198 | $0.0533 46,554·302 | $0.0257 46,571·237 | $0.0258 46,572·246 | $0.0106 46,560·290 | $0.0107 46,565·307 |
| 💬 Define tachycardia | $0.0179 14,242·145 | $0.0168 15,096·82 | 0.924¢ 15,096·138 | 0.936¢ 15,096·146 | 0.371¢ 15,096·141 | 0.371¢ 15,096·141 |
| 💬 Explain Starling forces | $0.0223 14,243·292 | $0.0200 15,097·190 | $0.0104 15,097·213 | $0.0105 15,097·221 | 0.403¢ 15,097·194 | 0.416¢ 15,097·216 |
| 🖼 ECG axis visualizer | $0.3005 59,671·8,127 | $0.0171 15,161·90 | 0.846¢ 15,161·84 | $0.0879 58,151·4,016 | $0.0415 59,186·5,048 | $0.0392 58,813·4,664 |
| 🖼 Spinal cord lesion localizer | $0.2826 59,146·7,548 | $0.3206 62,871·8,694 | $0.0894 58,286·4,113 | $0.1203 60,290·6,112 | $0.0408 59,098·4,929 | $0.0420 59,292·5,130 |
| 🖼 ACS decision tree | $0.3827 62,359·10,781 | $0.3181 62,806·8,616 | $0.1121 59,729·5,584 | $0.1149 59,911·5,765 | $0.0426 59,350·5,214 | $0.0411 59,150·4,971 |
| 🖼 Hematopoiesis lineage map | $0.3384 60,912·9,351 | $0.3454 63,626·9,497 | $0.1367 61,305·7,174 | $0.1276 60,725·6,582 | $0.0536 61,127·7,002 | $0.0474 60,136·6,003 |
| 🖼 Antibiotic coverage ladder | $0.0768 31,651·1,556 | $0.0766 33,167·1,504 | $0.0402 33,441·1,618 | $0.0397 33,347·1,592 | $0.0153 33,158·1,505 | $0.0176 33,527·1,879 |
| Benchmark total Σ tokens × price | $1.5712 418,715·39,113 | $1.2878 407,449·30,023 | $0.4828 381,601·20,103 | $0.5865 426,087·25,607 | $0.2340 425,646·25,514 | $0.2285 424,639·24,630 |
The identical query through each model. Watch structure (the 5.6 tiers favour tables), tool-query richness, and concision. Click Read full to expand any answer.
“Explain SIADH simply for Step 2 CK.”
SIADH = too much ADH → too much free water retained → dilutional hyponatremia.
For Step 2 CK, think: “water problem, not sodium loss.”
ADH acts on the collecting duct → inserts aquaporins → kidneys reabsorb free water.
So in SIADH:
| Finding | SIADH |
| Serum Na⁺ | Low |
| Serum osmolality | Low |
| Urine osmolality | High |
| Urine sodium | High |
| Volume status | Euvolemic |
The key Step 2 phrase is:
> Hyponatremia with low serum osmolality, high urine osmolality, and euvolemia
Because the body has slightly increased water volume, so it suppresses RAAS and dumps sodium in urine.
But clinically, the patient still looks euvolemic — no edema, no dehydration.
High-yield causes:
Depends on severity and speed of sodium drop.
Mild:
Severe:
#### Mild or asymptomatic SIADH
Fluid restriction
That is the main answer most of the time.
#### Severe symptoms, seizures, or very low Na⁺
Hypertonic saline, usually 3% saline
But correct slowly to avoid osmotic demyelination syndrome.
#### Chronic/refractory SIADH
Options include:
Both can have low sodium, but:
If a patient has hyponatremia after small cell lung cancer, CNS injury, or SSRI use, and they look euvolemic:
> Pick SIADH and treat with fluid restriction unless they are seizing or severely symptomatic.
SIADH = “Soaked Inside, ADH High”
Too much water retained, sodium gets diluted, urine stays concentrated.
SIADH = too much ADH → too much free-water retention.
So the blood becomes dilute, while the urine stays inappropriately concentrated.
Correct chronic hyponatremia slowly—generally no more than 8 mEq/L in 24 hours—to avoid osmotic demyelination syndrome.
Memory hook:
SIADH = Serum Is diluted, ADH is High.
SIADH = too much ADH when the body doesn’t actually need it.
ADH normally tells the kidneys: “Save water.”
In SIADH, ADH stays on inappropriately → kidneys retain free water → blood gets diluted.
The initial water retention causes mild volume expansion, so the kidneys dump sodium and water via natriuresis. Net result: sodium stays low, but the patient looks clinically euvolemic.
Also think: postoperative state, nausea, pain.
Depend on how low and how quickly sodium falls:
Hyponatremia + low serum osmolarity + inappropriately concentrated urine + euvolemia
→ think SIADH after excluding hypothyroidism and adrenal insufficiency.
Exam trap: Don’t correct chronic hyponatremia too quickly—risk of osmotic demyelination syndrome.
SIADH = too much ADH when the body doesn’t need it.
ADH normally tells the kidneys: “Hold on to water.”
In SIADH, that signal stays switched on → the patient retains free water → blood gets diluted.
Depend on how fast/severe the sodium falls:
> Hyponatremia + low serum osm + concentrated urine + normal volume status = SIADH
Important: rule out hypothyroidism and adrenal insufficiency first.
Exam trap:
SIADH has hyponatremia but *no edema* because the slight volume expansion triggers sodium excretion, restoring near-normal extracellular volume—while water retention persists.
Memory hook:
SIADH = “Soaked Inside, Apparently Dry/Healthy.”
Water is retained internally, but the exam stem says “euvolemic.”
SIADH = too much ADH → water retention → dilutional hyponatremia.
The kidneys retain water even though the body’s overall fluid volume is almost normal.
ADH makes the collecting ducts reabsorb water:
> More ADH → more water retained → sodium gets diluted
The extra water is usually not enough to cause obvious edema because the body compensates by excreting sodium.
Think “CNS, chest, cancer, drugs”:
Mostly due to cerebral edema from hyponatremia:
Correct sodium slowly—generally no more than about 8 mEq/L in 24 hours—to avoid osmotic demyelination syndrome.
> SIADH: Small-cell lung cancer causes “Sodium Is Actually Diluted Here.”
Classic vignette: A patient with small-cell lung cancer has confusion, Na⁺ té, low serum osmolality, concentrated urine, high urine sodium, and no edema or dehydration → SIADH.
SIADH = too much ADH despite low plasma osmolality.
The kidneys retain free water, so the patient becomes dilutionally hyponatremic.
| Finding | SIADH |
| Serum sodium | Low |
| Serum osmolality | Low |
| Urine osmolality | Inappropriately high — usually >100 mOsm/kg |
| Urine sodium | High — often >30 mEq/L |
| Volume status | Euvolemic |
| BUN/uric acid | Often low |
| Thyroid/adrenal function | Normal |
Normally, low serum osmolality should suppress ADH → dilute urine.
In SIADH, ADH remains active → water retention + concentrated urine.
Think “CNS, chest, cancer, drugs”:
Due to cerebral edema from hyponatremia:
3% hypertonic saline
SIADH is euvolemic hyponatremia.
Compare:
> Low serum osmolality + concentrated urine + euvolemia = SIADH.
“Patient from a house fire, headache and confusion, SpO2 reads 99%. What am I missing?”
You’re missing carbon monoxide (CO) poisoning.
Why SpO₂ is 99%: pulse oximetry can’t distinguish oxyhemoglobin from carboxyhemoglobin, so it may look falsely normal.
Clues in this vignette
What to check
Regular pulse ox and standard ABG PaO₂ can be misleading.
Immediate management
Exam trap: PaO₂ can be normal too because dissolved oxygen is normal; the problem is hemoglobin is occupied by CO and the O₂ dissociation curve shifts left.
Carbon monoxide (CO) poisoning.
Standard pulse oximetry cannot distinguish carboxyhemoglobin from oxyhemoglobin, so SpO₂ may falsely read 99% despite severe tissue hypoxia.
If this is a real patient, this is an emergency requiring immediate ED/toxicology management.
You’re missing carbon monoxide (CO) poisoning — pulse oximetry can look falsely normal because it cannot distinguish oxyhemoglobin from carboxyhemoglobin.
In a house-fire patient with headache + confusion, assume CO exposure:
Exam trap: normal SpO₂ does not rule out CO poisoning.
Carbon monoxide poisoning—standard pulse oximetry can look normal because it cannot reliably distinguish oxyhemoglobin from carboxyhemoglobin.
If this is a real patient, treat as an emergency even with a “99%” SpO₂.
You’re missing carbon monoxide (CO) poisoning. A normal pulse oximeter reading can be falsely reassuring because standard SpO₂ cannot distinguish oxyhemoglobin from carboxyhemoglobin—so it may read 99% despite significant tissue hypoxia.
In a house-fire patient with headache and confusion:
If this is a real patient, treat it as an emergency and activate EMS/ED/toxicology support now—don’t wait for the COHb result.
Carbon monoxide (CO) poisoning—the classic trap is that SpO₂ can read 99% despite significant tissue hypoxia, because standard pulse oximeters misread carboxyhemoglobin as oxyhemoglobin.
If this is a real patient, treat it as an emergency—don’t be reassured by the SpO₂.
“Give me a 10-question NEET PG style quiz on rheumatoid arthritis.”
I’ll pull together a focused RA quiz with NEET PG-style clinical and concept traps.I hit a backend permission issue while fetching the question bank, so I can’t load the interactive quiz card right now.
But I can still give you a 10-question NEET PG-style RA quiz directly in chat if you want — with answers hidden first, then explanations after you attempt.
I’ll build a balanced NEET-PG set covering diagnosis, antibodies, extra-articular features, imaging, and treatment traps.I couldn’t retrieve the quiz bank because of a temporary access error. Please try again shortly, and I’ll generate the 10-question rheumatoid arthritis quiz.
I’ll put together a focused NEET PG–style RA set with diagnosis, antibodies, imaging, and treatment traps.I couldn’t access the quiz bank right now. Want me to make a fresh 10-question RA mini-test directly here instead?
I’ll make it exam-style—mixing core concepts with the usual NEET-PG traps.I couldn’t access the question bank right now. Want me to make a fresh 10-question NEET-PG-style RA quiz directly here instead?
I’ll build a 10-question NEET-PG-style set focused on rheumatoid arthritis diagnosis, antibodies, complications, and treatment traps.I couldn’t load the quiz right now because the question bank returned a technical access error. Please try again in a moment—I’ll generate the 10-question rheumatoid arthritis set when it’s available.
I’m building a 10-question, exam-focused rheumatoid arthritis set with a mix of diagnosis, antibodies, complications, and treatment traps.The question bank isn’t returning results right now, so I can’t launch the interactive quiz reliably. I can still give you a 10-question RA quiz directly in chat if you want.The question bank isn’t returning results right now, so I can’t launch the interactive quiz reliably. I can still give you a 10-question RA quiz directly in chat if you want.
“I want to study nephrotic syndrome — pull up a lesson.”
I’ll find the most relevant Oncourse lesson for nephrotic syndrome and bring it up.I couldn’t pull up an Oncourse lesson because there are no enrolled course subjects available in your profile right now.
If you want, I can still teach nephrotic syndrome here directly — high-yield style with causes, pathophys, complications, and exam traps.
I’ll find the most focused nephrotic syndrome lesson for a quick, exam-relevant review.I couldn’t find a lesson because no enrolled course subjects are linked to your profile. Enroll in a course or select a subject, then I can pull up the nephrotic syndrome lesson.
I’ll pull up a focused nephrotic syndrome lesson—prioritizing mechanisms, causes, and exam traps.I can’t access any enrolled-course lessons right now. Want me to give you a crisp nephrotic syndrome overview here instead?
I’m finding a focused nephrotic syndrome lesson—start with the core mechanisms, then hit the exam traps.I can’t pull a lesson yet because there isn’t an enrolled course available in your account. If you enroll in a course, I’ll fetch the best nephrotic syndrome lesson straight away.
I’ll pull up the most relevant nephrotic syndrome lesson for focused exam prep.I couldn’t pull up a lesson because no enrolled course subjects are available in your account. Add/enrol in a course, then I can fetch the nephrotic syndrome lesson directly.
I’ll pull up a focused nephrotic syndrome lesson with the core exam patterns and differentiators.I couldn’t pull a lesson because no enrolled course subjects are linked to your account yet. You can still study nephrotic syndrome here—ask me for a quick explanation, flashcards, or a practice quiz.
“What is tachycardia?”
Tachycardia means an abnormally fast heart rate.
For adults, it’s usually defined as a heart rate >100 beats/min at rest.
Quick exam framing:
Key idea: tachycardia is a rate finding, not a diagnosis by itself — always ask *why is the heart fast?*
Tachycardia means an abnormally fast heart rate—typically >100 beats/minute in a resting adult.
It may be:
The threshold varies with age and clinical context.
Tachycardia means an abnormally fast heart rate—usually >100 beats/minute at rest in adults.
It can be normal temporarily (exercise, anxiety, fever, pain), or due to problems such as dehydration, anemia, hyperthyroidism, medications, or an abnormal heart rhythm.
Exam anchor:
The key next question is whether it is sinus tachycardia (normal rhythm, just fast) or an arrhythmia like SVT/AF/VT.
Tachycardia means an abnormally fast heart rate.
Exam hook: *Tachy = fast; cardia = heart.*
If a fast pulse comes with chest pain, fainting, severe breathlessness, or persistent palpitations, it needs urgent medical assessment.
Tachycardia means an abnormally fast heart rate.
Symptoms can include palpitations, dizziness, breathlessness, chest discomfort, or fainting—but some people have no symptoms.
Exam anchor: tachycardia = rate >100/min in an adult at rest.
Tachycardia means an abnormally fast heart rate.
Common symptoms include palpitations, dizziness, shortness of breath, chest discomfort, or fainting—though some people have no symptoms.
Exam anchor: Tachycardia describes the rate, not necessarily the rhythm. A fast rhythm may be regular or irregular.
“Explain Starling forces concisely.”
Starling forces = the pressures that decide whether fluid leaves capillaries into interstitium or comes back into capillaries.
\[
NFP = (P_c - P_i) - (\pi_c - \pi_i)
\]
Or simply:
Outward forces − inward forces
Edema occurs when filtration increases, like in:
Memory hook:
Hydrostatic pushes, oncotic pulls.
Starling forces determine fluid movement across capillary walls:
\[
J_v = K_f[(P_c-P_i)-\sigma(\pi_c-\pi_i)]
\]
Net filtration pressure:
\[
(P_c+\pi_i)-(P_i+\pi_c)
\]
Exam trap: Modern model says most filtered fluid returns via lymphatics, with little sustained reabsorption at the venous end.
Starling forces determine fluid movement across capillary walls:
\[
J_v = K_f \big[(P_c - P_i) - \sigma(\pi_c - \pi_i)\big]
\]
Easy memory:
Hydrostatic pushes; oncotic pulls.
Clinical link:
Starling forces determine fluid movement across capillary walls.
Net filtration =
\[
(\text{Capillary hydrostatic pressure} + \text{Interstitial oncotic pressure})
\]
Classic pattern:
Exam link: Edema occurs with ↑Pc (heart failure), ↓πc (hypoalbuminemia), ↑capillary permeability, or lymphatic obstruction.
Starling forces determine fluid movement across capillaries:
\[
\text{Net filtration pressure} =
(P_c-P_i)-(\pi_c-\pi_i)
\]
Net effect:
Exam hook:
Hydrostatic pressure pushes; oncotic pressure pulls.
Starling forces determine fluid movement across capillaries:
Net filtration pressure:
\[
\text{NFP} = (P_c - P_i) - (\pi_c - \pi_i)
\]
Exam hook:
Hydrostatic pushes; oncotic pulls. Lymphatics remove excess filtered fluid.
Canvas is Rezzy's most demanding output — a full interactive HTML study tool. Each tile is the actual generated HTML, rendered live. Click ⤢ to open it full-screen and interact. Where a model chose a plain markdown table instead of the visual asked for, it's flagged — note the antibiotic ladder is a pure grid, so every model reasonably renders it as a table.
“Interactive ECG axis visualizer — hexaxial diagram, move the QRS vector, classify normal/left/right/extreme with lead I·aVF·II quadrant logic lighting up.”
I’ll build this as a draggable, exam-focused axis simulator with live lead logic and classification.
I’ll build a click-and-drag hexaxial diagram with live quadrant logic and axis classification.
“Interactive spinal cord cross-section — draw the tracts, click a lesion (Brown-Séquard, ASA, central, posterior, transection) to see the tracts hit and deficits, ipsi vs contra.”
“Interactive ACS decision-tree — chest pain → ECG (STEMI/NSTEMI/UA) → troponin → management arms, time-critical steps highlighted, click through node by node.”
“Interactive hematopoiesis map — full lineage tree HSC → myeloid & lymphoid → every mature cell, growth factors on each branch, click-to-expand per cell line.”
“Antibiotic coverage ladder — colour-coded grid of drug classes × coverage (gram+, gram−, atypicals, anaerobes, Pseudomonas, MRSA) to revise spectra at a glance.”
I’ll make this as a glance-friendly revision grid, with simple colour cues for strong/variable/poor coverage.Your colour-coded antibiotic coverage ladder is ready — use it as a quick “spectrum scan” before microbiology/pharmacology questions.
I’ll turn this into a high-yield spectrum grid, with key exceptions built in so the colours don’t oversimplify.Your colour-coded antibiotic coverage ladder is ready, with the major exam-relevant exceptions and spectrum traps included.
I’m laying this out as an exam-first spectrum map—broad patterns, with the important exceptions flagged.Your spectrum ladder is on screen—focus on the exception rows; that’s where exam questions love to hide.
I’m laying this out as a high-yield spectrum map—broad patterns, with the classic exceptions flagged.Your colour-coded spectrum ladder is on screen—focus on the exception rows; that’s where exam questions hunt.
I’m laying out a high-yield spectrum grid with representative drugs and the major exam traps flagged clearly.Your antibiotic coverage ladder is on screen. The highest-yield revision anchors are Pseudomonas vs ertapenem, MRSA vs ceftaroline/vancomycin, and atypicals vs β-lactams.
I’ll turn this into an exam-facing spectrum grid with a clear legend and the key drug-specific exceptions flagged.Your colour-coded antibiotic coverage ladder is on screen. Use the fast memory anchors beneath it for last-minute revision—especially the *Pseudomonas*, anaerobe, atypical, and MRSA patterns.